Every Energy AI Initiative Stalls in the Same Three Places. Robots & Pencils Names Them.
Robots & Pencils
Three organizational failures. One misdiagnosis. A three-part series that tells energy leaders exactly where to look.
Robots & Pencils, an applied AI engineering partner known for high-velocity delivery and measurable business outcomes, published The Fault Line, a three-part series examining the organizational failures keeping energy AI trapped in perpetual pilot mode.
Forty percent of utility control rooms will deploy AI-driven operators by 2027, according to Gartner. Yet fewer than seven percent of energy organizations have gone live with even one AI use case, according to IDC and AWS research. The gap between investment and execution continues to widen across the sector.
The Fault Line argues the problem lives in three specific organizational breakdowns that repeatedly prevent AI from reaching production environments and generating operational learning at scale. Scott Young, EVP of Growth and Strategic Alliances at Robots & Pencils, wrote the series for the energy executive who has approved the budget, built the pilot, and is still waiting for AI to run.
“Energy executives are moving faster through decisive action that turns AI investment into operational advantage,” said Young. “Every quarter spent in evaluation is a quarter of compounding operational learning moving somewhere else. That is the fault line. And it is solvable.”
Three Articles. Three Failures. One Compounding Reality.
Part 1 – Going Live with Energy AI Starts with One Decision. The applications energy executives are waiting on are already ready to deploy. They have been for years. The first article examines the one thing standing between investment and results, and it is not technology.
Part 2 – Energy AI Operator Trust Is Earned by Design. When AI stalls in the control room, the default explanation is operator resistance. The second article argues that explanation is aimed at the wrong problem entirely and that the organizations making the most progress stopped trying to manage adoption and started doing something else.
Part 3 – The Energy AI Architecture Decision That Outlasts Every Tool. Most energy organizations are not building AI. They are accumulating it. The third article names the difference between a collection of tools that cannot learn from each other and an architecture that compounds and explains why no one selling AI tools has a financial incentive to close that gap.
Why This Matters Now
Investment, urgency, and operational pressure are converging quickly across the industry. The DOE’s Genesis Mission mobilized $293 million to advance AI in grid operations. ERCOT launched a dedicated Enterprise Data and AI organization in January 2026. At the same time, many organizations are adding AI systems faster than they are building the operational foundations required to scale them effectively.
The Fault Line identifies three areas where that gap consistently appears including executive decision velocity, operator-centered system design, and architectures capable of compounding intelligence across the enterprise. The series also addresses the regulatory and operational realities utility leaders face while advancing AI initiatives within NERC CIP environments.
Each article stands on its own. Together, the series presents a clear argument for how energy organizations move from isolated pilots to operational AI systems that improve through live deployment.
“The energy sector is entering a period where AI advantage compounds faster than most executives expect,” Young said. “The organizations deploying now will be operating systems shaped by thousands of hours of real-world learning while others are still refining pilots. The opportunity belongs to the organizations willing to move.”
Part 1 – The Fault Line: Going Live with Energy AI Starts with One Decision
Scott Young
This three-part series examines the three organizational failures that keep energy AI in perpetual pilot and what leaders who have moved past them did differently. Each article stands alone. The full series is the argument.
The energy executives I talk to have already committed to generative and agentic AI. The budgets are approved. The strategic plans name it. What most have not committed to yet is treating that deployment as an operational decision rather than an ongoing evaluation. That distinction is the entire ballgame.
Gartner projects that 40 percent of utility control rooms will deploy AI-driven operators by 2027. Nearly all energy CIOs plan to increase AI investment, at an average spending increase of 38 percent. The DOE’s Genesis Mission mobilized $293 million targeting AI’s role specifically in grid operations and reliability. The conditions for large-scale energy AI deployment have never been more aligned.
An IDC and Amazon Web Services (AWS) study of more than 900 organizations found that fewer than 7 percent have reached full production with even one AI use case.
The standard explanation is that energy is a uniquely complex operating environment. Legacy systems, fragmented data, strict regulation, and safety-critical infrastructure are real constraints. Energy leaders reach for them first when AI stalls. They are the setting, not the cause.
The actual reason is less comfortable. The applications energy leaders want most are already ready to deploy. Most organizations are waiting for a technology problem to solve when the problem is organizational.
The Energy AI Readiness Gap Nobody Is Naming
According to the Federation of American Scientists’ assessment of the Department of Energy’s priority AI applications, nearly half are high-impact and ready to deploy today. Operations and reliability use cases score 3.6 out of 5.0 on deployment readiness, the highest category in the entire assessment.
The most urgently needed applications are also the most architecturally mature.
That creates a specific kind of organizational trap. When technology readiness runs ahead of organizational readiness, leaders rarely recognize the gap for what it is. An initiative stalls, and the natural assumption is that something technical still needs improvement. The model needs more training data. The data environment needs more work. The pilot needs another quarter before it can prove itself.
What actually needs improvement is what we call the decision architecture gap. Most energy organizations have not built the organizational capacity to evaluate, commit to, and scale AI applications based on evidence of operational value rather than proof of technical completion.
What the Data Is Already Telling You
Energy companies already have the data. They are waiting on the decision to act on it.
NREL’s Open Energy Data Initiative hosts 2.6 petabytes of data across more than 2,000 datasets from 227 providers. Utilities already hold enormous volumes of AMI telemetry, SCADA signals, outage history, maintenance logs, and weather correlations. The question is not whether useful data exists. The question is whether it is being treated as institutional memory or as archived history.
These are not the same thing. Archived history answers questions when asked. Institutional memory learns continuously, surfacing patterns, updating predictions, and sharpening with every new cycle of operational data. We call this the institutional memory framework. The architectural commitment to treat operational data as a living learning system rather than a reference archive is what separates organizations that compound AI advantage from those that accumulate AI cost.
The data foundation is already there. The decision about what to build on it is the only variable left.
The Compounding Cost of the Wait
The energy sector is entering a period where AI advantage compounds. Organizations that go live now will be running systems that have learned through thousands of hours of real operating conditions by the time their competitors are still refining pilots.
Grid operations, reliability, and predictive maintenance are the applications energy leaders typically pursue first. They are also the ones that compound most sharply with continuous learning. A predictive maintenance system that has processed two years of real failure data across a fleet of transformers is qualitatively different from a system that has processed none. That gap does not close when the second organization eventually decides to start. It widens.
This is the real cost of treating AI deployment as a technology problem to be solved rather than an operational commitment to be made. The loss is not a single delayed quarter. It is the accumulated learning gap that grows while organizations wait for a breakthrough that is not coming.
Where the Decision Lives
The energy leaders making the most meaningful progress on AI are the ones who answered a harder question. Which operational outcomes matter enough to organize the entire effort around?
The starting point is simple. Grid load forecasting, AMI analytics, outage prediction, and field operations automation are all deployable today as agentic AI teammates that act on operational data utilities already own, execute decisions, and learn from every cycle. They are the foundation that makes every more complex application possible because each one builds the organizational infrastructure for learning, not just for experimenting.
The right question for energy executives is not whether to invest in AI. That investment is already moving. The right question is whether the organization is built to learn from what it deploys, or whether each initiative will generate insight for one team instead of compounding advantage across the enterprise.
Going live with AI in energy begins with a decision about what the organization is building toward and the commitment to treat every deployment as a step in that direction rather than a standalone test of the technology.
That decision is available right now. The technology has been ready for a while.
Building AI that operators will actually use requires a different kind of design than most energy organizations are attempting. Read Part 2: “Energy Operator Trust is Earned By Design”.
About the Author
Scott Young is EVP of Growth and Strategic Alliances at Robots & Pencils, where he works with energy executives to move from decision to live. Connect with Scott on LinkedIn.
Key Takeaways
The AI applications energy leaders want most score highest on deployment readiness. The gap between investment and going live is organizational, not technological.
Energy companies already have the data. Most are treating rich operational data as archived history rather than institutional memory. That distinction determines whether AI compounds or stalls.
Energy companies already have the data. The decision about what to build on it is the only variable left. Organizations that make that decision and treat every deployment as a step toward compounding intelligence rather than a standalone technical test are the ones that go live and stay live.
AI advantage in energy compounds over time. Organizations that go live now will hold a learning gap over later movers that closes only with years of real operational data, not with a better pilot program.
The right leadership question is not whether AI is ready. It is whether the organization is built to learn from every deployment rather than evaluate each one in isolation.
FAQs
What does organizational readiness for AI mean in energy?
