After a year on the road speaking to private-equity firms, their portfolio companies, and audiences at Millennial Alliance, REMA, and similar venues, one pattern has become impossible to ignore: most enterprises are still failing at AI adoption. They are failing because they are running the wrong playbook and blaming the technology, bad use cases, or bad data context.
They are treating AI like a tech project. It isn’t. AI is a digital workforce transformation, and the businesses that win the next decade will be the ones that stop trying to bolt AI onto how they already operate and start re-architecting themselves as operating systems — organizations capable of running like a program: consistently, repeatedly, and at a scale no human-only company can match.
What follows is a synthesis of three models I use in those engagements — the Stewardship Model, the Adoption Model, and the Operating Model — plus the four most common fallacies I see leaders fall into, and a candid conversation about the human consequences of getting this right.
The Paradox: Massive Progress, Marginal Results
AI capabilities are advancing at a pace most organizations cannot absorb. Models, tools, and platforms are evolving faster than the budget cycles built to fund them. Yet despite the hype and the investment, the majority of organizations still fail to create measurable, sustainable value from their AI initiatives.
The root cause of that failure is rarely the technology itself. It is in the operating models, governance, and human systems built around the technology.
Executives keep asking “What can AI do?” when the real question is “What should AI do?” — and what should the organization look like once it does it?
The fatal assumption underneath most stalled programs is that AI systems will somehow intuit business context. They don’t. AI executes exactly what it is told, no more and no less. It cannot self-correct for strategic misalignment, and it has no inherent awareness of your goals, your culture, or your constraints. The human layer of judgment, ethics, and organizational context is irreplaceable — and that is precisely the layer most transformation programs underfund.
“We’re not building AI tools anymore. We need to shift to building digital workforces.”
Four Fallacies That Are Stalling the AI Industry
Before describing the three models, it is worth naming the recurring patterns I see kill AI initiatives across industries and company sizes. If any of these sound familiar, you are not alone, and you are not too far gone to fix it.
Fallacy 1: Treating AI Like a Tech Problem Instead of a Workforce Problem
Most organizations are running their old ERP-deployment playbook on AI: pick a tool, stand up an implementation team, declare a go-live date, train the users, and claim victory. That playbook works when you are introducing a tool. Stop thinking of AI as only a tool. AI is a digital workforce.
The energy that would have gone into change management for a software rollout has to be redirected into something far more ambitious: designing how your human and digital workforces work together within a single operating model that produces results at orders of magnitude greater than what humans and tools could ever produce alone.
If your AI program is owned by IT and measured in deployment milestones, you are running the wrong playbook.
Fallacy 2: Targeting the Highest-Impact Step First, Without Mapping Up- and Downstream
This is the fallacy I see kill more pilots than any other. Leaders look at their value chain, identify the highest-impact bottleneck, and point AI at it. It works. The metrics improve. And then the program fails.
Why? Because every process has an input and an output, and businesses are interconnected systems. When the AI-enabled step starts overproducing, the downstream step gets buried — it cannot consume the new throughput, so it slows down, and the perceived ROI collapses. Meanwhile, the upstream step cannot feed the new super-efficient AI-enabled step fast enough, so the AI sits idle, and the business concludes that the technology underperformed.
The metrics were overestimated because no one factored in the connective tissue between departments. The project didn’t fail because the AI failed, but because the organization treated a process step as if it were independent of the rest of the business.
The strategy of rollout matters as much as the rollout itself. AI initiatives must be sequenced with the system in mind.
Fallacy 3: Treating Headcount Reduction as the Goal
Almost every AI business case I see opens with an FTE-savings line. This is, in my experience, the single most expensive mistake in the program.
The AI Steward — the human who supervises agents, validates their output, and translates evolving business reality into updated instructions and guardrails — is a brand-new muscle. Higher education is four to six years away from producing this role formally. There simply isn’t a labor market for it yet. The stewards in your organization are the ones you grow.
That growth requires real investment in three skills:
- Declarative communication. Stewards must be able to task AI orchestrator agents clearly enough to eliminate vagueness. Vague tasking is the single biggest invitation to hallucination.
