AI Strategy Is a Category Mistake

Most of what is being written about AI ROI right now is an argument about the technology. It is an argument about the wrong thing.

I spent nearly fifteen years at McKinsey and Deloitte advising Fortune 50 companies on programs that look exactly like the ones now being labeled AI transformations. The labels change. The structural failures do not. Whether the program produces P&L impact has very little to do with the model you picked, the vendor you signed, or how good your prompts are. It has to do with how you bought it, who owns it, and what the contract says happens if the number does not move.

This is the part of the AI conversation that the technology press is missing.

The 95 percent are a procurement story

In August 2025, MIT published the number that became unavoidable in every AI strategy meeting: roughly 95 percent of generative AI pilots delivered no measurable P&L impact. The 2026 data has only made the picture more uncomfortable. Writer surveyed 2,400 executives this spring and found 79 percent reporting real friction in their programs and only 29 percent reporting meaningful returns. S&P Global counted 42 percent of companies abandoning most of their AI projects, more than double the prior year. Morgan Stanley found that only 21 percent of S&P 500 companies could cite a measurable AI benefit at all.

The common reading: the technology is overhyped, or the timing is off, or organizations are not ready. The data does not support any of those readings.

The 5 percent who got returns were running the same models on the same data, against the same problems. They did three things differently, and all three are procurement choices.

They named a single dollar baseline for a single workflow before the contract was signed. Days sales outstanding. Renewal-quote turnaround time. Support resolution time. Revenue per sales rep per quarter. One workflow, one number, one before-and-after.

They capped the engagement at a length where the number could actually move. Most AI contracts are sized for the vendor’s roadmap, not for measurement. Six to twelve weeks is the right cap for a workflow-level proof point. Eighteen months is a platform purchase wearing a pilot costume.

They tied price to the number. The vendors who will engage with outcome-linked pricing are a different population than the ones who will not. Vendors who refuse to discuss it are not avoiding risk. They are telling you their internal probability that the number moves, and they are sharing it with you for free if you read the contract.

The procurement choices explain most of the variance in the outcome data. The model choices explain almost none.

The category mistake

The phrase “AI strategy” is doing damage. When a program is called an AI strategy, the team optimizes the AI. Better models, better fine-tuning, a steering committee on governance, a working group on responsible use. None of that changes a P&L line.

What you actually have is a workflow strategy that uses AI. The strategy lives at the workflow. The AI is the instrument. Programs that move numbers are organized around the workflow. Programs that produce activity reports are organized around the technology.

This is also the answer to why so many AI initiatives feel like vendor portfolio management. They are. Once you name the work “AI strategy,” every vendor with an AI feature becomes a stakeholder. Decision rights migrate to procurement. The workflow that was supposed to change becomes a slide in a quarterly review.

If your steering committee charter is about AI, you are in the 95 percent. If the charter is about a specific operating metric, you are not. That is the diagnostic, in one sentence.

The mid-market advantage is incentive structure, not org chart simplicity

You have heard a version of this argument before, and the version you have heard is wrong. The conventional take is that mid-market companies have simpler org structures, so they can move faster. That is true. It is not the most important thing.

The actual advantage is incentive structure.

At a Fortune 100 company, senior leaders are paid to manage complexity. They are evaluated on the size of their organization, the number of cross-functional initiatives they sponsor, and the breadth of their portfolio. Building governance is a career move. Naming a single dollar number for a single workflow is not. The political incentive is to expand the program until everyone is a stakeholder and nobody owns the outcome. I have watched this happen from inside the firms that sold it.

At a $90M to $2B mid-market company, the senior team is paid to deliver outcomes. Building governance is overhead. The political incentive is to compress the program until one person owns it and the number is visible. This is a structural advantage and it is perishable. The fastest way to give it back is to hire a consultant who wants to build you a Center of Excellence.

Decision distance

A useful internal metric is what I call decision distance. Count the number of sign-offs an AI initiative crosses between a use case being identified and a production budget being released. Count the systems the data has to cross. Count the times the business case has to be re-explained to someone new. Count the people in the chain whose compensation does not move if the initiative works.

A typical mid-market initiative scores four to seven. A Fortune 100 initiative scores twelve or higher. The score is a strong predictor of whether the program ever ships. The most common mid-market failure mode is copying enterprise governance and re-creating enterprise distance voluntarily.

If you are running the program with the structure of a Fortune 100 transformation, you are paying for the overhead and giving up the advantage at the same time.

Three questions before Q3 budget planning

For each active AI initiative, ask which single workflow it is supposed to change, and write down the dollar baseline that exists today, before the program.

For any initiative that cannot answer both parts, ask whether it is a pilot that proves the technology or a program that changes a number. Pilots that prove technology do not produce budget justification. A pilot is a defense mechanism if it has no kill date and no quantified failure threshold. Most pilots are defense mechanisms.

For the next initiative on the roadmap, ask what the contract says happens if the number does not move. The answer tells you the vendor’s confidence in their own pitch. It also tells you whether your finance team should put the spend in the AI budget or in the legal budget.

The cleanest version of this advice is unbillable, lives in twenty-minute calls, and is how I get most of my real work. If you are in the middle of a board conversation about AI ROI and the framing feels off, reply to the dispatch when it lands in your inbox or find me through the contact page.