It means the organization has defined which operational outcomes matter most, built the data infrastructure to support continuous learning against those outcomes, and established the decision process to evaluate and scale AI based on operational evidence rather than technical completion.
Why do so many energy AI initiatives stall after a successful pilot?
Pilots succeed at the local level because they are designed to prove technical performance. They stall at scale because scaling requires organizational infrastructure — shared data foundations, clear outcome definitions, and the governance to move from proof to live. Most organizations have not built those yet.
What is the difference between archived data and institutional memory for AI?
Archived data answers questions when asked. Institutional memory learns continuously, surfacing patterns, sharpening predictions, and improving with every cycle of operational data. The distinction determines whether AI compounds across the enterprise or produces isolated results for individual teams.
How do utilities close the gap between AI pilots and live deployment?
The fastest path from decision to live is standardizing the data foundation before scaling the AI system. Organizations that treat operational data as a shared institutional asset rather than system-specific input compress deployment timelines significantly and avoid the fragmentation that keeps most pilots from going live.
How long does it actually take to go live with energy AI?
It depends almost entirely on data infrastructure readiness, not model complexity. Organizations that have standardized their data foundations and committed to treating operational data as institutional memory have gone live with AI in 90 to 120 days. Organizations that treat each deployment as a custom integration build take two to three times as long and often stall before going live.
Which energy AI applications are ready to deploy today?
Operations and reliability use cases score highest on deployment readiness across the DOE’s priority applications. Grid load forecasting, AMI analytics, outage prediction, demand response optimization, and field operations automation are all deployable now using data utilities already collect. The barrier is organizational commitment, not technology availability.
What is the cost of waiting to deploy AI in energy?
The primary cost is the compounding learning gap. AI systems improve through real operational data. Organizations that go live now will be running materially smarter systems in two years than organizations that delay. That gap widens with time and does not close simply by starting later with better technology.
Part 2 – The Fault Line: Energy AI Operator Trust Is Earned by Design
Scott Young
This three-part series examines the three organizational failures that keep energy AI in perpetual pilot and what leaders who have moved past them did differently. Each article stands alone. The full series is the argument.
The default explanation for why AI stalls in energy operations goes something like this: operators resist change. They are comfortable with how things work, skeptical of technology they did not choose, and protective of the expertise they have spent decades building. The prescription that follows is predictable. Train them. Communicate more clearly. Involve them earlier. Manage the change.
This explanation has merit. It is just aimed at the wrong problem.
These are agentic AI systems, ones that surface recommendations, trigger actions, and learn from every operator decision. That distinction determines how trust gets built. Operator trust is earned through design. The organizations achieving live AI deployment in energy have stopped treating operator skepticism as something to overcome and started treating it as the signal that shapes how they build.
The Confidence Paradox
AI is most valuable in precisely the decisions where experienced utility operators are most confident. This is not a coincidence. It is the nature of complex operational environments. Grid stability calls, equipment risk assessments, and outage response sequencing are the decisions where utility operators carry the deepest accumulated judgment. In many organizations pursuing grid modernization, that knowledge is not documented anywhere. It retires when the operator does. These are also the decisions where AI can process patterns that no individual, regardless of experience, can evaluate at the speed and scale that modern grid operations demand.
This creates a specific problem. When an AI system surfaces a recommendation that contradicts an experienced operator’s intuition, the operator does not typically pause and reconsider. They override. Sometimes they are right to do so. Often, neither side ever finds out, because the correction disappears into a workflow without becoming feedback. The AI does not learn from the override. The organization does not learn from the pattern. The system gets evaluated on whether operators accepted its recommendations, not on whether acceptance or rejection produced better outcomes.
A Dalhousie University review published in Energy identified building human operator trust as the primary open challenge in the field, ahead of model accuracy, computational requirements, and integration complexity. That ranking matters. It reflects what researchers studying the most advanced energy AI deployments believe is holding back the most promising applications.
What Change Management Gets Wrong
The standard response to operator skepticism focuses on the operator. Train them differently. Explain the model’s reasoning. Show the accuracy data. Demonstrate value over time.
What this approach misses is that operator confidence is earned through repeated, verifiable demonstrations at the specific decision types operators care about most. Those demonstrations require something most implementations do not provide: a visible, credible track record at the local level before the system asks for broader authority.
Gartner warns that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing unclear business value and inadequate risk controls as the primary causes. In energy operations, inadequate risk controls and operator trust are the same thing. An operator who does not trust a recommendation will not act on it. An organization that cannot get operators to act on AI recommendations cannot demonstrate business value. The cancellation follows from the design failure, not from the technology.
Alsaigh et al., writing in Frontiers in Energy Research, analyzed 3,568 academic papers on AI governance in energy and found that explainability is one of the most significant and least developed barriers to operator trust. The systems being deployed in energy are largely not designed to give utility operators what they need to verify, challenge, and ultimately rely on AI recommendations. That is a design gap, not a training gap.
In regulated utility environments operating under NERC CIP standards, this design gap carries a second consequence. AI systems that cannot show their reasoning, support human override, and maintain audit trails fail both the trust requirement and the compliance requirement simultaneously. The design approach that earns operator trust in control room operations is also the one that satisfies regulatory expectations for human oversight of safety-critical decisions.
Designing Energy AI for Operator Trust, Not Adoption
The organizations deploying AI that reaches production in energy are not persuading operators. They are proving themselves to operators, one decision category at a time.
Research from Argonne National Laboratory’s GridMind system and the University of Vermont’s PowerDAG framework illustrates this principle at the applied research level. Both were built explicitly for expert decision-support augmentation rather than operator replacement. PowerDAG achieves a 100 percent task success rate specifically because it incorporates just-in-time human supervision as an architectural feature, not as a fallback. The operator-in-the-loop is not a limitation of the system’s current capability. The operator in the loop is what makes the system trustworthy enough to act on.
This design commitment is consistent across every advanced energy AI system in the current research landscape. Each of the following was built with operator augmentation as the primary design requirement, not an afterthought:
Argonne GridMind: conversational power system analysis built for expert decision-support augmentation, not operator replacement
University of Vermont PowerDAG: 100 percent task success rate via just-in-time human supervision as a core architectural feature
University of Toronto Grid-Agent: sandboxed execution with operator-controlled rollback before any AI-recommended action is implemented
Texas A&M X-GridAgent: natural language queries with human feedback loops built into the three-layer architecture
Every production-grade energy AI system identified in the current research literature shares this design commitment. That is the finding. The approach starts AI deployment at narrow, verifiable decision categories, builds a track record utility operators can see and challenge, and earns expanded scope based on demonstrated accuracy rather than elapsed time or training hours. It treats operator confidence as something AI must demonstrate, and organizational readiness as something that follows from the design.
Progressive trust architecture is the design approach of starting AI deployment at narrow, verifiable decision categories, building a track record utility operators can see and challenge, and earning expanded scope based on demonstrated accuracy rather than elapsed time or training hours. It treats operator confidence as something AI must demonstrate, not something organizations must develop.
A Tampere University study published in February 2026 found exactly this pattern in practice, conducting 16 interviews across nine departments of a Nordic energy company and identifying 41 AI-related use cases. Employees described successful AI introduction through incremental steps that aligned with existing workflows. They described it consistently as an evolution, one that fit the existing shape of the work rather than demanding the work reshape itself.
The Operator as Feedback Architecture
When the design takes hold, the dynamic inverts. Operator skepticism becomes the most valuable signal in the system.
Every time a utility operator reviews an AI recommendation, accepts it, overrides it, or flags it as wrong, that interaction carries information the system needs to improve in an operator feedback loop. In an agentic AI system, every human interaction with a recommendation is training data. That is what makes operator trust an architectural requirement, not a change management task. Organizations designed to capture and act on those signals are going live with AI that compounds in intelligence over time. Organizations that treat operator involvement as a transition phase on the way to full automation are managing adoption in perpetuity.
EPRI’s RADAR Initiative treats human capital development as a deployment prerequisite, not a follow-on activity. That sequencing reflects an understanding that the system’s intelligence and the operator’s intelligence need to develop in parallel, each informing the other, before the combination is ready to take on the decisions that matter most for grid modernization and operational reliability.
The organizations that earn operator trust design AI around the rules operators already follow. The operator’s existing process becomes the specification. Trust follows from the design.
Why Energy AI Operator Trust Is a C-Suite Problem
Energy AI operator trust is an architecture decision, and it belongs in the executive conversation alongside every other architectural decision the organization is making.
