Agentic AI Transformation Is Not a Program

Agentic AI transformation is a weekly practice run by named people, plus a layer that watches every agent in real time. A program can fund that work, but it cannot do the work, and companies that treat the program as the transformation end up with roadmaps and very few systems in production.

In most large companies, I would bet, the agents are already shipping without the executive team. The routes in are ordinary ones: a vendor switches on agent features at renewal, a team connects an assistant to the CRM, and nobody adds either to a list. The CTO who has been asked for an "agentic AI transformation strategy" may not know how many are running.

That leaves a board between two kinds of failure. The first is loud: an agent mistreats a customer and becomes a news story. The second is quiet: an expensive transformation program produces decks and roadmaps while very little reaches production.

The base rates are not encouraging. Over 40% of agentic AI projects will be canceled by the end of 2027, Gartner predicted in June 2025, citing escalating costs, unclear business value or inadequate risk controls. In BCG's September 2025 study of more than 1,250 firms, 5% were achieving AI value at scale and 60% were not achieving material value at all.

I have spent more than fifteen years watching this failure pattern. It used to be called digital transformation, then AI transformation, and now it is being relabelled agentic AI transformation. The label keeps changing and the mistake stays the same.

What the McDonald's drive-thru actually teaches

In June 2024 McDonald's told its franchisees it would end its test of automated order taking with IBM, which had run in more than 100 restaurants. By then, videos of the system getting orders wrong had spread online: one customer found it adding bacon to ice cream, and another watched it total up 260 Chicken McNuggets. IBM built the system and McDonald's deployed it.

Most coverage read it as a technology story. I read it as a practice story. I found no public account of who at either company owned what the system said to customers, or of a routine that would catch its errors before customers filmed them. That gap is my inference from the public record, and it is the gap this piece is about.

I expect the 2025 and 2026 cohort of enterprise agent failures to be larger and harder to see, because most post-mortems stay private. That is a view, not a measurement. The public case studies may take years to appear.

Underneath sits a familiar pattern. An executive committee funds a transformation. A deck describes it. A program manages it. Meanwhile the agent, the system actually making decisions in production at three in the morning, runs on whatever defaults its vendor shipped, and watching it is nobody's job by name.

Two architectures, run in parallel

In the organizations I see getting this right, two architectures run side by side.

The first is how the organization transforms: slowly, deliberately, cycle by cycle, on a human clock. I call this the 4-Step Loop, and it borrows from Lean Startup's build-measure-learn, Kotter's dual operating system, Toyota's kaizen and Agile's "responding to change", applied to AI. The pattern is older than AI. What changes is the domain.

The second is how the agents are governed: continuously, in real time, on a machine clock. I call this the Observability Layer, and the agentic era is what makes it necessary. A drifting prediction model could wait for the next quarterly review. An agent can act before the message thread about it finishes loading.

You need both. Start with the human layer.

The Human Layer: The 4-Step Loop

Step 1: ASSESS

Your first loop starts at your lowest pillar: the constraint is where transformation begins, and constraints shift as you move them. That is Goldratt's Theory of Constraints. You do not move five pillars symmetrically. You find the one holding the others back, you move it, you reassess, and the constraint shifts.

The tool I use for this is the AI Readiness Canvas: 15 questions and 8 to 10 minutes. It scores you across five pillars (Strategic Alignment, Data Infrastructure, Talent & Culture, Operational Processes, and Ethics & Governance) and shows which one is your constraint today.

Expect your first honest inventory of AI systems to be longer than the list IT keeps. Agents arrive inside tools people already pay for, and a list built from procurement records will miss some of them. You cannot transform what you cannot name.

Step 2: FOCUS

The weakest pillar is the only pillar for 90 days.

This is where most programs die. A steering committee sees five pillars, authorizes dozens of initiatives, distributes budget evenly, and spends three quarters producing symmetric mediocrity. Every pillar moves a little. No pillar moves enough to matter.

Resist it. Put the name of the weakest pillar on a board slide. Tell your team it is the only thing that matters until the next Canvas. Tell your CFO the other four pillars stay exactly where they are until this one moves.

