One thing is easy to get wrong these days: the idea that the longer you use a tool, the better it gets to know you.
Three months ago, your team spent an afternoon walking AI through the company's pricing logic. Which kinds of customers can get how much of a discount, which accounts always need the owner's sign-off, which product line's margin can't be touched. That day it answered well, so well that someone in the meeting room said this thing is actually usable. Three months later, you ask the same question again, and the answer is so generic it's as if you never talked. The discount suggestions are boilerplate, the customer tiers are ones it made up, and it doesn't mention the margin line at all. That afternoon in between might as well never have happened.
Nobody did anything wrong. It's just that what was said that afternoon was stored somewhere that gets closed.
01It never remembered in the first place
There's a premise that often gets skipped here: it was never meant to remember for you. The model hasn't gotten dumber, and you didn't buy the wrong tool.
There's a chart in the industry that splits AI memory into six levels, from L0 to L5. At L0, every time is the first meeting: you repeat the question, repeat the explanation, and it forgets everything when the session ends. L1 remembers this conversation; close the window or switch devices and it's gone. L2 remembers your preferences and who you are, like a user profile, and can still pull them up when you change topics. In the L2 box, the chart has a line in red: ninety percent of products stop here, yet advertise that they have long-term memory.
What you actually want is further up. L3 is called semantic memory: knowledge settles in, and the same body of understanding can be called back across tasks and conversations. L4 remembers the full story behind specific events and can answer "why did this happen." L5 can adjust what it does next time on its own, based on past results; the chart labels this level "a true agent." The same chart also comes with a self-check question for each level, and the L3 question is the one best suited to asking about where your company stands today. When it moves across tasks, can it reuse knowledge from before?
In the bottom-right corner of the original chart there's a note worth copying down:
The value of memory isn't in how much gets remembered, but in how the past takes part in shaping the future.
Translated into a version an owner would recognize, it goes roughly like this: the AI you're using now probably sits at the level of "remembers this meeting," while what you want is for a new colleague, three months from now, to ask and get that day's conclusions. What sits in between is a layer nobody has built yet, and it has little to do with how strong the model is.
02You're buying uses, not capability
Run the numbers from another angle.
Suppose a twenty-person company gets paid accounts for everyone. In the first month, people really do save time: emails get written faster, slide drafts appear, even complaint replies no longer start from scratch, and the owner looks at the report and thinks it's worth it. By month six, the bill is the same, but it doesn't feel any better. The reason isn't hard to find. The opening pitch sales worked out, admin doesn't know about; the payment-request write-up admin figured out, finance has never seen; and when the colleague who was best at asking leaves, the prompts they spent six months refining walk out the door with their exit paperwork.
Someone has explained this in ten hand-drawn pages, more bluntly than anyone, under the title "Every company is using AI now. So why isn't it making the boss more money?" It shows three employees each using AI to write copy, make images, and rework sales scripts. Every one of them got faster, yet an equals sign links them back to the owner's profit statement, which has barely changed. The author calls this "spot" AI use: individual actions get faster, the business process as a whole stays the same, and what gets stuck is always the same three things — scattered data, experience that lives in people's heads, and post-mortems whose lessons never get captured.
So the crux is somewhere else. The time those twenty people saved is real. It's just that not one of the methods that saved it stayed with the company, and in month seven everyone still has to start asking from scratch.
03"We have a knowledge base" isn't the same as having answers
At this point most companies will say: we have a knowledge base. Shared folders, an internal wiki, cloud drives, thousands of files, reasonably well organized. That really is a good thing. It's just answering a different question.
One breakdown of why enterprise AI never gets off the ground lays this out as a table. The left column is what a knowledge base does: which folder the data lives in, whether a keyword search can find it, who uploaded it and who's responsible. The right column is the four answers AI still needs before it can be put to use, and most companies can't give a single one.
The same thing has three documents. Which one do you trust? The old version of the pricing policy, the revised version, and the cheat sheet sales put together all sit in the folder at once. A person picks the latest one from experience. AI doesn't. It reads all three, then blends them together in its answer. The breakdown gives an example that's easy to picture. The owner asks which customer was slow to pay last month, AI gives a convincing-sounding answer, and when someone checks, all of it is wrong. It didn't actually make anything up. Sales, finance, and the owner's own reports were already using three different definitions, so there were three different numbers. Whichever one it picked wasn't wrong, but none of them could be taken into a meeting.
What can this person see? When nobody manages permissions, the only thing people dare use it for is small talk. Salary structures, customer lists, cost breakdowns: as long as nobody has explicitly drawn the line, the usual outcome is to lock it all down and call it a day, and the rollout stops right there.
