An argument against our own category
We sell AI transformation. What follows argues that most companies asking to buy it should spend the next quarter on something else, and that the something else does not involve a vendor.
This is not modesty. It is that the failure pattern is consistent enough to be predictable, and selling into it produces the 95% rather than the 5%.
What the research actually found
MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025, built on 300 public deployments, more than 150 executive interviews and $30-40 billion of tracked enterprise spend. The headline: roughly 95% of enterprise GenAI pilots delivered no measurable P&L impact.
The finding that matters more is the diagnosis. The authors are explicit that the divide is organisational, not technological — data readiness, integration depth and the capacity to learn from deployment separate the two groups. Model choice is not on the list.
Two secondary findings sharpen it. Deployments built through purchasing or partnership succeeded roughly 67% of the time against about a third of that for solo internal builds. And over 90% of firms showed a 'shadow AI economy' — employees using personal AI tools while the official pilot stalled.
Prerequisite one — data the model can rely on
A model does not hedge. Given inconsistent records it produces an answer with the same confidence it would use on perfect data, and it produces it faster than a person could produce a wrong one. Speed is the risk.
The check is unglamorous and takes an afternoon: export the data the workflow depends on, count duplicates, count records stale beyond twelve months, count fields that should be pick-lists and are free text. That percentage predicts your outcome better than any model benchmark.
Prerequisite two — a process someone has written down
You cannot automate a process that exists only as convention. What gets encoded is one person's version of it, and the disagreement surfaces after go-live as 'the AI is wrong' when the AI is faithfully reproducing an ambiguity that was always there.
The check is equally cheap: two people who own the process write its steps and decision rules independently. Where the lists diverge is your specification. It is a morning of calendar time and it routinely saves a quarter.
Why the sequencing instinct is wrong
The instinct is to prove value first and fix foundations later, and it is entirely reasonable — foundations are hard to fund without a demonstrated win. It is also the mechanism that produces stranded pilots.
What happens: the pilot runs on a clean, narrow slice, demos well, and gets funded for production. Production is where variance lives. The data that was fine for fifty curated records is not fine for fifty thousand real ones, and the process that seemed obvious turns out to have four undocumented exceptions.
The alternative is not slower. Run the two checks above before choosing a model. If they pass, the pilot proceeds with a real chance of shipping. If they fail, you have learned it for the cost of an afternoon rather than a quarter.
What we would tell you to do instead of hiring us
If you do exactly this and nothing else, you will be ahead of most of the 95%:
- Export the data behind your candidate workflow and count duplicates, dormant records and free-text fields that should be structured
- Have two process owners independently write the steps and decision rules; compare
- Pick a high-volume, low-variance workflow — not the interesting one
- Write down the number you expect to move and its value today
- Only then evaluate models, and evaluate them on your tasks rather than on leaderboards
Four of those five cost nothing but attention. If the first two fail, the honest recommendation is a data and process engagement, and if you would rather run that internally we will tell you what to look for.
The uncomfortable corollary for vendors
MIT's buy-versus-build finding cuts in our favour and we are not going to pretend it does not — partnership outperformed solo internal builds by roughly three to one. But the same study makes clear that partnership succeeds when the organisational prerequisites are in place, not instead of them.
A partner who sells you an AI programme without checking those prerequisites is selling you a place in the 95%. That is the part of the category we are arguing against.