The Buyers Brought a Machine
Two things moved, and they were reported as unrelated.
Start with the buyers. Channel Dive covered new research from the AI Revenue Institute — and before the number, know where it comes from. The AI Revenue Institute is a research and consulting firm that sells advisory work on exactly this problem, and this study is the research it launched itself with. No funding is disclosed. Weigh it accordingly. They surveyed five hundred and twenty-one decision-makers. More than half of them said they had already removed a vendor from consideration after an AI assistant pointed out a shortcoming. Three-quarters use it to go through vendor documentation. Eighty-two percent said that if the AI turned up problems with a vendor somebody had recommended to them, they would be more skeptical of the person who made the recommendation — or avoid them altogether.
Now the second number, which came from a completely different part of the business press.
OpenAI cut the developer price on its frontier model by more than twenty percent last week. Reuters has the figures — five dollars per million input tokens down to four, thirty dollars per million output down to twenty. That is the top of the market moving, not a discount tier. Google did the same thing to Gemini 3.7 Flash: a fifty percent introductory cut, running until the thirty-first of December, at which point the price doubles again.
And here is what happened the last time those prices came down. At the end of July, OpenAI cut its cheap model by eighty percent — seven dollars per million tokens down to a dollar forty — and its mid-tier model by twenty percent. Business Insider, working from analysis by TD Cowen, went back and tracked what followed. Usage on the cheap model rose roughly fourteenfold. And revenue from that model went up afterward, not down.
Take those as a sequence of events, without deciding yet what they mean. The price of asking a machine to read something for you fell across the board — by eighty percent at the cheap end. Consumption did not settle at the new price. It went well past it.
And in research fielded across that same stretch, more than half of the people buying technology said a machine’s reading of a vendor had already taken one off their list.
One of those numbers is causing the other one.
Cheaper Made It Bigger
The reason consumption outruns the price cut has a name now, and Gartner is the one using it.
They call it the Inference Paradox, and the mechanism underneath it is this. Better unit economics do not produce a smaller bill, because nobody spends the same money on the same amount of work. They spend the same money on more work. Gartner’s forecast, reported by ARN, is that the inference cost of a single agentic workflow rises more than fivefold by 2028 — not despite falling model prices, but through them. Their senior director analyst Will Sommer’s point is that every new generation of capability demands more tokens than the last, and there is no universal cheap model coming to rescue anyone from that.
Now look at where the extra tokens actually go, because that is the part that decides everything after it.
Gartner’s numbers, in Computer Weekly, put a simple chatbot question at about a penny. An agent performing a task — planning it, retrieving what it needs, checking the answer, going back for more — costs up to a dollar fifty. That is a hundred and fifty times the price for the same job. Nobody in this industry would tolerate a hundred and fifty times on anything they could see. This one arrives as a rounding error per query and only shows up in aggregate, at the end of a month, on somebody else’s invoice. The same analysis has those agents consuming five to thirty times more tokens than a chatbot to solve an equivalent problem, and models tuned for planning and learning generating tokens at eight to ten times the cost of the simple kind.
So a price cut does not buy the same question, cheaper. It buys a categorically different question. The long, patient, multi-step kind that was never worth paying for before.
Deloitte’s work on token economics shows what that looks like once it is loose inside an organization. Google processed something on the order of four hundred and eighty trillion tokens a month across 2025 — fifty times the year before. And one health care organization Deloitte describes grew its own consumption eight to ten percent every month, which sounds survivable right up until you run it out. Eight percent a month is two and a half times bigger in a year. Ten percent is more than three. That organization had run past six million dollars in unplanned annual cost before anyone in finance could work out what was driving it.
Nobody approved that. It grew because the cost of asking one more question fell below the level where a person stops to consider whether to ask it.
And the first thing anybody does with a question that costs nothing to ask is look somebody up.
And that changes what everything you have ever published is actually for.
Your Website Is a Deposition
Your website stopped being a brochure and became a deposition.