- Deep domain knowledge. Every gap an AI has to fill from its own latent knowledge — rather than from context the steward provided — is another opportunity for hallucination. Stewards need enough domain depth to supply the right context up front.
- Knowing what good looks like. Stewards don’t need to know the right answer in advance. They need to be able to look at a result — on a dashboard, in a report, in a customer interaction — and recognize “That’s not right.” Then they need observability skills to dig into the logs and determine whether the failure was due to a bad prompt or bad context data.
You don’t hire those people off the street today. You build them.
Fallacy 4: Cutting Heads Before the Stewards Exist
Closely related to fallacy three, and worth its own warning: do not let your finance organization pull headcount reduction forward in the program.
Stewards do not yet exist in the marketplace. If you spend tens of thousands of dollars per person training your workforce to be stewards and then push them out the door early to hit a savings target, you have just funded your competitors’ AI transformation with your own balance sheet.
Headcount reduction belongs at the end of the journey, and even then it should be the option of last resort. Stewards are usually transplantable inside your business. Move them. Don’t lose them.
ROI metrics also need to change. AI initiatives whose business case is anchored on efficiency are running a finite game. The infinite game is asymmetric growth — growing revenue and capability faster than your cost structure — and that is the game AI is built for. Private equity doesn’t win by owning technology. It wins by owning companies that scale faster than their cost structures. AI is the most powerful lever ever invented for exactly that.
The Three Models: A Coherent Framework for Getting It Right
Here is where most engagements I lead spend their time. The three models below are three perspectives within the same transformation:
- The Stewardship Model describes the end state — what your people do, and how they operate, after the transformation.
- The Adoption Model describes the journey from where you are today to that end state.
- The Operating Model is the manifestation of getting the first two right. If your end state and your journey are well-designed, the operating model emerges. If they aren’t, no operating-model diagram will save you.
Most leaders I meet have heard about one or two of these. Few have integrated all three.
The Stewardship Model: Designing the End State for Your People
Most organizations see AI adoption as a technology problem. The Stewardship Model reframes it as an operating-model problem with five layers, each of which has to be designed deliberately.
Layer 1 — Have a guiding star. Define what you will and will not do with AI. Establish how it is governed across the organization. This is a decision. Without it, every downstream choice gets made by accident.
Layer 2 — Measure impact on board-level metrics. Connect AI outcomes to the metrics that actually matter to leadership, the board, and investors. If your AI dashboard does not roll up to the same metrics your CEO presents at the board meeting, your AI dashboard is not the right dashboard.
Layer 3 — Automate low value to AI; squeeze high value to humans. Redirect human capital toward the work that requires judgment, creativity, and relationships. The point of automating busywork is to free humans to do the work they are uniquely good at, thereby maximizing outcomes, not to eliminate humans. Focusing on the correct goal and target is critical here.
Layer 4 — Train your human capital. Build balanced scorecards and tactical measures. Equip your workforce to operate effectively alongside AI agents. This is the layer most programs skip, and it is the layer most programs fail in.
Layer 5 — Identify agents and address the connective tissue. Design the orchestration layer that connects agents to each other and to the operating model. Isolated agents cannot compound outcomes; however, Orchestrated agents can.
These five layers are an integrated operating system, with guiding principles at the top and agent execution at the base, all held together by continuous human oversight. Stewardship is a permanent operating responsibility, a new role within the business landscape.
The Adoption Model: Designing the Journey
If the Stewardship Model is the end state, the Adoption Model is the path. And the path is determined less by your technology choices than by your culture.
There is no universal AI adoption playbook. The right strategy is the one calibrated to your organization’s risk tolerance, decision-making speed, and change capacity. I find it useful to place organizations on a four-point spectrum:
- Risk-Averse: Governance-first. The framework is fully implemented before any deployment. Compliance leads every decision. These cultures will not implement AI as-they-go.
- Pragmatist: Wait for proven ROI before committing. Let others take the early risk; adopt once the value proposition is de-risked.
- Early Adopter: Launch strategic pilots to learn while doing. Willing to invest ahead of certainty in exchange for competitive advantage.