Energy leaders who reframe it that way will find their AI initiatives stop requiring managed adoption programs. When a system proves itself in decisions utility operators already own, and when it visibly learns from every interaction rather than ignoring operator judgment, trust follows from the design rather than preceding it.
In the energy organizations getting this right, the technology earns the operators. That is the design commitment that everything else follows from.
Scott Young is EVP of Growth and Strategic Alliances at Robots & Pencils, where he works with energy executives to move from decision to live. Connect with Scott on LinkedIn.
Key Takeaways
Operator skepticism in energy AI is a design signal, not a barrier. Organizations that treat it as a change management problem are solving the wrong problem.
Gartner’s 40 percent agentic AI cancellation projection is not a technology forecast. It is an operator trust forecast. In energy, unclear business value and inadequate risk controls are the same failure.
Every production-grade energy AI system in current research shares one design principle: AI augmentation of operator judgment rather than routes around it. The approach starts narrow and earns scope through demonstrated accuracy. That is a prerequisite for going live, not a preference.
The operator feedback loop is the most valuable learning signal in an energy AI system. Capture it by design or manage adoption in perpetuity.
The organizations that earn operator trust design AI around the rules operators already follow. The operator’s existing process becomes the specification. When the design gets that right, trust follows from day one.
FAQs
Why do energy operators resist AI recommendations?
Utility operators do not resist AI because of technophobia. They resist recommendations they cannot verify, from systems that do not operate by the same rules they do. The organizations making the most progress treat operator skepticism as a design requirement rather than a change management problem.
How does design earn operator trust in energy AI?
Progressive trust architecture is the design approach of starting AI deployment at narrow, verifiable decision categories, building a track record utility operators can see and challenge, and earning expanded scope based on demonstrated accuracy rather than elapsed time or training hours. It treats operator confidence as something AI must demonstrate, not something organizations must develop.
How do we implement AI in NERC CIP-regulated control room environments?
NERC CIP compliance and energy AI operator trust are co-dependent in utility control room environments. AI systems that make their reasoning visible, support human override, and maintain full audit trails satisfy both requirements simultaneously. The design approach that earns operator trust in control room operations is also the one that meets regulatory expectations for human control over safety-critical decisions.
How do you design AI that energy operators will actually use?
The most consistently successful approach is designing AI around existing operator workflows rather than alongside them. That means incorporating the actual rules, constraints, and judgment criteria operators use, making AI reasoning visible in terms operators can evaluate and challenge, and starting with decisions where the AI can build a verifiable track record before expanding its scope.
What is the connection between operator trust and AI ROI in energy?
They are the same thing. A utility operator who does not trust an AI recommendation will not act on it. An organization that cannot get operators to act on AI recommendations cannot demonstrate business value. Gartner projects more than 40 percent of agentic AI projects will be canceled by end of 2027. Inadequate risk controls is one of the primary causes, and in energy operations, risk control and operator trust are inseparable.
How do we capture retiring operator knowledge before it is lost?
AI systems designed to learn from every operator interaction are uniquely positioned to capture institutional knowledge from experienced utility operators. Each acceptance, override, and correction the system receives from a senior operator encodes judgment that would otherwise retire with that person. Organizations that deploy AI before their most experienced operators leave are building a knowledge base that survives the workforce transition.
Is operator trust in AI a technology problem or a leadership problem?
It is a design problem, which makes it a leadership problem. Technology teams will build what they are asked to build. If they are asked to minimize operator friction rather than earn operator trust, that is what gets built. The framing of the requirement determines the outcome. Energy leaders who put operator trust into the design specification rather than the change management plan get fundamentally different results.
Part 3 – The Fault Line: The Energy AI Architecture Decision That Outlasts Every Tool
Scott Young
This three-part series examines the three organizational failures that keep energy AI in perpetual pilot and what leaders who have moved past them did differently. Each article stands alone. The full series is the argument.
The energy AI market offers no shortage of compelling grid modernization use cases, from predictive maintenance and load forecasting to DER orchestration and outage detection. Every one of them is real, proven, and deployable today.
None of them, taken individually, produces the result energy executives are actually trying to achieve.
What every grid modernization strategy is ultimately pointed toward is generative and agentic AI that gets smarter over time and compounds advantage across the organization. What most energy organizations are building is a collection of agentic AI tools that cannot learn from each other. The distinction between those two outcomes is the energy AI architecture gap, and no one selling AI tools has a financial incentive to close it.
The Fragmentation Consequence
Forrester’s 2026 predictions report projects that vendor fragmentation will force the majority of enterprises to compose what the firm calls agentlakes. These are composable architectures designed to manage and orchestrate fractured AI deployments that individual teams built without a shared foundation. That is not a forecast about a future problem. It describes what most energy organizations are constructing right now, one use case at a time.
An IDC and Amazon Web Services (AWS) study surveying more than 900 organizations found that 50 percent have deployed ten or more AI agents. Fewer than 7 percent have reached full production with even one use case. The math tells a clear story about AI scalability in energy: most organizations have more AI tools in flight than AI value to show for it. The agents are accumulating. The intelligence stays flat.
Gartner warns that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing unclear business value and inadequate risk controls as the primary causes. In most of these cases, the tools performed as designed. The AI architecture that would have allowed them to compound never existed.
What Energy AI Architecture-First Implementation Means
Architecture-first is not a technology preference. It is a design discipline that asks a different question before any tool is selected, any use case is prioritized, or any pilot is launched.
Most organizations start by asking what an AI system should do. The organizations achieving compounding AI advantage in energy start by asking what an AI system needs to know in order to get smarter every time it operates.
Those two starting questions lead to fundamentally different implementations. The first produces a tool. The second produces a learning system.
An AI tool solves a discrete problem and stays there. An AI architecture connects solutions so that each one makes the next smarter. The difference determines whether AI investment compounds into enterprise advantage or accumulates into enterprise cost.
In energy operations, this distinction matters because reliability planning, DER coordination, and asset investment prioritization are all continuous processes that should improve with every cycle of real operational data they touch.
The design discipline connects operational data across assets, decisions, and time so that every deployment makes the next one faster, smarter, and more valuable. It treats generative and agentic AI as an organizational capability that compounds with use, not a collection of tools to be procured.
The Four Layers Energy Organizations Skip
At Robots & Pencils, we work from a four-layer energy AI architecture framework that has emerged consistently across production-scale deployment research and our own engagement experience. It is the architecture that turns agentic AI into enterprise infrastructure, the kind that acts, learns, and coordinates across the organization rather than operating in isolation. Most energy organizations invest heavily in two of the four layers and skip the other two. That sequencing error is the primary reason AI teammates fail to become intelligent infrastructure.
The Business Context Layer is where operational data becomes institutional memory. SCADA signals, historian databases, market feeds, maintenance records, and workforce systems need not be consolidated in one place. They need to be unified in shared meaning, so that AI agents across every layer of the organization operate from the same understanding of what the data represents and what decisions it should inform. Connecting these data layers does not require opening OT environments or replacing existing control systems. The OT-IT integration approach that unifies shared meaning operates within current security boundaries and NERC CIP frameworks, making it compatible with even the most sensitive operational technology environments.
The Agent Execution Layer is where AI teammates perform the real work of forecasting, optimization, anomaly detection, and dispatch routing. These are agentic systems that act on data, coordinate across workflows, and improve through every operational cycle. Most energy organizations invest here first and most heavily. Without the Business Context Layer underneath, every AI teammate operates on local data with local context, unable to learn from what agents in adjacent systems are seeing or doing. The result is precisely what most energy AI programs produce: isolated wins that do not reinforce each other.
The Evaluation and Optimization Layer is where AI systems improve through operational feedback. Digital twins, physics-informed models, and continuous calibration convert operational experience into model intelligence. This is the layer that turns a static deployment into a learning system. It is also the layer most frequently absent from energy AI implementations, because it requires the first two layers to be functioning before it can deliver its value.
The Apps Layer is where utility operators interact with AI through conversational interfaces, dashboards, and decision-support tools that surface AI intelligence in human terms. This is often where energy organizations begin, because it is the most visible and the most straightforward to demonstrate. Starting here without the layers beneath it produces AI that surfaces recommendations operators cannot verify and cannot trust.