The board will be uncomfortable. Good. That is the signal you have focus.

Step 3: PRACTICE

One practice for the weakest pillar, run once a week and protected on the calendar like a board meeting.

Strategic Alignment weakest: the Weekly Alignment Check. That is the whole meeting. With agents in production, it is the one place you catch strategy drift. The roadmap item you funded in Q1 is not the agent running in production in Q3, and the practice is how you notice.

Data Infrastructure weakest: the Data Honesty Meeting. What data do we claim to have? What can we actually access, right now, without a cleanup project? If you do not know what your agents can reach, they know more about your data than you do. The meeting surfaces the gap before an agent finds it.

Operational Processes weakest: the Weekly Inventory. Every AI system running in your organization, written down, owned by a name. With agents, the inventory has to include the ones people bought or switched on without telling anyone. An inventory that matches IT's list exactly has probably missed some.

Talent & Culture weakest: the Uncomfortable Upskilling. Five people whose decisions matter, ahead of company-wide training for 5,000. They are the people whose judgment agents will challenge, so pick them accordingly: your best product manager, your most senior underwriter, your lead clinician. The agent will either augment their judgment or quietly replace it. Train them first, and deeply.

Ethics & Governance weakest: the Monthly Stop. Once a month, the named owners decide whether any live system should be paused, and at least once a quarter one actually is: a system stopped in production after going live, which is a different thing from a proposal rejected in committee. With autonomous systems, MVG's accountability matrix becomes the load-bearing wall, because shipping without named ownership is the gap the McDonald's case points to. A governance process that has never stopped anything is documenting, and calling it governing.

That practice draws on Minimum Viable Governance, which I have written about separately: its three questions (which AI systems do we have, what could go wrong with each, and who can pause each one) and the accountability matrix they produce. The Loop's governance practice needs a named person with the authority to pause, and MVG is how you assign one. The gap is wide. Deloitte's 2026 State of AI in the Enterprise found that close to three-quarters of companies plan to deploy agentic AI within two years, and only 21% of those report a mature model for agent governance.

Step 4: MEASURE

Three questions every 90 days. Ask them of yourself, in a room, with the answers written down.

1. Did the weakest pillar score improve on the Canvas?
2. Did the practice actually happen every week, or did it get rescheduled?
3. What did the practice surface that we didn't know before?

If any answer is "no" or "nothing," the practice is theater. Change it.

Step 4 fails most often on the practice that happens faithfully and surfaces nothing. That is a meeting, whatever the status report says. A practice that uncovers nothing the organization didn't already know has become a ritual. Kill it and design a sharper one.

What the Loop doesn't cover

Before the practitioners pile on in the comments: I know what this framework doesn't do.

The Loop is delivery discipline. It does not replace platform rationalization decisions. It does not resolve multi-year vendor lock-in. It does not settle the org-design fight with your COO about whether AI capability sits inside the business units or in a shared service. Those need program structure: named owners, named budgets, named timelines, auditable decisions.

What the Loop retires is the myth that the program is the transformation. The program allocates capital and the practice changes how people work. Run both, and keep them distinct.

The Machine Layer: The Observability Layer

The 4-Step Loop is how your organization transforms. On its own, it does not watch the agents.

Agents act faster than any weekly practice. A team can run a perfect Tuesday meeting and still have an agent make a costly commitment between Friday evening and Monday morning. The Loop moves the organization. Something else has to watch the agent.

That something is a second architecture running in parallel, the Observability Layer. It has three components.

1. Instrumentation. Every agent's decisions logged in real time, queryable and reviewable. A weekly report is too slow and a quarterly audit slower still. The named accountable person needs a view they can open at any hour. If your agent cannot answer "what did you decide at 2:47 on Sunday morning, and why," it is running you.

2. Thresholds. Three stop conditions per agent, written before deployment. If the agent triggers any one of them, it stops automatically, with no meeting and no escalation chain. This takes MVG's pause threshold and deployment gate down to the level of a single agent. The conditions are written in advance because a threshold debated in the moment will not hold.

