What happens when the content is AI-written
A system that answers questions about your business has to be given something to answer from. The temptation, and it is a strong one because it is fast and free, is to generate that material with AI.
Do that and the system is reasoning from its own output. Ask it something specific and it will answer from generic text that was written to look like your business rather than from your business. You get generic answers delivered with the confidence of specific ones, which is worse than an obviously wrong answer, because an obviously wrong answer gets caught.
It is a quiet failure. Nothing errors. The dashboards populate. The answers are fluent and plausible and shaped correctly, and they are the same answers any business in your industry would get, because that is exactly what they were built from.
The knowledge only your people have
The material that makes the difference is not written down anywhere yet, which is precisely why no model has it. It is the reasoning your people apply without noticing they are applying it.
- Why you price that job type the way you do, and the one case where you do not.
- Which supplier you go to when the timeline is tight, and what it costs you to do that.
- The three questions you ask a new enquiry that tell you whether it is worth quoting.
- What went wrong on the job everyone still refers to, and what changed afterwards.
- The rule somebody made eleven years ago that everyone still follows and nobody has written down.
None of that is in a manual. Some of it is not in anyone's head completely. It exists across four people who have each got part of it. Getting it out is an interviewing problem before it is a software problem, and it is the part that cannot be bought in.
It comes out in corrections, not in an interview
On one engagement the knowledge that made a stack of documents checkable at all was in one person's head, and it did not come out by being asked for. It came out as twenty-seven rules over six weeks, each one written the moment the system got something wrong.
They are not the rules anybody would have thought to write down:
- One product code covers two positions rather than one, so a count that looks short is correct.
- Drawings three months old and stamped preliminary are normal on a live job, not an error to flag.
- One whole category of order never needs checking at all.
- Arithmetic mistakes are real and nobody actions them, because sign-off happens somewhere else entirely.
And one the reviewer had never thought to check himself: nothing in his process compared a supplier's required-by date against the install date. On a real job, two orders were due after the installation, one of them by a month.
What that engagement actually taught us
That the model was not the product. The twenty-seven rules were the product, and the model was the way of getting them out of somebody's head and into a form the business owns. That is the same claim this page makes, arrived at the expensive way.
What we can extract, and what we cannot
A fair amount can be recovered from what you already have. Quotes, job records, email threads, the spreadsheets somebody maintains privately. Those carry a great deal of real decision-making, and they are the first place to look because they cost you nothing to provide.
What cannot be recovered that way is the reasoning behind the exceptions, and the exceptions are usually where the value is. That part comes out of conversation with the people who hold it, and there is no way to do that without their time.
The size of that ask, before you agree to it
It is a real ask. It is hours from people who are already busy, and it lands at the start of the project rather than the end. We would rather say so now than discover it together in week six.
What this asks of you, in hours
Roughly: a handful of structured conversations with the people who actually make the decisions, and someone on your side who can say which version of a disputed answer is correct. Not a committee. Not a documentation project.
If that is not available, if the people who hold the knowledge genuinely cannot be released for it, then a connected system built on your own material is not the right thing to buy yet, and we will say so. A smaller piece of work on one process is a better use of the same money, and it is a real answer rather than a polite one.
What it looks like when it is done properly
The Kaima website is the example we can show you, because it is ours and needs nobody's permission.
Every page here traces to a recorded founder interview. The transcripts are kept. Each page carries what it is allowed to claim and what it is not, and the pages we could not yet write are marked as unwritten, not filled with something plausible. Where a fact behind the plan turned out to have expired, the correction is recorded alongside the original rather than quietly replacing it.
That is the same discipline applied to a website rather than to a business system, and it is why we are willing to make the claim on this page at all. We are asking you to do something we do ourselves.