Fifteen years of articles, documentation, and release notes were not enough to make our knowledge usable by AI. The problem was never the volume of text: it was saying what counts as true today. Here is how we built that reference, how we query and maintain it, and what we refuse to delegate.
At GoodBarber, a Mobile App Builder where I’ve led the marketing team since 2011, our first uses of AI were the obvious ones: finding ideas, rewording a paragraph, summarizing a document. The wow effect was real. The effect on how we actually worked was much smaller.
The turning point came for us in late 2025. Models were getting sharply better, agents were starting to work from durable context, and our experiments could finally become a production system. It was no longer enough to get a good answer in a conversation. We needed to get, weeks later and across several people, a result that was consistent, verifiable, and owned.
During my research in late 2025 and early 2026, as we worked out how to scale AI, I kept seeing the same pattern: marketing usage trails engineering by about three months. Watch what developers do today, and you’re reading marketing’s near future. Mathieu, our Head of Frontend, has just written that near future from his own seat: what AI shifted in an engineer’s job. From that, a simple scale:
Assistance: the human produces, and AI helps here and there to search, reword, or critique.
Documented production: AI produces from identified sources, then a human checks, decides, and signs.
Supervised automation: certain bounded changes propagate under explicit rules and controls, with a named person accountable.
We wanted to reach the second level. And the first project was not choosing a tool.
Why fifteen years of writing wasn’t enough for AI
The seven language editions of our blog add up to thousands of articles. Our knowledge base is thick. Our release notes tell fifteen years of product history. Text was never the missing piece.
But all of that text is production, written for readers and customers. It doesn’t necessarily contain the knowledge that governs it: the voice we want to use today, the proof behind each claim, the rules we never break, the customers we’re writing for, what still holds and what we abandoned along the way.
That knowledge lived mostly in heads, habits, and veterans’ reflexes. An oral tradition, sitting on a mountain of text.
And an AI doesn’t read oral tradition. The mountain of text can’t replace it either, because fifteen years of content forms strata. Positionings succeed one another, prices move, arguments die. Plug a model into the whole thing without telling it what counts, and it averages the eras. It resurrects an abandoned tagline, quotes a three-year-old price, blends the 2015 voice with today’s, all with perfect confidence.
So the problem was never writing. Writing was all we had done. The problem was telling the archive apart from the current truth. We had written everything, except ourselves.
Turning oral tradition into an AI-ready knowledge base
We built a shared, versioned marketing reference, what developers call a repository. It doesn’t claim to hold every answer itself. For each subject, it says what counts as authoritative and where to read it.
It holds our voice and tone, the customer profiles we write for, every proof we’re allowed to use, and our guardrails, the list of things we never say, whatever the temptation. Data that moves, like pricing, is not copied into a document: the reference points to the system that serves it live.
This is the context the AI reads before it works. It no longer has to guess our positioning from texts written in different eras. It knows which sources to consult, which claims to verify, and which lines not to cross.
The surprising part is that extracting this knowledge was not the chore I expected. The unlock was reversing the roles: before asking AI to help me write, I asked it to interview me.
For over an hour, it asked questions under one explicit instruction: contradict my answers, dig into every “that’s obvious” and force me to explain the rules I had stopped noticing. I often ask my AI to be my adversary; that’s when it’s most useful to me. I answered out loud, with Wispr Flow, the way I would have talked to a journalist.
That conversation surfaced distinctions and guardrails that no brief written alone would have contained. Implicit knowledge rarely shows up on demand. It shows up when someone asks one question too many.
The reference also produced a side effect I hadn’t planned for: it changed the team’s culture. To keep it alive, colleagues who had never touched Git learned branches, pull requests, and review before publication. Engineering practices walked into marketing. And the traffic runs both ways: my colleague Muriel taught herself to build with AI and drew lessons from it I couldn’t have written. The borders between crafts are getting more porous, not because everyone does the same job, but because each side is learning to handle part of the other’s tools. The crafts are trading tools.
Fix the source, not just the output
If we had stopped once the reference was written, we would have had a fine piece of documentation, which is to say a dead document. What holds the system together is the feedback loop.
Every piece of content tests the base. If the AI quotes an old price, we fix the sentence, but above all we ask why it didn’t consult the pricing source. If it produces an unverifiable claim, we add the available proof or ban the claim. If its voice goes generic, we enrich the reference with a rule or a counter-example.
When an output is wrong, fix it. When it’s wrong for a reproducible reason, fix the source.
Through use, we identified what I call our backbone documents: the positioning, the proofs, the guardrails, and the voice rules that everything else depends on. We maintain them like production code, because an error there propagates into everything produced afterwards.
Quality is no longer a piece-by-piece affair. You don’t just proofread the texts; you maintain what generates them.
Automate the production, not the responsibility
The one thing the reference cannot carry is responsibility.
For now :), nothing ships here without a named human owner. AI proposes drafts, and the first draft is always the hollow one: it has the shape, rarely the idea. Someone has to decide whether the text deserves to exist, critique it, rewrite it, and agree to sign it.
In full transparency, the text you are reading was born exactly this way: an AI working from our reference, an interview to extract what I actually thought, several versions critiqued and rewritten, then my signature, with everything it commits me to. If the input and the iterations are good, the final piece is too. That’s our method in one sentence.
The reference reduces repeatable errors. It replaces neither judgment nor signature. On what humans keep as their own, judgment and taste in particular, my colleague Mathieu just wrote it better than I could. I would only add one team-lead’s nuance: what we kept isn’t a set of tasks. It’s a set of responsibilities.
Fifteen years is only an advantage if it’s legible
Fifteen years of history bring experience, proof, and reflexes that a younger company cannot fake. They also bring contradictions, habits gone invisible, and truths that quietly stopped being true.
AI doesn’t turn that memory into an advantage on its own. For that, you have to tell what lasts from what changed, make the choices explicit, and maintain the sources that carry them.
The established companies that get the most out of AI won’t simply be the ones that adopt the tools fastest. They’ll be the ones that know what they know, what counts as true today, and who answers for what goes out.
AI didn’t change what we know. It changed what we had to write down.