3. Named Escalation. One human, on call around the clock, with documented authority to pull the plug in real time. A rotation of three people sharing the duty spreads the accountability until nobody holds it. One name, one number, one decision. This takes MVG's accountability matrix down to the agent. A committee cannot be woken at 11pm on a Friday by the General Counsel and asked to pull a system down. A person can.

Without the Observability Layer, the Loop produces transformation theater: you are "assessing, focusing, practicing, measuring" things you cannot actually see. With it, the Loop has eyes.

The pressure is visible in the numbers. A drifting prediction model used to give you weeks to notice; an agent can transact, escalate or commit your company to a position in seconds. In S&P Global's surveys of companies in North America and Europe, the share abandoning most of their AI initiatives before production rose from 17% in the 2024 edition to 42% in 2025. My expectation is that the agents still running in two years will be the most closely watched ones.

The 30-Day Start

This is what I tell executives to do on Monday morning.

Day 1. Take the Canvas. You will know your weakest pillar by lunch.

Day 2. Confirm the weakest pillar with your team. Say it out loud. Write it on a slide. Put it in the all-hands deck. Tell the other four pillar owners they are on hold for 90 days. They will not like it, and that is the work.

Days 3 to 7. Pick your one weekly practice for the weakest pillar. Schedule it as recurring on the calendar: same day, same time, same attendees, every week. Protect it like a board meeting, because it is one, just smaller and more frequent.

Days 8 to 30. Run the practice three times. Don't skip. Don't reschedule. Don't delegate. Show up.

In parallel, if you already have agents in production: inventory them today. Which ones have instrumentation? Which ones have stop thresholds written down? Who, by name, can stop each one on the spot? If the answers are no, no and no, start the Observability Layer on Day 1, because those agents are not waiting for your Tuesday meeting.

Day 30. First measurement. What moved? What surfaced? What's theater?

If you have run three meetings faithfully and learned nothing new, the practice is wrong. Redesign it. If the weakest pillar hasn't moved but the practice has surfaced three things the executive team didn't know, you are doing this correctly. The score moves last. The awareness moves first.

One honest caveat on the 5-versus-5,000 question

The Uncomfortable Upskilling practice concentrates AI capability in 5 high-agency people first. Expanding it to the 5,000 is a second-order CEO responsibility, and the Loop does not solve it. Without an expansion plan, you have built an AI elite inside a workforce that didn't get the training. That is a governance failure, whatever the practice achieved.

Start with 5. Expand by pull. Track who in the 5,000 is asking the 5 for help, and build your second cohort from that list. The Loop's job is to create the seed. The CEO's job is to make sure the seed grows into a forest and stays open to everyone.

The close

Full disclosure: I advise on exactly this work. Take the argument with that context.

The organizations that get through this will have small practices, run every week, and clear sight of the agents already in production. The size of their transformation program will matter much less.

So the question for your next executive meeting is narrower than a strategy: what are we doing on Tuesday, and who is watching the agent we shipped on Friday?

If you want a place to start, take the Canvas and name your constraint. The AI Transformation Loop Workbook (PDF, 6 pages) has printable pages to run the Loop with your team.

Sources and changes

• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, 25 June 2025.
• BCG, The Widening AI Value Gap, September 2025.
• CNBC, "McDonald's to end IBM AI drive-thru test", 17 June 2024. The New York Times, "260 McNuggets? McDonald's Ends A.I. Drive-Through Tests Amid Errors", 21 June 2024.
• Deloitte, State of AI in the Enterprise 2026, press release, 21 January 2026.
• S&P Global Market Intelligence, "Generative AI shows rapid growth but yields mixed results", 27 October 2025.

Changes on 5 October 2026. The argument, the Loop and the Observability Layer are unchanged. Removed because no public source could be found: a client story of 12 AI systems that turned out to be 41, the claim that an inventory "typically surfaces 3x what IT reports", and the "$40 million" figures, which were illustrations written as if they were costs. The opening scene of named departments running unlisted agents was rewritten as my view, because it described no real company. The claim that the Canvas needs no email was removed, because it asks for one after the fifth question. The Monthly Stop now says what it is: a monthly decision, with at least one real pause a quarter. Three custom charts were replaced with library visuals, and the companion workbook was corrected to match.