AI Adoption

Everyone Gets the Same Model

By John J. BakerJuly 29, 20266 min read

A better model shipped recently. Your competitor got it the same afternoon you did, at the same price, without asking anyone's permission and without doing anything to earn it. That is the odd thing about this technology. The part everyone talks about is the part nobody owns.

The part you own is context: what the system knows about your business. Your pricing logic. Which customers get handled differently and why. What a good proposal looks like here as opposed to anywhere else. None of that arrives in a release. It accumulates, or it does not.

This piece is about how that context changes shape over time, from a sentence you retype every morning into something that runs part of your operation.

Two halves of the same setup

One of them arrives for everybody on the same afternoon. The other one is yours.

The model

Improves on a schedule you do not set.

Who improves it: A lab you have never spoken to
When you get it: The day it ships, automatically
Who else gets it: Every competitor, same day, same price
Advantage it creates: None. It moves everyone at once

Your context

Improves only if you build it, and then it keeps improving.

Who improves it: The people already doing the work
When you get it: Gradually, starting the week you begin
Who else gets it: Nobody. It does not exist anywhere else
Advantage it creates: Compounds, and widens the longer you run it

The half everyone shops for is the half nobody owns. Choosing a tool is a one-afternoon decision. Building context is the only part of this that is actually yours.

Why most of this stalls in year one

The most cited number in the last two years is that 95 percent of corporate generative AI pilots produced no measurable return. It comes from MIT's Project NANDA and its 2025 report on the state of AI in business, built from roughly 150 interviews, a survey of 350 employees, and a review of 300 public deployments. It is one report rather than settled evidence, and it has been misread almost as often as it has been quoted.

What is worth taking from it is not the headline. It is the diagnosis. The researchers named the failure the learning gap: the tools could not retain feedback, adapt to context, or improve over time. The complaint they heard from users was not that the model was not smart enough. It was that there was too much manual context required each time.

Every owner I talk to has already lived that sentence without having a name for it. You got a genuinely useful answer on Tuesday and a generic one on Thursday, from the same tool, on a similar question. The difference was almost never the model. It was how much you had told it before you asked.

The pilots did not fail on model quality

The MIT researchers called it the learning gap: tools that could not retain feedback, adapt to context, or improve over time.

Of corporate generative AI pilots showed no measurable return: 95%Of corporate generative AI pilots showed no measurable return
Reached real value, by integrating deeply enough to learn from use: 5%Reached real value, by integrating deeply enough to learn from use

What the users actually said

They do not learn from our feedback.
Too much manual context is required each time.

Neither one is a complaint about intelligence. Both are complaints about memory.

Source: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, drawn from roughly 150 leader interviews, a survey of 350 employees, and a review of 300 public deployments. Coverage of the findings. One report rather than settled evidence, and widely misread. The diagnosis is the useful part, not the headline number.

The five shapes context takes

Context does not go from nothing to asset in one move. It passes through stages, and each one is a real improvement on the last. Most businesses are living on the first or second and have concluded that AI is mildly useful, which is the correct conclusion from where they are standing.

The prompt

Most start here

You type what you need, you get an answer, and the session ends. Tomorrow starts from zero again, because nothing about the exchange was kept.

Where it lives: Nowhere
How far it reaches: One person, one question

The paste

Most stop here

You keep a scratch document of background and paste it in before the real question. The first time your context exists outside your head, and the last time it is convenient.

Where it lives: A file only you maintain
How far it reaches: One person, every question

The brief

The background becomes a written standard the tool reads every time: how you price, what good looks like, which rules are not negotiable. Now it is a company asset rather than a personal habit.

Where it lives: A shared, owned document
How far it reaches: Anyone on the team

The connection

The system reads your live systems instead of a snapshot, so it works from the job costs entered this morning and the pipeline as it stands today rather than whatever was true when somebody last updated the brief.

Where it lives: The systems you already run
How far it reaches: Anyone, always current

The record

Compounds

Every time a person overrides the output and says why, that reason goes back in. The standard stops being something you wrote once and starts being something the business is still writing.

Where it lives: A record with an owner
How far it reaches: Anyone, and better each month

The move that matters

Every stage is a genuine improvement, which is exactly why so many companies stop early.

Stage two works well enough to feel like success and quietly caps the whole company at whatever one person remembers to paste. The step worth planning for is three, because it is the first one anybody else can use.

Stages one through four each get better once. Stage five is the only one that gets better on its own, because it is the only one where using the system also improves it.

The first two stages are personal. They live in your head or in a document only you maintain, which means the quality of every answer in the company is capped by whoever happens to be asking. The third stage is the first one with leverage, because a written brief gives your newest hire the same starting point as you.