Everything you have published rests on an assumption that no longer holds — that a person would arrive, spend ninety seconds, and form an impression. The thing arriving now has unlimited seconds and forms no impression at all. It is looking for retrievable, checkable specifics. It does not get persuaded; it extracts and it compares. And where there is nothing to extract, it does not record uncertainty. It records absence. Absence is a finding.
The obvious response is to publish more, and that is precisely the response the evidence closes off.
Pew Research went and measured it. In a random sample of roughly ten thousand webpages taken this July, about one in ten showed strong signs of having been written by a machine. Among pages published since ChatGPT launched in late 2022, it is more than a third. And the place it concentrates is commercial dot-coms — which is to say, sites exactly like yours, doing exactly what you would do. Generic competence is now the most abundant substance on the internet. You cannot differentiate inside it, because it is the medium.
And where it is thickest, the platforms have started pushing back. More than a million people have now clicked LinkedIn’s “seems like AI slop” button, and LinkedIn reports that views on flagged posts drop by roughly forty percent. That sits against an analysis from Pangram finding forty-one percent of long-form posts on the platform were fully machine-generated. So the move most providers would reach for — more posts, more content, more presence — now carries a measurable penalty on the one network where this audience actually lives.
Which leaves a single lever, and it is an odd one for this industry to have to pick up.
Publish the things about your business that are true, checkable, and that nobody else is able to say. Your real response times, measured, with the month attached to them. What you cover and specifically do not cover, in the language of the agreement. How you price and on what basis. Named outcomes with dates.
Or keep the site you have — well built, carefully worded, aimed at a human being who is no longer the first one through the door — and get taken off lists you were never told you were on.
And there is one item on that list you are going to argue with me about.
Because of everything on that list, the one you will resist publishing is the pricing, and that is the exact field the machine reads as blank. You do not have to post a number — you have to post the basis. What you charge per, what moves it up, what moves it down, and what is out of scope at any price. The provider who publishes the logic gets retrieved and quoted inside a buyer’s research pass. The one protecting the number gets summarized as “pricing not disclosed,” and that is a sentence that has never won anybody a deal.
What to Consider
Find out what it says about you now, before you change a word. Open three or four different models and ask each one to evaluate your company as a managed services provider for a sixty-person business, then ask specifically what it can and cannot determine about how you price. Read the answers without defending yourself. That output is your current first sales call, it has been running without you for months, and it will tell you which particular absence is costing you rather than leaving you to guess.
Publish the basis in a form built for extraction, not for reassurance. The instinct is to write a warm paragraph about how every client is different and pricing depends on their needs, and a retrieval system gets exactly nothing out of that sentence. What works is flat and declarative, with the unit named — we price per user, servers counted separately at this ratio, after-hours included to this threshold, onboarding billed this way. You are not writing to persuade a reader anymore. You are writing so a machine can quote you accurately to someone you have never met.
Write down what you do not cover, because that is the field being checked. The finding in that buyer research was not that AI made vendors look expensive — it was that more than half of them dropped a vendor after AI surfaced a shortcoming. Something it went looking for and found. If you never state your exclusions, they get inferred from your silence, or borrowed from a competitor’s page that does state theirs, and you get compared against a boundary you did not set. Publishing your limits is the only way you get to define what counts as a shortcoming in your own listing.
Think about what happens when everybody catches on to this. The specifics get published, the machine starts comparing them against each other, and the only thing that cannot be manufactured at that point is a record with dates on it. A response time you have been publishing since last spring is evidence. The same number posted the week a buyer asks for it is a claim.
Picture the provider who did this in a quiet week. Six months from now a buyer they have never spoken to runs the same research pass across four shops, and theirs is the only one that comes back with an actual pricing basis, actual coverage limits, and a number with a date attached. They do not win because they were cheapest. They win because they were the only one the machine could describe.
If this trend continues, within the next twelve months the first quote most buyers ever see from you will be one a model assembled out of your public record without asking your permission — and the providers who never published a basis will discover they are being priced by inference, against a competitor who wrote theirs down.