Is this rule still in effect? Last year's shipping process was changed once in March this year, then changed back in June. The files are all still there, and none of them is marked obsolete. A person knows to ask. AI doesn't ask.
How do you keep the veterans' experience? Someone who has been in purchasing for fifteen years knows which supplier always delivers late in the rainy season, and that isn't written in any file. Our July piece, "Your AI Has No Raw Material," was entirely about this gap: what's left in the system is only the final conclusion, and nobody wrote down how it was reached.
If these four questions can be answered one by one, the layer you can draw on exists. If they can't, those thousands of files are still just a very neatly kept filing cabinet.
04The difference between a filing cabinet and an answer
The filing-cabinet metaphor comes from another write-up, about why nobody uses the knowledge base. It opens by putting in a fair word for employees: they actually did search, it's just that what they found didn't look like an answer.
Its example is an everyday one. An employee searches "expenses" and gets four files: Expense Policy V1.0.pdf, Expense Process Flowchart.jpg, Expense FAQ.docx, Travel Expense Guide.pdf. They still have to read them all, decide which one is valid, and piece the steps together on their own. Too many documents to know which to read, content too long to find the point, formats so inconsistent it's a headache just to look, and still not sure after searching. At this point, most people choose to @ the colleague who's sure to know.
And the version that looks like an answer goes like this. Yes, you can claim it. The process is: submit an expense form, upload the invoice and receipts, department approval, finance review, payment. An expense form template is attached, with a one-click copy button next to it.
Same data, two forms. The first leaves the final step of pulling it all together to the person; the second has already done it. What the "layer you can draw on" does is move that act of pulling it together from every time someone asks to the moment the knowledge gets written in.
Engineering has a set of names for this, and the one you'll hear most is RAG. But a name doesn't help an owner; something you can test does. Our August piece, "The Model Is Rented. So What Do You Actually Own?", used a criterion that happens to apply here too: with a brand-new conversation and a brand-new colleague, relying only on the files you've left in the company, can the job be taken over and finished? If the answer is no, that layer hasn't been built yet.
05Where to start, and where not to
This layer sounds big. In practice, it isn't a project that takes six months to build.
A guide to organizing this for businesses lays out the order in five steps. First, gather existing material: product descriptions, service processes, FAQs, internal policies, and past tickets all count. Next, clear out duplicate, outdated, and contradictory content. Then sort it by business scenario; a folder tree split by file format won't help. Then set access permissions, making clear which roles can look things up and which can edit. Finally, set boundaries on AI's answers: don't make things up without a basis, say so when it can't tell, and pass important matters to a person to confirm.
The order can't be reversed. Opening up queries before permissions are set is handing AI the keys to every door in the company.
Keep the scope small, too. The breakdown of enterprise AI mentioned earlier gives this rhythm: diagnose and plan first, then pick one lighthouse scenario and push it all the way through with real data until it produces results you can check, and only then copy it to other roles. Doing it the other way around almost always means rework. For small and mid-sized businesses, the lighthouse scenario is usually the ten questions customer service gets asked most, or the pricing policy sales looks up most often. Run a round with real questions, get it running smoothly, then move it to other departments. Going company-wide with everything from day one usually stalls at the data-cleanup stage, then quietly fizzles out.
The same breakdown also names three kinds of things to set aside for now: those where the rules themselves haven't been settled, those where the data is still on paper, and those where a mistake can't be rolled back. It also gives three one-minute self-check questions, well suited to reading aloud at your next management meeting. How many definitions does the same number have? What data can AI see right now? Is there a record of what it has done?
Before A company's assets were its equipment, its customer list, and a set of processes that ran
Now There's one more, and you can't see it: a layer of knowledge others can ask, AI can draw on, and departing staff can't take with them
Back to that afternoon
Our August piece, "AI Can't Read Your SOP," was about what shape the raw material needs to take; the July piece was about how the raw material was never in the files to begin with. This one is about where it goes once it's ready. Get the location right, and everything after it can build up.
Back to that afternoon at the very beginning. If that day's conclusions had been written back somewhere the company owns, asking again today wouldn't get you a generic answer. It would get you one more accurate than three months ago, because in between there were a few more closed deals, a few exceptions, and a few discounts turned down, all stacked on the same layer.
This is also what we're really looking at when we run an AI workflow diagnostic. We look at whether you can answer those four questions, and whether the answers are written in files or live only in someone's memory. Which model you use isn't really the point.
Let AI move fast. Let people hold the anchor. Tools can be replaced. What you've built up can't.
Rooted in craft. Built for the new wild.