The fifth is the only one that compounds. I have written about what happens to the corrections nobody writes down: the estimator moves the number, the rep rewrites the email, the work goes out correct, and the reason evaporates. A context record is simply the place those reasons land. Do that for a year and the system is meaningfully better than it was, and you did not buy anything to make that happen.

Do you need to train your own AI on your business?

Almost certainly not, and this is the question I get asked most.

Training a model, in the sense of fine-tuning it, teaches it a style or a narrow skill. It is slow, it costs real money, it requires a volume of clean examples most companies do not have, and it goes stale the moment your pricing changes. Context does the thing you actually wanted: it hands a very capable general model the specific facts of your operation, and you can change those facts on a Tuesday afternoon with no engineer involved.

The reason this shifted is structural and fairly boring. Early on, the working memory of these tools was tiny. The 2022 and early 2023 versions held a few thousand tokens, a few thousand words, so feeding a model your standards genuinely was a technical problem. Today the flagship models hold around a million tokens. What used to be a research constraint is now a filing question.

One caution, because the number gets oversold. What a model accepts and what a model reliably uses are not the same thing. It will take everything you hand it, never complain, and bill you for all of it, but attention thins out across a very long input. A short, current, well-organized brief beats a dump of every file you own. The goal is not more context. It is right context.

What actually belongs in the record

The instinct is to point the thing at the shared drive. That produces a confident average of your own inconsistency, which is the garbage in, garbage out problem wearing a new outfit. A context record is curated, not archived.

Four things worth writing down

A context record is curated, not archived. These four earn their space.

01
Decisions, with the reasoning
Not the call you made. Why you made it, in one sentence, while it is still obvious.
We walked away from that job because the schedule risk sat with us, not the owner.
02
Standards, with both examples
What good looks like here, paired with something that missed. The line is learned from the contrast.
This proposal won. This one did not, and here is the part that lost it.
03
Corrections, as they happen
What the output got wrong and why the person changed it. The only input that grows on its own.
Nobody quotes that supplier in Q4. Lead times move and the number goes stale.
04
Boundaries, stated plainly
Where it must stop, who it hands to, and which numbers have to come from a system rather than a guess.
Never send pricing without a human. Pull the rate from the job costing, never estimate it.

What does not belong

Everything else.

Handing over the whole shared drive feels thorough and is not. Ten years of files contain every version of how you used to do it, and the model has no way to know which one you still believe.

Illustrative examples, not real client records. The pattern is what transfers: short, dated, reasoned, and owned by a named person.

Four things are worth the effort, and the fourth is the one people skip. Decisions are only useful with the reasoning attached, because the decision tells the system what you did once and the reasoning tells it what to do next time. Standards need both a good example and a bad one, since a model learns the line from the contrast. Corrections are the renewable input, the only category that grows on its own. And boundaries are what let you actually use the output, because a system that knows where it must stop and who to hand off to is one you can leave running.

The honest limits

Stale context is worse than no context. A brief still carrying last year's pricing will produce last year's pricing, fluently and without hesitation, and a person will send it. Anything you write down needs a name attached to it and a date it gets reviewed. That is the whole governance requirement, and skipping it is how a good system quietly becomes a liability.

None of this substitutes for judgment, either. A written standard shortens the ramp for a new person and gives a machine something real to work from, but it is not the same as the judgment that produced it. Capture makes the standard portable. It does not make it self-aware.

And it is worth knowing where this lives. Your context record is a plain description of how your business decides, which is close to the most sensitive document you own. That is not a reason to avoid building it. It is a reason to know which system holds it and who can read it.

Why the start date matters more than the tool

Take two companies with the same tools, the same budget, and the same industry. One has spent a year writing down why it does what it does. The other starts next quarter. Both get the next model upgrade on the same day for the same price, and it does nothing to close the distance between them, because the upgrade was never the thing that separated them.

This is the part the enterprise world has started saying out loud. At the Gartner Data and Analytics Summit this year, Distinguished VP Analyst Rita Sallam put it as context becoming a cost-control and trust strategy rather than a nice-to-have, and Gartner's published position is that organizations prioritizing this layer will see materially better accuracy at materially lower cost. Translated out of enterprise language: the companies that wrote down what their business means get better answers and pay less for them.

You do not need a data team to start. You need one workflow that matters, the reasons behind how it gets done, and somewhere durable to put them.

Where ClearOak comes in

Most of our work with owners is this work. Not picking a tool, since everyone gets the same tools. Finding the workflow where context is the actual bottleneck, getting the reasoning out of the people holding it, and building the loop that keeps the record current once the novelty has worn off.

If you have tried AI, found it useful in flashes and unreliable in practice, that gap is almost always the context, and it is fixable. Schedule a call and bring the workflow you have already tried to automate once.

Prefer email? Reach me at john@clearoakconsulting.com and tell me what you find yourself re-explaining every single time. That is usually the first thing worth writing down.

Back to Frameworks