The DOE’s Genesis Mission, which mobilized $293 million to advance AI for grid operations, is structured specifically around the integration layer. Its primary working groups address data integration standards, shared computational infrastructure, and cross-system interoperability rather than individual use cases. The federal government’s most significant AI-for-energy investment is funding the architecture that makes use cases compound, not the use cases themselves.
What Compounding Looks Like at Scale
ERCOT created a dedicated Enterprise Data and AI organization in January 2026. Rather than establishing an AI team or center of excellence, ERCOT created an enterprise function that treats AI as organizational infrastructure rather than a departmental capability. That organizational move signals a shift from ad hoc AI experimentation to systematic, enterprise-wide architecture. ERCOT is building the foundation, not accumulating the tools.
The economics of getting this right at scale are significant. The Department of Energy (DOE) projects that virtual power plant (VPP) deployment at scale could reduce overall grid costs by $10 billion per year by redirecting spending from peaker plants to participants. Separately, DOE analysis projects that VPP deployment could avoid $17 billion in annual power sector expenditure by displacing new generation build-out. VPPs already provide peaking capacity at roughly 40 to 60 percent lower cost than conventional alternatives. NREL’s Autonomous Energy Systems program is designed to manage hundreds of millions of distributed energy resources through reinforcement learning and distributed decision-making. None of these outcomes are achievable with a collection of point solutions. They require AI that can coordinate across assets, learn from aggregated behavior, and improve through every dispatch cycle.
The same principle holds at the operational level. When workforce scheduling data, dispatch rules, real-time outage events, and multi-channel delivery connect into a single intelligent workflow, no individual component produces the result. The value lives in the connections between layers, not in any single tool operating independently.
The Energy AI Architecture Question to Ask Before the Next Vendor Call
The energy AI market will continue producing use cases, point solutions, and vendors faster than any organization can evaluate them. That pressure does not ease.
For utilities operating on regulatory capital cycles of three to five years, this matters more than it does in almost any other sector. The cost of the wrong architectural decision is not one quarter. It compounds across the next rate case.
What energy leaders can change is the question they ask before any solution enters their environment. Not whether a tool solves a problem they have. Whether adding that capability makes the rest of their AI smarter, or adds another isolated system their organization has to manage separately forever.
That question is harder to answer and slower to commercialize, which is why most vendors will not help energy leaders ask it. The answer might be that their tool does not belong in your architecture yet, or that it belongs in a different layer than the one they are selling it for.
This design discipline is not a product category. The organizations that adopt it as a discipline rather than a procurement checklist are the ones that will look back in five years and understand why the gap between them and their competitors only widened. The tools they deployed got smarter with every cycle. The tools their competitors deployed stayed exactly where they started.
The right partner makes progress inevitable. Robots & Pencils builds the four-layer architecture that connects your operational data, earns operator trust, and compounds intelligence across your energy business. Request an AI Briefing and find out what AI teammates live inside your operations look like.
About the Author
Scott Young is EVP of Growth and Strategic Alliances at Robots & Pencils, where he works with energy executives to move from decision to live. Connect with Scott on LinkedIn.
Key Takeaways
Fifty percent of organizations have deployed ten or more AI agents. Fewer than 7 percent have gone live with even one use case. The agents are accumulating. The intelligence stays flat. The gap is an architecture gap, not a technology gap.
The design discipline that closes the architecture gap connects operational data across assets, decisions, and time. An AI tool solves one problem. An AI architecture makes every deployment smarter than the last.
The four layers separating production-scale AI from perpetual pilots: Business Context, Agent Execution, Evaluation and Optimization, and Apps. Most energy organizations invest in layers two and four while skipping one and three. That sequencing error is why isolated wins never compound into enterprise advantage.
The DOE’s Genesis Mission mobilized $293 million to advance AI for grid operations. Its primary working groups address data integration standards and cross-system interoperability. The federal government’s most significant AI-for-energy investment is funding the architecture, not the tools.
The value in an energy AI deployment lives in the connections between layers, not in any individual component. Workforce data, dispatch rules, real-time outage events, and operational feedback need to connect into a single intelligent workflow before the architecture compounds. No individual tool produces that result on its own.
FAQs
What separates an AI architecture from a collection of AI tools?
An architecture connects operational data across assets, decisions, and time so that every deployment makes the next one faster, smarter, and more valuable. A tool solves a discrete problem and stays there. The distinction determines whether AI investment compounds into enterprise advantage or accumulates into enterprise cost.
What are the four layers of energy AI architecture?
The four layers of the Robots & Pencils energy AI architecture framework are:
Business Context Layer: operational data unified into institutional memory across SCADA, historian, market, maintenance, and workforce systems
Agent Execution Layer: AI teammates performing forecasting, optimization, anomaly detection, and dispatch routing
Evaluation and Optimization Layer: continuous improvement through digital twins, physics-informed models, and operational feedback loops
Apps Layer: conversational interfaces, dashboards, and decision-support tools for utility operators
Most energy organizations invest in the Agent Execution and Apps layers while underinvesting in the Business Context and Evaluation layers. This is the primary reason AI wins remain isolated rather than compounding into enterprise advantage.
What is the difference between an AI center of excellence and an enterprise AI function for utilities?
A center of excellence is a capability hub that individual teams draw from on request. An enterprise AI function treats AI as infrastructure that the entire organization runs on. ERCOT’s decision to create a dedicated Enterprise Data and AI organization in January 2026 reflects the latter model. The organizational distinction matters because enterprise infrastructure receives the investment, governance, and architectural discipline that shared service centers rarely sustain at scale.
Why do energy AI tools fail to compound into enterprise advantage?
Tools fail to compound when they are deployed without the architectural foundation that would allow them to share context and learn from each other. A predictive maintenance system that cannot access outage history cannot improve its predictions based on failure patterns across the fleet. A load forecasting system that cannot connect to DER dispatch cannot refine its models based on how demand response actually performed. Compounding requires connection, and connection requires architecture.
How does the DOE Genesis Mission inform energy AI architecture decisions?
The Genesis Mission is structured around data integration standards, shared infrastructure, and cross-system interoperability rather than individual use case development. Energy leaders can interpret this as a clear signal: the federal government’s most authoritative AI-for-energy initiative concluded that integration architecture is the primary bottleneck, not model capability. Organizations building their AI strategy around individual use cases are solving a second-order problem.
How do we evaluate whether our current AI architecture is designed to compound?
Ask three questions. First: can AI agents in different parts of the organization access and act on the same operational data with the same shared meaning? Second: does each AI deployment improve in accuracy and value over time based on operational feedback, or does it perform at the same level it was trained to? Third: when a new AI use case is deployed, does it make existing systems smarter, or does it operate in isolation? If the answer to any of these is no, the architecture is not designed to compound.
What should energy leaders ask vendors before selecting an AI solution?
Ask how this solution connects to the operational data the organization already has, how it shares learning with other AI systems in the environment, and which of the four architectural layers it operates in. If a vendor cannot answer the second question, their solution is a tool rather than an architectural component. That does not make it wrong to buy, but it does mean the organization needs to understand which layer it belongs in and what foundation needs to be in place before it will deliver compounding value.
Robots & Pencils Expands Retail and Consumer Goods Leadership with Appointment of Saul Delage as Client Partner
Robots & Pencils
As AI reshapes how Retail and Consumer Goods businesses compete, Robots & Pencils plants its flag in the vertical and brings in a 30-year industry veteran to lead the charge.
Robots & Pencils, an applied AI engineering partner known for high-velocity delivery and measurable business outcomes, today announced the appointment of Saul Delage as SVP, Client Partner, Retail and Consumer Goods (RCG). Based in Chicago, he brings 30 years of experience building executive partnerships and driving growth across some of the most recognized names in digital transformation. Delage joins a proven leadership team at Robots & Pencils with extensive experience delivering for more than 100 of the world’s most recognized consumer brands, including dozens of Fortune 500 companies.
The appointment is a deliberate move. Robots & Pencils is investing with intention in Retail and Consumer Goods, including CPG, eCommerce, Restaurants & Everyday Essentials, and Retail, an industry under mounting pressure to move from AI experimentation into generative and agentic AI that performs in production and delivers measurable business outcomes. Delage will lead client relationships across the vertical, helping enterprise leaders get more from their AI investments and more from their investments in AWS, while strengthening the company’s presence in key markets and building closer, more embedded partnerships with clients.
The Right Leader for the Moment
“Saul is the kind of leader clients trust before the contract is signed and can’t imagine working without after,” said Len Pagon, CEO of Robots & Pencils. “He has worked alongside some of our team before. He knows how we operate, and he knows this vertical inside out. Retail and Consumer Goods is a large growing industry vertical for us, and we went out and got the right leader.”
