It has read almost everything ever published. It has never met your company.
Organisational AI context is everything about how your company actually works that a model would need and cannot find. Your standards, your history, your constraints, and the reason one particular customer gets handled differently from all the others.
Seven enterprises in a hundred say their data is completely ready for AI. Eighty-five in a hundred say they have a clear data strategy, and nearly eighty in a hundred say their AI work is still held back by not being able to get to their own information. Those last two are true at the same time.
Try this first
None of these is a hallucination. Every one of them is a sound, well-argued answer to the question that was asked, produced by a system that had no way of knowing the one thing that decided it.
What was asked
Draft the renewal email for this account.
What came back
A clean, friendly email with a light nudge about the price increase.
What it could not know: that account is three weeks from a merger, everyone involved knows it, and nobody mentions price until that closes.
What was asked
How should we structure the new support tier?
What came back
Three tiers, usage-based, with a self-service layer at the bottom.
What it could not know: you tried the self-service layer in 2024, it cost you two of your five largest customers, and the founder has said once that it is never coming back.
What was asked
Write the onboarding checklist for a new engineer.
What came back
A sensible fourteen-point list that any engineering team would recognise.
What it could not know: eleven of the fourteen are already automated here, and the three that matter are the ones nobody has ever written down.
In each case the person reads the answer, thinks that is about right, and quietly fixes the missing part themselves. That works, and it means the correction lives in one person's head and has to happen again next week.
Out in the wild
Left is what any model can see about a company like yours. Right is the part that decides almost every real answer. Nothing on the right is secret. It is just not written down anywhere a machine can get to it.
Defining it
Most companies, when they hear that their AI needs context, point it at the shared drive. Then they are disappointed, and they conclude the technology is not ready. The drive holds outputs: the finished proposal, the signed contract, the published policy. What decided any of it is nowhere in there.
Take one example you will recognise. Your proposals all look similar and one of them was priced eleven per cent higher, and the reason is a conversation somebody had in a car park. A model reading twenty proposals sees the pattern and averages it. A person who has been here four years sees the exception and knows why. The difference between those two is the whole of this page.
The drive is full of what you decided. The reasons are in three people, and two of them are busy.Paul Musters
There are three layers of it, and companies usually have the first and almost never the third. The facts: numbers, customers, contracts, mostly in systems already. The method: how work actually gets done here, which is partly written and mostly not. And the judgement: what you would refuse, what you have already tried, and where the line sits between fine and not fine. Only the first layer is a data problem. The other two are a writing problem, and they are the ones that make an answer feel like it came from inside the company.
Which is why this sits where it does on the AI Culture Levels. Writing down how the work is really done is Level 3 behaviour, and it is the first level at which a company has anything of its own to hand a machine.
Below that line, every person is prompting from their own head, so the quality of an answer depends on who happened to ask. That is survivable while AI is a side activity and it stops being survivable the moment an agent produces something a customer sees.
What the research says
Two surveys from the same company, six weeks apart, tell the story better than either does alone. The first was run by Harvard Business Review Analytic Services with Cloudera among more than 230 people involved in their organisation's AI data decisions.1 The second polled 1,270 IT leaders at companies with more than a thousand employees.2
Cloudera has a name for the middle one and it is a good name: the AI readiness illusion. Eighty-five in a hundred have a strategy on paper. Four in five are still blocked by the thing the strategy was about. Nobody in that survey is being dishonest, and the two answers sit in different parts of the same head.
The first study asked what makes preparing data for AI hard. The answers are worth reading in order, because the top one is not a technology problem in the way people expect.1
Siloed information at the top, and the absence of a strategy second. Those are both organisational answers wearing technical clothes. Somewhere in your company two systems hold overlapping versions of the same thing and nobody has ever had the argument about which one is right, because until now nothing forced it.
All of the above is about data. The harder half is method, and Microsoft measured it. Asked whether agent workflows, human handovers and quality standards are documented and repeatable at organisation level, fourteen in a hundred said yes. Among the most advanced AI users in the same survey it was twenty-five in a hundred.3
Read that next to the seven per cent and you have the honest picture. The facts are in systems and hard to join. The method is in people and not written at all. A company that fixes only the first one ends up with beautifully governed data and answers that still do not sound like it.
Two cautions on the evidence. Both Cloudera studies were commissioned by a company that sells data platforms, and the questions are framed accordingly. And the second one surveyed companies with more than a thousand employees, which is well above the size emaho usually works with. The direction is useful, the exact percentages are about somebody else's company.
How it starts
Every company documents what it decided, because that is what a contract, a spec and a policy are. Almost nobody documents why, because until recently the why lived comfortably in whoever made the call and you could just ask them. That worked for decades. It stops working the moment the thing doing the work cannot walk over to their desk.
The context you need most is concentrated in three or four people, and they are three or four people precisely because they are the ones everything routes through. Asking them to spend a week writing it down is asking the busiest part of the company to stop. So it gets scheduled, moved, and quietly dropped, every quarter.
The person who could write it down is the person you cannot spare, which is how you know it is worth writing down.Paul Musters
Somebody gets a slightly generic answer, spends four minutes correcting it, and moves on. Four minutes is nothing. It is also invisible, uncounted and repeated by eleven people a week, and every one of those corrections is a piece of context that existed for a moment and was then thrown away. That is the mechanism that keeps this exactly where it is.