- Innovator: Rapid experimentation at pace. Embrace ambiguity, move fast, build governance iteratively as the program matures.
The honest answer to “How do you define your culture?” determines your entire implementation strategy. Mismatches between strategy and culture are a leading cause of stalled programs — an Innovator strategy in a Risk-Averse culture will be killed by the immune system; a Risk-Averse strategy in an Innovator culture will be abandoned in a quarter.
The five-fundamentals checklist applies regardless of where you sit on the spectrum:
- Augment people before replacing them. Replacement should never be the starting point — it destroys trust and undermines adoption.
- Invest in organizational change management. Technology adoption without OCM funding is a recipe for failure.
- Define success metrics before deployment. Know what winning looks like before you begin; it is impossible to address this without bias afterward.
- Automate busywork first. Low-value, high-volume tasks build organizational confidence and demonstrate ROI quickly.
- Continuously steward AI agents. The job isn’t done at deployment.
The journey then follows the AI maturity curve — from experimentation, to bolt-on automation, to integrated operations, to true agentic operations. Most organizations waste enormous time and capital traversing the first three stages by trial and error. The Stewardship Model exists to short-circuit that, by starting with the right measurements and the right operating principles from day one.
The Operating Model: What Naturally Emerges
The Operating Model is what you get when the first two are right. Don’t treat it as a separate initiative. It is the strategic recognition that you are no longer running an organization of people-using-tools; you are running a process engine — an architectural operating system capable of executing consistently and repeatedly at machine speed under human supervision.
A functional AI operating model has five interconnected layers:
- Leadership sets strategy, governance, vision, and resource allocation.
- Human Stewards supervise AI agents in real time. They are the human managers of the digital workforce — a new and critical role.
- Agents and sub-agents execute operational tasks autonomously within defined boundaries. Orchestrator agents manage specialized sub-agents across workflows.
- Internal structure defines ownership, accountability, and culture. Who owns AI outcomes? How is success recognized and rewarded?
- Feedback loops drive continuous improvement by surfacing performance signals from agents back to stewards and leadership for action.
Ownership must be distributed and clearly defined. Ambiguous accountability is one of the leading causes of stalled AI programs. The CEO defines the AI vision and champions transformation at the board. The CIO/CTO owns the architecture and platform. The Data Office leads data governance, quality, and access. Business units identify use cases, own the ROI justification, and drive adoption within their domains. AI Stewards monitor agents in production and flag anomalies. Risk and Legal own the governance and risk framework and ensure compliance.
When this is wired correctly, the business begins to behave like a program: predictable, observable, instrumented, scalable. That is the operating-system shift.
When Are You Actually Ready to Scale?
The gap between a successful proof of concept and enterprise-scale agentic operations is vast, and most organizations fall into it. Pilots succeed in controlled conditions. Bolt-on automations solve isolated problems. Few organizations reach true agentic operations — and the path isn’t blocked by technology but by the absence of the operating-model conditions required to scale with confidence.
Enterprise-readiness goes beyond a feature set. It is an operational condition. Five criteria must be met:
- Reliability: Agents complete tasks end-to-end without drift or unexpected behavior under real-world conditions.
- Observability: Full transparency into what agents are doing, why, and when — so problems can be detected and corrected before they escalate.
- Governance: Formal policies, audit trails, and oversight mechanisms are active and enforced.
- Human Stewardship: Trained stewards actively monitor agents, respond to escalations, and continuously align system behavior with evolving business intent.
- Measurable Business Value: Clear, repeatable ROI is demonstrable and attributable to specific agent operations.
Five operational metrics tell you whether you are there: Task Completion Rate (TCR), Human Escalation Rate (HER), Hallucination Rate (HR), Time to Resolution (TTR), and Cost Per Transaction (CPT). The thresholds I recommend before flipping the agentic-operations switch are TCR above ~95%, HER below ~10%, hallucination rate below ~2%, decreasing cost per transaction, and 100% audit compliance. Don’t scale what you can’t measure.