A Career Built on Trust and Delivery
Delage arrives with a career forged across Cognizant, Isobar, Havas, Razorfish, and Fry, where he built high-performing growth teams, secured long-term relationships with Fortune 500 companies, and earned a reputation as one of the most technically fluent executives in the industry, equally effective in the boardroom and in the details of delivery.
His hire reflects the company tenet that winning in Retail and Consumer Goods requires leadership with deep business experience, the technical fluency to speak the language of AI, and the delivery discipline to back it up.
“Retail and Consumer Goods businesses have made significant investments in AI, and too many have little to show for it in production,” said Delage. “Robots & Pencils builds enterprise AI systems — generative, agentic, and production-ready — that move fast and tie directly to revenue and customer experience. That is exactly what this industry needs right now, and Robots & Pencils is built to deliver it.”
Built on AWS. Driven by Outcomes.
Robots & Pencils is unabashedly aligned with AWS and is building its Retail and Consumer Goods vertical around that conviction. The goal is helping enterprise leaders drive measurable business outcomes on AWS, from first deployment to full-scale production. For AWS co-sell teams and enterprise leaders who need a partner that moves fast and delivers, that commitment is the differentiator.
Robots & Pencils Goes All in on AWS with Appointment of Adrian Bird as Vice President of AWS Partnership
Robots & Pencils
Bird brings two decades of alliance leadership, including five years inside AWS, to accelerate co-sell engagement and expand enterprise AI delivery with AWS
Robots & Pencils, an applied AI engineering partner known for high-velocity delivery and measurable business outcomes, today announced the appointment of Adrian Bird as Vice President of AWS Partnership.
The timing is deliberate. Bird joins as the company deepens its investment in the AWS ecosystem. He will lead the company’s AWS Partner strategy and execution, expanding joint customer engagement, and strengthening alignment with AWS teams.
Two Decades of AWS and IBM Partnership Leadership
Bird brings direct experience from AWS, where he was Partner Sales Leader from 2020 to 2026. In that role, he managed comprehensive channel strategy and partner program initiatives across ISVs, global systems integrators, and technology partners, collaborating with AWS field sellers across regions and industries to support joint go-to-market initiatives. He built partner success frameworks that drove exceptional year-over-year growth in partner revenue, created revenue operations tools adopted across multiple AWS business units, and led initiatives that significantly expanded the security partner ecosystem. He earned AWS’s Innovation All-Star recognition in 2024.
Prior to AWS, Bird spent fifteen years at IBM in progressively senior partnership and alliance roles. He grew the Watson Media worldwide partner channel from inception to more than a third of business unit revenue within two years, scaled the IBM Commerce partner business several-fold over four years, and led the integration of Sterling Commerce’s partner ecosystem following its acquisition, retaining the vast majority of partners while substantially growing combined revenue. He is a recipient of IBM’s Industry Solutions Successful Partnering Award and the IBM 100% Club.
Accelerating AWS Partnership Strategy
“Adrian has spent his career building partner ecosystems that generate real, compounding results,” said Scott Young, EVP of Growth and Strategic Partnerships. “Having led partner strategy inside AWS, he knows exactly how AWS field teams operate and what it takes to be a partner they actively bring into deals. That perspective, combined with his track record of execution, is precisely what we need right now. Clients who need enterprise AI at speed will benefit directly from what Adrian builds.”
“We are unabashedly all in on AWS,” said Len Pagon, CEO of Robots & Pencils. “We have tremendous traction and momentum. Adrian is another key investment in taking our AWS partnership further, faster.”
Driving Enterprise AI Adoption and AWS Consumption at Scale
Bird’s appointment builds on recent company milestones including earning AWS Advanced Tier Services Partner status, selection as one of 11 inaugural AWS Pattern Partners globally, and the launch of its Studio for Generative and Agentic AI in Bellevue near AWS headquarters. The company is also actively engaged with AWS through its collaboration with the AWS Generative AI Innovation Center to support joint enterprise initiatives. Bird’s mandate is clear: align partnershipstrategy with Robots & Pencils’ ability to deploy AI into production at speed and drive measurable AWS consumption and customer value at scale.
“Robots & Pencils has built what most companies only claim to have. They have the engineering depth, a proven record of deploying AI into production at speed, and the scale to serve enterprise clients globally,” said Bird. “The AWS partnership is the force multiplier that connects those capabilities to the clients who need them most. My job is to make sure that potential becomes performance, for Robots & Pencils, for AWS, and for the clients we serve together.”
Request an AI briefing to evaluate how applied AI can deliver velocity and impact within your organization.
Robots & Pencils Brings Enterprise AI Platform Expertise to ASU-GSV Summit Panel
Robots & Pencils
CEO Leonard Pagon joins education industry discussion exploring how leading universities move from AI pilots to enterprise-scale platforms with governance and speed
Robots & Pencils, an applied AI engineering partner known for high-velocity delivery and measurable outcomes in complex institutional environments, today announced that its CEO, Leonard Pagon, will participate in a panel discussion at the ASU-GSV Summit, a premier global event focused on the intersection of education, technology, and workforce innovation.
As a seasoned operator who has built and scaled technology-driven companies, Pagon will share practical insights on how universities are operationalizing AI across institutional systems while maintaining governance, security, and accessibility.
Pagon will be joined by Kyle Bowen, Deputy CIO, Arizona State University; Matthew Gee, Director, U.S. Program Data, Gates Foundation; Stephanie Khurana, CEO, Axim Collaborative; and Elizabeth Reilley, Chief AI Officer, University of North Carolina. The panel will be moderated by Lev Gonick, Enterprise CIO, Arizona State University.
The session centers on the real-world mechanics of enterprise AI in higher education, including:
How institutions connect AI across SIS, LMS, CRM, and operational systems
Governance models that allow rapid experimentation without losing institutional control
Real examples that enable non-technical users to design and deploy AI solutions
The growing role of agentic AI in strengthening institutional resilience and operational efficiency
“A lot of AI activity in higher education is happening at the use case level, which creates fragmentation across the institution.” said Pagon. “The real shift is building platforms that bring those efforts together and make AI work inside everyday institutional constraints. The impact shows up when AI moves from pilots into production systems that perform, with the right guardrails in place.”
Scaling AI Across the University with Governance and Control
Robots & Pencils has worked with with Arizona State University since 2019 on cloud-native architecture, platform modernization, and applied AI initiatives that improve the student experience and strengthen institutional operations. This collaboration has produced practical insight into how universities move from small pilots to enterprise AI programs through platforms that support rapid innovation while maintaining strong governance, security, and accessibility standards.
From Experimentation to Enterprise AI
Institutions are increasingly looking to AI to improve student support, reduce manual processes, and extend staff capacity without increasing headcount. The ASU-GSV Summit brings together entrepreneurs, investors, and education leaders who are shaping the future of learning and workforce development through technology.
Higher education leaders are moving quickly to modernize operations, improve student experience, and prepare graduates for an AI-driven economy. Enterprise AI platforms give universities a structured way to design, deploy, and govern AI solutions across departments while enabling faculty, staff, and researchers to use AI in their day-to-day work with security, compliance, and accessibility built in. This approach enables institutions to scale AI in higher education with clarity, control, and speed.
Why higher education AI governance frameworks fail after approval and who is responsible for closing the gap.
Across higher education, AI is no longer theoretical. It shows up in advising offices, finance teams, registrar systems, and IT backlogs every day. Not long ago, the conversations felt divisive. Leaders debated risk, approved tools, and moved forward with cautious optimism.
Today, many of those same leaders are sitting with a different feeling. The systems technically work. Progress feels uneven. Accountability feels scattered. And no one can say with certainty whether the institution is truly advancing or simply carrying new technology without a clear owner of the outcome.
That uncertainty now lives with presidents, provosts, and CIOs expected to defend AI investment, manage institutional risk, and show results inside universities designed to move carefully, by consensus, and without urgency. The technology is working. The institution is not.
The gap between those two facts is structural.
Today, Robots & Pencils, an applied AI engineering partner known for high-velocity delivery and measurable outcomes in complex institutional environments, announces the release of The Institutional Intelligence Crisis, a three-part research series examining why AI adoption fails at the departmental level and what senior leadership must address to change that trajectory.
