There is a small discipline I try to apply to any slide that ranks things. Read the headline, then stop, then count the rows underneath it, then go back to the headline. Most of the time the two agree and you have spent thirty seconds. Occasionally they do not, and what you have found is not a typing error. It is a claim that entered the document through the prose rather than through the evidence, and it will travel further than anything in the table, because a headline is what gets quoted into the next deck.
I went back this month to a sourcing review I worked on with a large European public-sector organisation that was preparing to re-tender its IT service desk. One module of that review scanned how the provider market charges for this kind of work. Its central exhibit is a comparison grid: twelve providers down the left, seven commercial models across the top, a mark in a cell where the provider was recorded against that model. Above the grid sits a single sentence of interpretation.
The sentence says that fixed price and unit-based pricing are the models predominantly in use at present, and that outcome-based pricing is the trend, with clients who are looking for value beyond cost savings.
Count the grid.
The tie the sentence steps over
Reading down the columns, unit or consumption pricing is marked for eleven of the twelve providers. Fixed price is marked for eight. Outcome-based pricing is marked for eight. Gain sharing drops to four, time and materials to three, revenue and profit sharing to two, and the staffing column to one.
That distribution has a clear shape. Consumption pricing is close to universal because it moves volume risk onto the buyer, a comfortable place for a provider to stand. The models that push business risk the other way thin out fast, and the two that push the most of it sit at two marks and four.
What the distribution does not contain is the ordering the headline asserts. Fixed price and outcome-based pricing are at the same number. One of them is described as the current state of the market and the other as a direction of travel, and the only thing separating them on that page is which side of the word "however" they were written on.
I want to be careful about what I am accusing anyone of, because the sentence is probably true. Fixed price almost certainly did carry more signed contracts in that market than outcome-based pricing did. Anyone who had sat through a dozen of these tenders would have said so without hesitation, and whoever wrote the headline had sat through more. The problem is not that the claim is wrong. The problem is that the exhibit underneath it cannot tell you whether it is right, and the page is arranged so that it looks as though it can.
Two questions wearing one exhibit
The grid's columns map onto the pages that follow it, one block per provider, and those blocks are written in each provider's own voice: the models it says it works with, described in its own vocabulary. So a mark is best read as the provider having put that model on its own list. It records willingness to sell. The module never says that anywhere, which is the first thing wrong with the page.
The headline is about what clients currently have. That is a different measurement entirely, taken on a different population, and no column in the grid holds it. A provider can offer outcome-based pricing to everybody and have sold it to nobody. One of the provider blocks elsewhere in the same module makes exactly that shape visible: it volunteers that four fifths of its contract book runs on three constructs it names, two of them priced per full-time equivalent and none of them outcome-based, while its row in the grid is marked across every column on the board.
So the module holds both numbers, an availability grid and scattered fragments of an adoption picture, and never puts them side by side. The headline borrows the authority of the first to make a statement about the second.
This is the failure I care about, and it is not a counting failure. Nobody miscounted. Somebody wrote a true sentence from experience, placed it above a chart that measures a neighbouring thing, and the reader takes away a trend line the evidence in front of them does not contain. Six months later that trend line is in a business case, and by then it has no provenance at all.
What the fill rate says about flexibility
The page before the grid carries its own headline, and that one asserts that most providers are flexible about the pricing model they will work with.
Thirty-seven marks in eighty-four cells is forty-four percent. Read row by row, one provider is marked on all seven models and one on a single model, and the middle of the field sits at two or three. That is not a flexible market. That is a market where the average participant will comfortably discuss three constructs and has to be pushed on the rest.
Both headlines are the same move. Each is a reasonable summary of what an experienced person believes, sitting above a table that says something narrower, and each is generous in the same direction, toward a market more mature and more accommodating than the marks record. When the prose and the exhibit disagree, the disagreement is almost never random. It runs toward the answer that makes the next stage of the work easier to recommend.
- 8 and 8
- Providers marked for fixed price, and for outcome-based pricing, out of twelve
- 11 of 12
- Providers marked for unit or consumption pricing, the actual leader
- 37 of 84
- Grid cells carrying a mark, a fill rate of 44 percent
- 1 of 12
- Providers marked for the staffing model, the thinnest column in the grid
The repair is not more data
None of this needed another questionnaire round. It needed three things on the page, the same three a RealAI Platform review insists on before a pack reaches a decision.
Say what a mark means. One line under the grid: this cell is marked where a provider named the model in its response. That sentence makes every count above it honest, and turns the gap between offering and adoption from invisible into visible.
Separate the two claims physically. If you want to say what clients are buying, that is a second exhibit with its own source, even if the source is a paragraph admitting the estimate is judgement. A named estimate that carries its own uncertainty is worth more than an unnamed one borrowing certainty from the chart below it.
Let the counts write the headline. "Consumption pricing is near-universal, fixed price and outcome-based pricing sit level, and everything else is thin" is a duller sentence and a better one. It is also more useful to a buyer, because it leaves the real question standing: two thirds of the field will quote outcome-based work, and nothing in the module says how much of it has ever been signed.
Why I am rehearsing a slide nearly ten years old
Because the same page is being produced again right now, with different column headings.
The AI market scans crossing my desk this quarter are built the same way. A grid of providers down the side and capabilities across the top: autonomous agents, orchestration across several models, retrieval-augmented generation, copilots embedded in the line-of-business tool, graded autonomy with a human approval step. The marks come from vendor responses. Above the grid sits a sentence about what enterprises are actually running in production.
The gap between those two is wider now than it was for service-desk pricing, not narrower. Nearly every provider marks the retrieval column, because retrieval is a pattern rather than a product. Nearly every one marks the agent column too. What those grids do not carry is a column for the evaluation set and the rubric it is scored against, a column for what the agent did the last time it was wrong, or a column naming the person who owns the autonomy threshold. Those are the questions that would separate the field, and they are absent because the capability taxonomy gets fixed before the responses arrive.
There is a second reason to be strict about it now. With the EU AI Act in force, a procurement file is no longer only a commercial artefact. A headline about market adoption sitting above a chart of vendor self-declarations is the kind of thing model risk management functions are being asked to stand behind, and the MLOps evidence that would settle it, the run logs and the evaluation history, sits with the provider rather than the buyer. That evidence is worth asking for at questionnaire stage, when a provider still wants the work.
A grid records what a provider will sell you. A headline about what clients are buying cannot be read off it, and the moment the two sit on one page the reader stops being able to tell which question was answered.
The discipline is cheap and it does not require any expertise in the subject. Read the headline. Count the rows. If the ranking in the sentence is not visible in the marks, the sentence came from somewhere else, and the only responsible thing to do is say where.
Counts are as recorded in the market-scan module of a sourcing review run with a large European public-sector organisation ahead of an IT service desk re-tender: a comparison grid of twelve providers against seven commercial models, sitting above per-provider pages written in each provider's own voice. That review produced findings and recommendations for a buying decision, not delivered results. Reading the grid against its own headline is ours.
“A grid records what a provider will sell you. A headline about what clients are buying cannot be read off it, and the moment the two sit on one page the reader stops being able to tell which question was answered.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.