The Hard Truth: 30% Organic Churn — and Why It Isn’t a Headcount Strategy
I want to address something that almost no one says out loud at these engagements, but every leader needs to internalize: a business implementing AI seriously across its organization will naturally experience roughly 30% employee churn. This isn’t because those employees can’t do the work or can’t become stewards. It’s because of how humanity has been educated and trained for the last sixty-plus years.
We have trained generations of people to target repeatable excellence. Get the right answer. Don’t be creative. Be consistent. We have, frankly, dimmed the light of human creativity in a way that should bother us more than it does. And we have built a culture in which many people measure their self-worth — their entire identity — by the thing their work produces.
The world is changing the locus of value. The value is shifting from the thing to the ability to produce a thing.
Consider a 3D printer. What is more valuable: the design the printer uses, or the object it prints? The schematic is the value. The object has gone almost to zero. I can send a schematic to anyone with a 3D printer, and they can reproduce the object. But can they change the schematic? Can they invent a new schematic for the needs of the day? The architects, the designers, the stewards who can produce any “thing” — that is where the value has gone. The output is worth zero. The outcome is worth everything. And the ability to make more outcomes is where human ingenuity and creativity are going.
Not every human will grasp this shift. Many have tied their personal worth to the thing for so long that they are unable to see it any other way. They will reject the new paradigm and flee to a business that hasn’t yet adopted AI, hoping to preserve the cognitive frame that has defined their working life. That is the source of the 30%. It will happen organically.
This is precisely why there is no need to include headcount reduction in the strategy. It will occur as a natural consequence of the transformation. Your job as a leader shouldn’t be to push people out. Your job is to create the conditions under which the people who can make the shift do make it — and to retain them, retrain them, and reposition them as architects of outcomes rather than producers of things. That is the asymmetric-growth play.
A Note on Cybersecurity in the AI Era
A short but essential aside, because no AI conversation is complete without it. AI is being weaponized by adversaries to automate, accelerate, and scale attacks at a level we have never seen — automated vulnerability discovery, hyper-personalized AI-generated phishing, deepfake voice and video impersonation, AI-driven reconnaissance. The same capability also enables defenders to detect, respond, and neutralize threats faster than any human team alone — creating a new asymmetry in favor of well-prepared organizations.
The largest vulnerability remains humans, and traditional security awareness training is no longer sufficient when attackers can generate perfectly tailored deception at an industrial scale. Five actions that can be initiated immediately and that materially reduce AI-era risk: train employees on AI phishing using AI-generated simulations; deploy AI-based monitoring for continuous, real-time visibility; secure your open-source dependencies against AI-driven vulnerability discovery; audit AI agents and enforce least-privilege permissions rigorously; and establish governance policies that define what AI is permitted to do.
Governance is your first and most durable line of defense.
The Closing Argument
AI will transform every industry. The question is whether your organization will lead it or be left behind. The window for establishing competitive advantage is open now, and it is closing.
Three things determine whether you are on the right side of that window:
- Stop running the tech-deployment playbook. This is a digital workforce transformation, not an ERP rollout. Architect your operating system.
- Sequence your rollout for the system, not the step. Up- and downstream connective tissue is where ROI lives or dies. The metric that matters is the system’s throughput. Avoid focusing solely on the step’s throughput. It’s a very common trap.
- Invest in stewardship, and let churn happen organically. Stewards don’t exist in the labor market yet. Grow them. Don’t reduce headcount before they exist; don’t worry about reducing it after, because the people who can’t make the shift will leave on their own.
We are in the early Internet phase of AI. When the internet arrived, businesses bolted it on; the utility wasn’t obvious, and the early playbooks underperformed. Today, no one can imagine a business without internet connectivity. AI is at exactly that moment. The companies that re-architect themselves as operating systems will do more than just outperform their peers — they will be playing a different game entirely.
“Down-scaling is a finite game. Asymmetric growth is an infinite game. If your goal is down-scaling, you’re not using this tool right.”
When you’ve grown as far and as wide as you can, by all means, right-size your business. Until that day, get the ROI out of the people you’ve trained. Those who adapt, keep. Those who won’t or can’t, replace.
The output is worth zero. The outcome is worth everything. And the ability to make more outcomes is the only value worth building a company around.