Drawing on research and operational experience across universities and complex organizations where AI adoption is already underway, the series identifies a set of recurring patterns that appear once AI moves beyond experimentation and into daily operations.
The series is authored by Jess Martin, Principal Delivery Manager at Robots & Pencils, and is written for university presidents, provosts, CIOs, and boards of trustees. It treats AI adoption as an institutional design challenge, not a technology procurement problem, and focuses on the post-pilot phase: the period where accountability structures and human dynamics determine whether AI becomes a reliable capability or quietly rots.
“AI doesn’t create accountability problems.” says Martin. ”It exposes the ones you already have.”
Why AI Governance Fails in Higher Education: Three Failures That Compound
The series is built around three failures that compound in sequence:
The Intelligence Leak (Part 1): When institutions fail to provide operational pathways for AI, high-performing staff build their own, exporting institutional problem-solving logic to personal accounts at third-party vendors. The sector now calls this Shadow AI. The university does not get smarter. The vendor does. When institutions leave a gap between policy and practical access to AI tools, staff close that gap themselves, often outside the visibility of supervisors or institutional governance.
The Redistribution of Expertise (Part 2): AI makes institutional expertise portable. The specialized knowledge that senior staff in advising centers, registrar offices, and financial aid departments have spent decades accumulating and making indispensable can now be replicated by a junior colleague and a well-prompted AI. What leadership often experiences as operational friction is frequently a rational response from professionals whose expertise has defined their role inside the institution.
The Brittle System (Part 3): When no one is accountable for output quality, performance degrades without announcement. Errors become plausible enough that staff quietly work around them rather than report them. The system continues running while confidence in the results quietly erodes. In many institutions, leaders lack a clear view into whether AI systems are improving outcomes or introducing new operational risk.
Higher education leaders are encouraged to read the full series and engage with a data-driven perspective grounded in accountability, execution, and institutional readiness.
Part 3 – The Institutional Intelligence Crisis: The Brittle System
Jessica Martin
This article is part of a three-part series examining why AI adoption stalls in higher education and what senior leaders must address to restore momentum. Each article stands alone. Reading the full series is recommended.
Execution, Quality Drift, and the Cost of Looking Away
For four months, Donna’s enrollment verification tool looked flawless.
As Registrar, she oversaw the deployment, ran the tests, and watched it process thousands of student records without a single flag.
Then an IT team upstream changed how transfer credits were coded as part of a routine update, and the change never surfaced in any channel that reached Donna’s office or AI.
The tool did not fail loudly. It started producing plausible errors, correctly verifying about 90% of students while quietly mishandling a subset of transfer students. With no performance owner assigned to audit for drift, the errors went unnoticed for weeks.
When a student finally flagged the discrepancy, Donna’s staff investigated, found the issue quickly, and stopped trusting the tool.
They left it running, but they also rechecked every single verification by hand. The institution now pays for the AI license, and the full manual workload it was meant to reduce.
This is the AI ROI problem in higher education: a tool that looked like it was working, until someone finally checked.
Why AI Systems Fail After Launch: The Day-Two Problem
Across higher education deployments, Robots & Pencils has consistently observed that the most dangerous phase of AI adoption arrives about six months after launch. Initial energy fades, the project team moves on, and the tool is left in day-to-day operations without a named owner, a monitoring protocol, or a working feedback loop.
Without clear accountability, quality drifts as vendors ship updates, prompts that worked in September fail in February, and upstream data formats change. If nobody owns day-two oversight, those issues accumulate quietly until trust collapses and staff begin working around the tool.
Most institutions are not measuring whether AI is actually paying off. Kiteworks and EDUCAUSE report that only 13% are tracking ROI for AI investments, which leaves the rest funding tools that look like progress on a dashboard without delivering sustained value. The EDT Partners AI Impact Study (2026) found that only 2% of institutions have secured new funding specifically for AI projects, with 30% having no cost accommodation plan at all. When AI is funded by redirecting existing budgets rather than new investment, accountability disappears along with the original budget line.
Algorithmic Bureaucracy vs. Human Bureaucracy
Higher education runs on human bureaucracy. It is slow and imperfect, but it can flex around messy reality: a registrar notices an unusual student situation, applies context, and makes an accountable exception.
Algorithmic bureaucracy trades that flexibility for speed. It is brittle, and when it breaks, it often does so quietly, producing outputs that look compliant until someone checks the edge cases.
THE CRITICAL DISTINCTION
We used to trust “The System” because we trusted the people running it. Now we are being asked to trust the math. That shift requires a different kind of institutional accountability than higher education has built before.
When an AI hallucinates compliance, it delivers a wrong answer with the confidence of a policy manual, with no hedging, no uncertainty, and no indication that something may have gone wrong. A slow human bureaucracy fails loudly and individually. An algorithmic one fails quietly and at scale. Without someone specifically tasked with auditing for brittleness, the system will eventually fail in ways a slow human bureaucracy never would.
How Unchecked AI Trust Becomes Institutional Liability
The Donna incident is not an edge case. It reflects a documented pattern of how AI trust degrades in operational environments.
Policy Awareness Without Confidence of higher education staff are aware of their institution’s AI policies. Of those, only half feel confident using AI tools for work. Having a policy on paper is not the same as having a workforce that trusts it.
Zombie Pilots of institutions are measuring the ROI of their AI investments. The rest are operating on assumption.
In higher education, the 66% non-validation rate reported by KPMG (2025) matters because the consequences are real. A wrong degree audit recommendation can delay graduation, a miscoded financial aid calculation can trigger federal compliance issues, and an enrollment verification error can ripple into accreditation reporting. That pattern of unchecked trust, at that scale, creates genuine institutional liability. This is how adoption degrades in practice: the tool stays “Active” on a dashboard while staff quietly stop believing it and rebuild manual checks around it.
Defining Acceptable Variance
Sustainable AI impact requires honesty about what the technology is. AI will not be 100% accurate, so the institutions that get value out of it define acceptable variance up front and are explicit about which tasks can tolerate errors and which cannot.
MDPI (2026) found that AI achieves higher scoring consistency than humans in 66% of assessment cases, but 50% of those systems fail the Transparency Test and do not adequately disclose how the decision was reached. Consistency without transparency is hard to trust.
Defining acceptable variance before deployment is an ethics and accountability decision that belongs with academic leadership, not an IT implementation detail, and if that conversation hasn’t happened, the institution isn’t ready to deploy.
The Four Requirements for Durable Adoption
Many higher-ed institutions measure the wrong things, like licenses assigned, daily active users, or how much text was generated. Those are activity metrics, and they say nothing about trust, accuracy, or whether the work is actually improving. Across higher education deployments, Robots & Pencils has found that the difference between AI that compounds value and AI that quietly degrades is rarely the technology. It is whether someone is named, empowered, and evaluated on what happens after launch.
The institutions modeling this well are not the ones that moved fastest. Stanford, MIT, Harvard, UC Berkeley, and Arizona State have each implemented named governance structures – ethics boards, oversight committees, regular audits – that make accountability visible and operational. The technology at those institutions is not meaningfully different from what is available to everyone else. The governance surrounding it is.
Four conditions have to be present for AI to move from perpetual pilot to institutional infrastructure. Institutions that are missing any one of them will recognize the Accountability Vacuum opening again. Together they form the core of a durable AI governance framework for universities serious about moving from experimentation to operational accountability.
Named Accountability: Every AI implementation needs a single person accountable for output quality, with that responsibility reflected in their role expectations and evaluation.
Feedback Cycles: Staff need a direct way to flag bad output and see corrections made. If the reporting path requires navigating a ticketing system, it won’t get used, and errors will accumulate quietly until they surface somewhere you can’t ignore them.
Operational Integration: AI review has to be built into the normal rhythm of departmental operations, with a standing owner and outcome metrics that get reported alongside everything else the department is accountable for.
Radical Transparency: Leadership must be honest about what the AI can and cannot do. Pretending an AI is 100% accurate destroys trust the moment the first error appears, and there will be one. The institutions that survive it are the ones that already told their staff it was coming.
The Accountability Vacuum: A Final Word
Every institution in this series was present at launch and absent when consequences arrived. That gap is where institutional credibility is won or lost.
Marcus used a personal AI account because the sanctioned process could not meet the deadline he was given. Diane stepped in because the institution gave her a directive and none of the infrastructure to do it. Raymond configured rules that reflected his judgment; the institution never validated them against policy. Donna stopped trusting the tool because no one was responsible for watching it once it was in production.