Three minutes, six questions, nobody sees your answers. Your level, what it costs you there, and what changes one up.
How to see it
Take a real task you would hand somebody in their first week. Write the brief exactly as you would write it for a person, with the same amount of background, and give it to a model instead. Then compare what comes back with what a good new hire would produce on day three.
Then the task genuinely does not need your context, and you have found something worth handing over permanently. Most companies have more of these than they expect and use fewer of them than they should.
Generic, plausible, missing the local rules. This is the useful result. The gap is your onboarding, not the model, and whatever you would have told the new person on Thursday is the first context worth writing down.
It contradicts something everybody here knows. That points at a specific piece of judgement, usually one sentence long, that has never left somebody's head. Write that sentence down today.
The second outcome is the common one, and it is oddly good news. It means the missing thing is knowable and writeable, and that the same gap has been quietly costing you on every new hire for years.
Run it on three different tasks from three different parts of the company before you conclude anything. One is an anecdote, and three is a pattern you can point at in a meeting.
Level by level
There is a version of this that fails, and it is the one where a company spends six months building one enormous knowledge base that everybody is supposed to use. It gets built, it gets stale, and within a year two people maintain it and nobody reads it.
An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.
The version that works is smaller and closer to the person. What a salesperson needs the machine to know is not what a controller needs, and neither of them will maintain a document written for the other. So we measure two things per person, how somebody thinks and works and how far along they are with AI, and then build their agents with the context that fits their actual week inside them.
That also solves the maintenance problem, which is the one that kills every knowledge base. Somebody keeps their own agent honest because it is theirs and because a wrong answer costs them the afternoon. Nobody has ever felt that way about a shared wiki.
emaho measures one Operating Profile per person: personality type and AI level in a single profile. On that we build a personal set of AI agents that fit how that person works, inside the tools they already use. Fifteen minutes to complete, first profile free, built for companies between 20 and 500 people.
Fifteen minutes per person. No credit card, no strings.
What to do
There is no project here and there does not need to be one. Your company already produces the raw material for this every week, in the form of the same explanations given over and over to different people.
Ask three or four of the people everything routes through to keep a note for two weeks. Every time they explain something that is obvious to them and was not obvious to the person asking, they write one line. Not a document, a line. At the end of two weeks you will have somewhere between thirty and sixty lines, and about ten of them will come up more than once. Those ten are your context, and they were free.
Write those ten properly. One paragraph each, in plain language, including the reason and the exception. The reason is the part that matters, because a rule without its reason gets applied to the case it was never meant for. Then put them where the tools your people actually use can reach them, which for most companies means a single file rather than a platform.
Keep it alive with the same trick that produced it. Whenever somebody has to correct a machine answer by hand for the second time, the correction becomes a line. That turns the four wasted minutes into the only maintenance mechanism that has ever survived contact with a busy quarter.
And one thing to leave alone for now. Do not begin by pointing anything at the shared drive. Fourteen in a hundred companies have their way of working written down at all,3 so for most of them the drive contains outputs and no reasoning. Indexing it produces a confident machine with the same blind spot and a larger vocabulary.
If you run the two-week note, send me three of the lines that came up more than once. Those three usually say more about a company than an org chart does, and I will tell you what I would do with them. I answer myself, usually the same day, and that is the end of it.
The cost of waiting
The visible cost is a tool that gets quietly abandoned. You buy something that promises to work with your own documents, it produces generic answers because the documents contain outputs and not reasoning, and within four months people go back to doing it by hand. Then the story inside the company becomes that this stuff does not really work here, which is expensive to undo.
Underneath sits the correction tax. Every generic answer costs somebody a few minutes and those minutes are never counted, never reported and never fixed, because each one is too small to raise. The same eleven corrections happen every week and the knowledge they contain is thrown away eleven times.
And there is the one that arrives on a Monday. One of the three or four people who hold the reasoning leaves, and what walks out with them is the answer to why, on about forty separate things. You find out which forty over the following year, one uncomfortable meeting at a time.
The common ones
Close by
All three are what happens when the machine has to answer without knowing anything about you.
The same gap seen from the receiving end: advice that is sound in general and wrong for you.
Read this one 20What happens across a whole market when nobody has anything of their own to put in the prompt.
Read this one 15What else goes missing when the way work really happens was never written down.
Read this oneTwenty-five things that break inside a company once people start using AI, each with the research behind it and the level where it starts to bite. This one starts to bite at Blueprint, and 7 of the others start there too.
Getting going
A single company-wide knowledge base goes stale in a year. A profile and a set of agents per person stays current, because the person who uses it is the person who fixes it.
First profile free · no credit card · built for companies of 20 to 500 · you decide what your team gets to see
Not ready to put your team in anything yet? Start with the level of the company instead. The Culture Level scan is six questions, three minutes, and asks nothing of you.
Fifteen years of leadership and team development in Dutch scale-ups. That practice now sits in software: one Operating Profile per person, with agents that actually fit. He writes these pages from what he runs into with clients, not from a research summary.
Numbers are quoted as published. The three examples at the top and the two-column comparison are composites drawn from client situations rather than transcripts. The share of companies per level comes from emaho's own work with clients.
