Ask an agency owner what their most valuable system is and you will hear about the CRM, the project tool, maybe the pipeline spreadsheet. Ask what actually runs the place and the honest answer is a skull.
I know because it was mine. For twelve years at This Page Amsterdam, the standards lived in my head and my partner's. What good work looked like, what this specific client would accept, why we had decided against that approach in 2014, which promise had been made in which meeting: none of that was in a tool. The agency had thirty people and two runtimes, and every important deliverable passed through one of them before it left the building.
Founder dependence is the most diagnosed condition in the agency industry, and the fixes have been the same for twenty years. Document your processes. Build SOPs. Delegate. Hire the second in command. I tried versions of all of them, and I watched dozens of other founders try them, and the pattern is remarkably consistent: the documentation gets written, goes stale within months, and the founder quietly remains the system of record. The number two absorbs some load and still walks into the founder's office for the judgment calls, because the judgment never made it into the documents.
For a long time I read that pattern as a discipline failure. Founders who would not let go, teams that would not read the wiki. I no longer think that. I think every one of those fixes failed for the same structural reason, and the reason is worth stating plainly: they all asked the founder to write things down for humans, and written-down things are dead the moment the business moves.
What the founder's head actually does
Watch what the skull-shaped system really provides, and it is not storage. Anyone can store things. A wiki stores things beautifully.
The founder's head does four jobs no document ever managed. It stays the same expert over time: the standard applied on Tuesday is recognizably the standard applied last March, without anyone reloading it. It knows what is currently true: that the client contact changed, that the pricing changed, that the old rule about revisions was superseded in the autumn, and it does not accidentally serve the 2023 version of any of those. It remembers what was decided, as distinct from what was merely discussed, and it treats decisions with more authority than table talk. And it knows which client it is standing in front of, so the judgment arrives scoped to this relationship, this history, this scope of work.
Continuity, in other words. Not recall. A wiki has perfect recall and zero continuity, which is why the wiki never became the founder.
Writing about AI memory earlier this year, I gave names to what happens when systems carry recall without continuity: identity drift, stale truth reuse, outcome amnesia, scope confusion. I wrote those as failure modes of AI systems. It took me embarrassingly long to notice they are also a precise description of an agency the week the founder goes on holiday. The team's output drifts off-standard. Superseded rules come back from the dead. Decided questions get relitigated. Two people give the team different answers and nobody decides which one is true. Client boundaries blur. The failure modes are identical because the missing thing is identical.
Why the chatbot is not the answer either
The obvious modern move is to point AI at the problem, and the obvious version of that move fails in an instructive way.
Give an assistant your documents and chat history, and you get recall: it can find what was said. Recall reproduces the wiki's limitation at higher speed. I have lived this in my own AI-heavy practice: sessions where a decision I had closed weeks earlier came back on the table, weighted the same as every live option, because the system remembered the discussion and had no concept of the decision. It remembered everything and knew nothing. A system like that does not relieve founder dependence. It adds a very fast intern to an agency that already had plenty of information and a shortage of judgment.
The distinction that matters is the one between what was said and what is settled. An agency runs on settled things: this is our standard, this is what we agreed with this client, this is current, that is superseded. Settled things need to outrank discussion, stay versioned as they change, and carry their history with them, so that when the rule changes, the old rule becomes historical instead of wrong-but-still-findable. And when two credible people disagree and nobody has decided, the honest state is not a guess. It is an open dispute, shown as one. None of that is retrieval. All of it is governance, and governance is precisely the thing the founder's head has been providing for free, invisibly, at the cost of being unable to take four consecutive weeks off.
The agency brain
So here is the claim, stated as directly as I can make it: the fix for founder dependence is not a document, not a hire, and not a chatbot. It is an expert system at the center of the agency's operations, holding what the founder's head holds, in the way the founder's head holds it. The standards as governed, versioned things. The decisions as records with authority above discussion. The current truth as current, the old truth as history. The client work scoped, so an answer knows which relationship it belongs to. I have started calling this an agency brain, because that is what it is: not an assistant bolted onto the side of the business, but the continuity of the business itself, finally living somewhere other than a commuting skull.
That is the destination, and I should say before describing it that I am building toward it, which you should read as a declaration of bias as much as a declaration of intent. In the version I am working toward, new people would learn from it instead of apprenticing on interruptions. Delivery would get checked against it before the founder's door, not instead of standards but as the standards. And the founder's judgment stops being the bottleneck without stopping being the source: the head still decides what becomes settled. It just stops being the only place settled things can live.
What I will say from inside the work is that the hard part is not the technology, and it is not storage. The hard part is the governance: deciding what earns the right to become durable, who validates it, and how truth gets retired when it stops being true. Which is fitting, because that was always the actual job. The founder was never really the agency's memory. The founder was the agency's judgment about what deserved remembering, and that job does not disappear into the system. It finally gets a system worthy of it.
The system has a name now, Constance, and a whitepaper that makes the argument in full, including what is not built yet. Both are at trustconstance.com. Its one rule is the whole idea in a sentence: no claim becomes truth without a decision a person owns. The brain holds what was decided. It still does not ask the questions that lead to deciding; that part stays with the person in the seat.
The agency was always run by a brain. It just happened to be biological, impossible to back up, and on holiday twice a year.