Your registrar’s office, advising teams, and financial aid staff have already formed an opinion about whether AI is part of the institution’s operating model or simply a pilot being performed for leadership. Those judgments will settle based on what happens after deployment.
The technology is not the test. Leadership is.
Punch List: Dismantling the Brittle System
#
Action
Owner / Timeline
1
Establish a Drift Audit: Quarterly, test 50 random AI outputs against a human expert’s assessment. Make the results visible to the Output Owner and their department head. If the error rate is rising, that is your early warning system.
AI Output Owner + QA – quarterly
2
Define Variance Tolerance Before Launch: State the Error Budget for every AI-involved task in writing before deployment. If you cannot define what an acceptable error rate looks like, you are not ready to deploy.
Provost Office + Deans – before launch
3
Link IT Change Logs to AI Owners: A database schema change or upstream system update should trigger an immediate AI review cycle; not a help ticket six weeks later. Build this handoff into standing IT protocol.
CIO – standing protocol
4
Create a One-Click Error Channel: Staff need a frictionless way to flag wrong AI output. If it requires navigating a ticketing system, they will not use it. They will use their judgment and quietly work around the tool instead.
IT + Department Heads – within 30 days
5
Report on Outcomes, Not Activity: Replace license counts and login metrics with error rates, exception handling time, and decision consistency scores. These tell you whether AI is working or degrading quietly.
Leadership – next reporting cycle
The pace of AI change can feel relentless with tools, processes, and practices evolving almost weekly. We help organizations navigate this landscape with clarity, balancing experimentation with governance, and turning AI’s potential into practical, measurable outcomes. If you’re looking to explore how AI can work inside your organization—not just in theory, but in practice—we’d love to be a partner in that journey. Request an AI briefing.
Key Takeaways
Most AI failures occur after deployment, not during launch. The greatest risk emerges months after implementation when oversight fades, ownership is unclear, and systems begin drifting as prompts, data sources, or upstream systems change.
Unchecked AI systems degrade quietly rather than failing visibly. Algorithmic systems often produce plausible but incorrect results that go unnoticed until someone manually verifies them, eroding trust and forcing staff to rebuild manual checks around the tool.
Many institutions measure AI activity instead of outcomes. Metrics such as logins or licenses assigned create the appearance of progress, yet few institutions measure ROI, accuracy, or operational impact, allowing ineffective tools to remain in place.
Unchecked trust in AI outputs creates institutional risk. When staff rely on AI responses without validation, incorrect outputs can propagate into compliance decisions, academic records, and student services at scale.
Durable AI adoption requires operational governance after launch. Sustainable impact depends on four conditions: named accountability, continuous feedback cycles, integration into normal operations, and transparent communication about AI’s limitations.
Frequently Asked Questions
1. Why do AI systems often fail months after deployment? Many institutions treat deployment as the finish line. Without ongoing monitoring, ownership, and feedback loops, changes in data sources, system updates, or prompts can quietly degrade output quality over time.
2. What is the “Day-Two Problem” in AI adoption? The Day-Two Problem describes what happens after the launch phase ends. When project teams move on and no operational owner is assigned, AI systems drift in quality and gradually lose staff trust.
3. Why is algorithmic bureaucracy more fragile than human bureaucracy? Human systems can adapt to unusual situations through judgment and context. Algorithmic systems prioritize consistency and speed, which makes them vulnerable to silent errors when conditions change.
4. How should institutions measure AI performance? Instead of focusing on activity metrics such as usage or logins, institutions should track outcomes such as accuracy rates, exception handling time, and decision consistency.
5. What governance practices help prevent AI quality drift? Organizations can reduce risk by assigning a clear output owner, defining acceptable error thresholds before deployment, creating easy error-reporting channels, and running regular audits of AI outputs.
Part 2 – The Institutional Intelligence Crisis: The Redistribution of Expertise
Jessica Martin
This article is part of a three-part series examining why AI adoption stalls in higher education and what senior leaders must address to restore momentum. Each article stands alone. Reading the full series is recommended.
Professional Identity, Resistance, and the Power Shift AI Creates
Diane didn’t wait for permission. She couldn’t afford to.
As Director of Advising, she had been told her department needed to find ways to absorb the impact of recent turnover, and leadership had suggested AI as a possible direction. But no tool was specified, and no governance ever reached her desk.
So her staff did what capable people do in a vacuum: they improvised. They started using a free LLM to draft student appointment summaries. It worked well until staff began uploading degree audits and academic plans. Diane recognized the security risk immediately. She also recognized that the university’s official AI policy, still in draft, was not going to arrive in time to help her.
Tired of waiting, Diane used her weekend to write her own policy. It was just one page and defined what could be uploaded, what required a human to double-check, and who to call if the tool produced an error.
Six months later, her document was the actual operating standard for her advising staff. The university’s 37-page policy was still sitting with the committee in a draft.
This is what AI adoption in higher education looks like when governance does not keep pace with operational reality.
It is easy to view Diane’s initiative as a simple win: a director filling a leadership vacuum to keep her department safe. However, while Diane was solving the governance problem, her senior staff were reacting to a different reality. These advisors began quietly slowing the AI pilot to a crawl, not because they wanted to be difficult, but because the tool threatened the value of the specialized expertise they had built over decades. Without an alternative role that turned them into ‘architects’ of the system, they protected their professional value by highlighting each and every edge case or exception the AI couldn’t handle, ensuring the tool remained too ‘risky’ to operate without them.
Why Adoption Fails at the Departmental Level
AI adoption in higher education rarely dies in a boardroom. It dies in the registrar’s office, the advising center, the financial aid office. Pilots are almost never formally rejected. They simply fade.
When a tool is introduced without a clear redesign of the workflow around it, usage becomes uneven. A few early adopters find value. The rest quietly route around the tool. By the time leadership reviews usage metrics, the adoption is a ghost. Logins may be high because of a mandate, but actual impact on daily work is negligible.
Leadership often describes this as change management friction or fear of technology. That is the wrong diagnosis. And the wrong diagnosis produces the wrong response.
Research from Frontiers in Education (2025) found that concern about AI ethics does not reliably predict whether faculty actually engage with AI tools, largely because most lack the means to critically evaluate AI-generated outputs. When people cannot assess whether the AI is right, avoidance is the rational response. They have not been given a reason to trust it.
That trust deficit plays out differently depending on where someone sits in the institution. For faculty, it is an epistemological problem. For senior administrative staff, it is an existential one.
When AI Expertise Becomes a Threat to Professional Identity
In many universities, power is held by those who know the rules: the exceptions, the workarounds, the edge cases that never made it into the policy manual because the only person who fully understood them was the one who created them.
For decades, this has been the primary currency of administrative authority in higher education. The gatekeeper holds informal power precisely because what they know is scarce, undocumented, and difficult to transfer.
When a junior advisor with a well-prompted AI can navigate a complex academic plan as accurately as a 20-year veteran, the social architecture of the advising center doesn't adapt gradually. It loses its foundation.
The social architecture of the advising center depends on that scarcity, and the AI eliminates it. The World Economic Forum (2025) identifies this emerging class of displaced knowledge workers as the AI Precariat: staff facing chronic insecurity and identity loss as their specialized roles are undercut by automation.
The numbers are moving faster than most senior administrators realize. Sixty-six percent of enterprises are already reducing entry-level hiring specifically because of AI, and 42% of employers believe most entry-level white-collar positions could disappear within five years. Higher education’s administrative workforce sits precisely in the crosshairs of that projection. These are specialized, knowledge-intensive, relationship-dependent roles. They are not safe from this.
Expansion Expected, Anxiety Rising of higher education faculty and administrators expect institutional AI use to increase over the next two years. Concern about AI-related role elimination has doubled year over year.
The Quiet Saboteur: What No One Will Tell You
Your most resistant senior administrators are not afraid of the technology. They are afraid of what the technology reveals.
Consider what it means to spend nearly two decades becoming indispensable in an institution that moves slowly. The winning strategy centered on becoming the person who understood how the system actually worked. The colleague people called when policy language became unclear. The person whose judgment translated the written rules into workable decisions. Someone the institution had, without ever realizing, built workflows around.
This was not laziness or territoriality. It was how the institution rewarded people. Longevity plus accumulated knowledge equaled authority. And authority, in higher education’s flattened salary structures, was often the only real compensation available. Salary bands in administrative higher education are often tied to the complexity and specialization of the role. The registrar who knows the exceptions is classified and paid differently than the one who processes straightforward cases. Their compensation and title rest on the same premise: the knowledge they hold is scarce and difficult to transfer, and the institution depends on it.
