AI Comes Last

Most companies are being asked the same question this year, and most are answering it in the same order: find the places artificial intelligence fits, then build the plan around them. I have come to the conclusion that the order is wrong, and I can point to the hour that convinced me.

The question on the table was about the next three years. The leader answering said nothing about AI; it was all about growth. They talked about how much more the business expected from their part of it, about the work as it stands today, and about the distance between the two. They talked for a while. Nobody steered them back.

That was the design, and I still was not ready for how plainly it worked. I lead the planning work that produced that room, I am partway through it rather than past it, and what follows is what I am finding, with the unfinished parts marked as unfinished.

Ask leaders where AI could help and they will produce answers, and nothing in what follows is a criticism of them for it. Someone has a license, or a demo, or a board member who asked why the company is not doing more with AI, and every function of the business gets asked where this could help. Leaders do what leaders do with an ask like that. They answer it. A list forms. It is long, it is full of things that would be nice to have, and almost nothing on it carries a number anyone will defend when budgets get set. The first question pointed everyone at the tool, and a list is what that question returns.

None of this started with the technology. Paul Nutt at Ohio State spent twenty years studying about four hundred real decisions inside organizations and found that half of them were not fully in use two years later, and that failure was four times as likely when the people deciding took the first fix that came along instead of first working out what was driving the decision. A tool looking for somewhere to go is the first fix that came along, with a license attached.

The pattern is not local to any one company. McKinsey's 2026 global survey found nearly nine in ten organizations using AI somewhere and about four in ten able to point to any effect on earnings, a share that has not moved in a year. The small group that does see real earnings impact looks different in two ways that matter here: most of them set growth or innovation goals for AI rather than efficiency alone, and about three quarters of them have redesigned the work itself, against a quarter of everyone else. Read one way, that is a finding about scaling. Read another way, it is a finding about where the planning started.

So the question in our rooms was never about the validity of artificial intelligence. It was where the function needs to be in three years, what growth looks like from where that leader sits, and what in the work would have to change to get there. After enough of those hours, the thesis I am working from, and still testing, comes down to one question.

Is the plan or roadmap an inventory of places the technology fits, or a plan for growth that is enabled by the technology?

Start from the tool and you get a wish list. Start from each function's three-year growth, then map that growth to the work the function actually does, and you can see where AI carries part of the load. The second way is slower to begin and harder to run. It also produces the only kind of plan a leader will defend in a budget meeting: the kind that was their argument before it was anyone's plan.

The second thing I did not expect was how little the functions had in common once they started talking.

Every function pictures a different three years.

I expected the functions to differ in what they wanted most. The differences were in what growth meant, not how much of it, and you can hear them in what each leader tests a plan against.

• A leader who thinks in volume tests for capacity: will this hold at twice the load.
• A leader who thinks in reach tests for distance: will this work in the places we are not yet.
• A leader who thinks in what their people could stop doing by hand tests for relief: what does this take off the desk.

Put three leaders who test that differently in one room and ask them a single question about AI, and the hour goes to one test while the other two never get asked. Ask them separately about their own three years, and each one tells you something the others did not know to ask. Kahneman, Sibony, and Sunstein found the same thing about judgment in general: get each person's answer on its own before you put the answers together, because once a group is talking, the first view in the room pulls the rest toward it. That is a fair account of why separate sessions produced better answers than one room would have.

That is the part I keep turning over: the function-by-function design was meant to make the sessions manageable, and it turned out to be the reason the answers were any good.

Now, there is a cost in that design, and I want to name it before someone names it for me. If the functions never sit together, nobody in any room is building the picture of all the functions together while they talk. That picture gets assembled afterward, and it is easy to assemble badly. Someone writes that summary, and the method does not say who, or what keeps their version honest.

The AI question gets small once the work is on the table. The second half of each session was built on that, and it held every time.

The growth answer had to go somewhere, and where it went was onto the work. Not onto systems, and not yet onto AI. Onto the work: the value streams a function runs, meaning the chains of work that turn a request into a result, the steps inside them, where the hours go, where things wait, where a person is doing something because no record exists that could do it for them. This is the part most AI plans skip, and the reason is plain. It is tedious. Nobody gets excited about writing down how a request moves from one desk to the next. But it is the only step that tells you where "augment" is a real word and not a hope.

Once the work is written down, the AI question has somewhere to land: does any of this change if a machine can read, summarize, draft, classify, or predict? For some steps the answer is obviously yes. For others the answer is a better process, or a system the function already owns and does not use, or a hire. The exercise has to be allowed to say no. A map where every line comes back as an AI opportunity is a map to trust less, not more.

Boston Consulting Group puts the weight of an AI effort at roughly a tenth on the algorithms, a fifth on technology and data, and the rest, seven tenths, on people and process. The digits are an estimate; the shape is the point, because it was the shape of every session: most of the hour went to the work, and the technology took the last ten minutes.

The way it comes out after enough of these sessions: a function needs its growth set against the work it does now and the work as it could be, and then a plain look at where AI carries some of that. Nowhere in that sentence does anyone adopt AI, which is the thing most plans are actually about.

Comparing the answers changes what they are. It is the last thing this piece can say with confidence, and the claim that needs the most care.

The workshops are done. The capture is done. So is the comparison, three ways: the themes that repeat across functions, each function's ambitions against the systems and data it has today, and value against effort and risk. Three comparisons, each asking the same material a different question. The results are not the point here. What the step did to the material is, because that is about the method and not the findings.

It turned a stack of hopes into a set of arguments. The first comparison asks whether a growth story holds outside the room it was told in. The second asks whether an ambition has the systems and data under it, or will need them built first. The third asks whether a piece of work that looks like a home for AI carries enough value to be worth the effort and the risk. Those are questions, and the method does not answer them. It makes the decisions arguable, which is a different thing and probably the more important one.

Here the summary comes due. The person who writes it has more say over how each function's three years reads on the page than that function's leader had in the room, and nothing in the method polices that by itself. My working answer is to show the comparison back to the people whose answers went into it before anything gets ranked, so each function sees how its three years fared against everyone else's while there is still time to say the comparison missed something. Nutt's study backs the instinct: bringing the people a decision affects into the making of it was the tactic most tied to decisions that held, and the one managers used least, in about one decision in five. I do not know yet whether that is enough. If you have a better check, that is the question I would most like answered.

The roadmap itself is the step ahead, and it is not written. When it is, the question changes from where AI belongs to who owns the estate it lands in, and what gets retired, sequenced, or starved to make room. That is a different piece, and it will be a harder one, because a roadmap is the first thing in this exercise that can be wrong in public.

For now I keep coming back to that first room. A leader asked about three years, talking about growth and about the work, and not once reaching for the word everyone had come to discuss. If the plan that comes out of this is any good, it will be because the first question let them do that.

And if you have run this kind of exercise before and found a way to keep the summary honest to the people who gave the answers, what did it look like?

Sources

Boston Consulting Group. (2026). Artificial intelligence and AI at scale: Our approach [Capabilities page]. BCG. https://www.bcg.com/capabilities/artificial-intelligence 

Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A flaw in human judgment. Little, Brown Spark.

McKinsey & Company. (2026, August 25). The state of AI in 2026: On the road to ROI. McKinsey Global Survey. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

Nutt, P. C. (2002). Why decisions fail: Avoiding the blunders and traps that lead to debacles. Berrett-Koehler.  
https://news.osu.edu/half-of-business-decisions-fail-because-of-managements-blunders-new-study-finds/