Year three: you discover the workaround for the transfer credit edge case no one else knows.
Year seven: you are the person they call when something breaks.
Year twelve: your institutional memory earns you a seat in rooms your title was never meant to enter.
Year sixteen: you are the policy, in every practical sense that matters.
Year nineteen: a junior staff member sits down with an AI and gets the same answer you would have given, in four seconds.
For someone whose professional identity is built on being the expert in the room, that kind of displacement doesn’t register as a career setback. It lands as something closer to erasure.
And here is the detail that makes it genuinely uncomfortable: some of that legacy knowledge, when the AI replicates it, turns out not to have been sophisticated governance wisdom. Some of the exceptions being gatekept for two decades were never actually correct. They were just unchallenged, because only one person fully understood them, and that person had every incentive to keep it that way.
AI does not just democratize the knowledge. In some cases, it audits it. And that audit can be brutal for someone who built a career on being the authority.
Your senior staff is not going to say they are afraid the AI will make their hard-earned expertise look common. But that anxiety is real. When a tool can perform in seconds what a veteran staffer spent decades mastering, it creates a crisis of professional identity.
Pilots often stall because the people expected to run them are protecting a lifetime of professional equity. They are using the tools they have left to remain indispensable, pointing out every tiny policy exception and procedural hurdle that AI isn’t yet “trusted” to handle.
When Resistance Hides Inside the AI Configuration
Diane improvised in good faith. Raymond did something different.
Raymond had nineteen years in the registrar’s office. He knew the exception credit process the way a watchmaker knows a movement. Not just what the parts did, but why they were arranged the way they were, and what happened when someone who did not understand that arrangement tried to change it.
When the AI degree-audit pilot launched, Raymond was the obvious choice to help configure the exception rules. He was cooperative. He attended every implementation meeting. He flagged edge cases the vendor’s team had not considered. Leadership took his involvement as confirmation that senior staff were bought in.
Raymond configured the exception logic to route any non-standard credit scenario to a human reviewer before the AI could resolve it. Transfer credits. AP overrides. Co-enrollment arrangements. Prior learning assessments. These cases were complex, he explained. The AI could not be trusted with them yet. His threshold flagged 40% of all degree audits for manual review.
The actual institutional policy, had anyone cross-referenced it, required human review on roughly 8%.
Leadership looked at the dashboard and saw what they expected: high AI utilization, appropriate human oversight, senior staff engaged with the process.
What they were actually looking at was Raymond, rebuilt in code. He had not resisted the AI. He had become its gatekeeper. His queue was full. His expertise was indispensable. And because the configuration lived in a system only he fully understood, no one thought to ask whether the threshold was right. Only whether Raymond had approved it.
He had.
This is the version of resistance that never shows up in adoption metrics. Raymond’s department showed 100% AI utilization. His pilot was considered a success. The Accountability Vacuum does not always look like failure. Sometimes it looks exactly like what leadership hoped to see.
From Gatekeeper to Architect
There is an alternative to both of these outcomes, but it requires leadership to move first. In engagements across higher education, Robots & Pencils has found that the institutions making the fastest progress on AI adoption are not the ones with the most sophisticated tools or the strictest policies. They are the ones that looked at what their staff were doing outside the sanctioned path, treated it as data about where that path was failing, and gave their most experienced people a meaningful role in redesigning it. Unauthorized AI use tells you exactly what the institution has not yet solved. Banning the tool addresses the symptom while leaving the underlying need completely intact. The question is not whether your staff are using AI. They are. The question is whether the institution is learning anything from how.
The person who spent decades learning every exception, every workaround, every edge case that the student information system cannot handle: that person is not your AI problem. That person is your answer to it. They are the only one in the building who knows where the institutional logic actually lives.
The person who knows where the bodies are buried is the only person qualified to tell the AI where not to dig.
The difference between a registrar with nineteen years in the office who quietly rebuilds their gatekeeping function inside your AI pilot and one who becomes its most rigorous auditor is not temperament. It is whether the institution made them an offer worth accepting.
This is a genuine repositioning of professional value: moving from a knowledge holder to a knowledge architect. Rather than maintaining individual indispensability through daily tasks, the institution is asking them to make their expertise permanent by building it directly into the institutional framework.
That is a different kind of legacy. And for the right person, it is a more compelling one.
But the timing is critical. If the institution waits until AI has already rendered a role redundant to propose a new path, the offer will likely be perceived as an afterthought. In higher education, where titles change slowly and salary bands are narrow, seniority is one of the few available signals of institutional standing. The transition needs to be presented as a proactive investment in expertise, not a reactive attempt to find someone a new place.
The challenge for leadership is to redesign the reward system that has favored individual gatekeeping.
Shadow AI as a Diagnostic
If 70% of a department is using an unauthorized tool, that is not a discipline problem. It is a map of where the sanctioned path failed them. Reading that map honestly is how institutions move past the Accountability Vacuum. But getting staff onto the sanctioned path is only half the problem. What happens after they get there is where most institutions stop paying attention.
Punch List: Navigating the Power Shift
#
Action
Owner / Timeline
1
Redesign the expert role before the pilot launches: Formally shift senior staff in the registrar’s office, advising center, and financial aid office from information gatekeepers to algorithmic auditors. Make the transition visible, titled, and compensated accordingly. Not a consolation prize.
Deans + HR – before AI rollout
2
Audit your AI configurations: If a senior staff member helped configure your AI tool, have someone independent verify that the exception thresholds match actual institutional policy. Not what they remember the policy to be. The policy.
Provost Office + Registrar – within 60 days
3
Draft the escalation protocol: Create a clear, published answer to the question every staff member actually has: if the AI gives a student the wrong information, who is responsible, and who has the authority to correct it?
Provost Office – within 30 days
4
Run friction interviews: Ask staff in each functional area directly: what part of your job does the approved AI tool make harder? That answer tells you where the resistance lives before it calcifies into something you cannot fix.
Functional leaders – quarterly
5
Formalize the workarounds: Identify the Diane-style departmental standards already operating across your colleges. Integrate them into central governance. They are solving problems your official policy has not addressed yet.
The pace of AI change can feel relentless with tools, processes, and practices evolving almost weekly. We help organizations navigate this landscape with clarity, balancing experimentation with governance, and turning AI’s potential into practical, measurable outcomes. If you’re looking to explore how AI can work inside your organization—not just in theory, but in practice—we’d love to be a partner in that journey. Request an AI briefing.
Key Takeaways
AI adoption often fails at the departmental level, not the leadership level. Most AI initiatives do not fail through formal rejection. They gradually lose momentum when daily workflows are not redesigned around the new tools, leading staff to quietly route around them.
Resistance to AI is often about professional identity, not technology. Senior administrative staff may slow or resist AI initiatives because the tools threaten the specialized expertise and institutional authority they have built over decades.
Institutional power in higher education is often tied to undocumented expertise. Many administrative roles derive influence from knowing complex rules, exceptions, and workarounds. AI can rapidly replicate this knowledge, disrupting long-standing social and professional hierarchies.
AI resistance can hide inside the system itself. Staff involved in configuring AI tools may unintentionally or deliberately embed gatekeeping logic into the system, preserving their role while appearing to support adoption.
Successful AI adoption requires redefining expert roles. Institutions that move fastest reposition experienced staff from knowledge gatekeepers to system architects and algorithmic auditors, embedding their expertise directly into the institutional infrastructure.
Frequently Asked Questions
1. Why do AI pilots frequently stall within departments? Adoption often slows when the introduction of AI tools does not include a redesign of the underlying workflow. Without clear operational changes, only a few early adopters use the tool while others continue existing processes.
2. Is resistance to AI primarily driven by fear of the technology? Not usually. Resistance more often reflects concern about professional displacement or loss of authority, especially for staff whose roles are built on specialized institutional knowledge.
3. What is the “AI Precariat”? The term describes knowledge workers who face growing insecurity as AI systems replicate or automate expertise that once required years of specialized experience.
4. How can institutions prevent hidden resistance inside AI systems? Organizations should audit AI configurations independently to ensure system rules reflect official policy rather than individual interpretations or legacy workarounds.
5. What role should experienced staff play in an AI-enabled institution? Instead of guarding knowledge through manual processes, senior experts can act as architects and auditors who encode institutional expertise into AI systems and oversee their accuracy and governance.
