What an agency brain can do, and the four things it cannot

Everybody arrives at this conversation with the same question, and almost nobody asks it out loud. It is not “how does it work”. It is “what would this actually do for me, specifically, on Monday”.

Fair question. Here is the full answer, including the parts that do not work, because a capability list without limits is a brochure and you have read enough of those.

If you want the definitional version first, we wrote the definition we published on the Shortlist blog and it covers what an agency brain is before it covers what it does. This post assumes you already accept the premise and want to know whether it earns its keep.

The short version

An agency brain reads everything your agency produces and receives, keeps it, and makes it available to agents that do work on top of it. In practice that splits into four things it does well and four it does not.

What it does well: retrieval, drafting against real history, watching things continuously, and answering questions about your own business that currently require asking a person.

What it does not do: judgement, relationships, anything requiring accountability, and any process you have not actually defined.

The rest of this post is that list with the detail attached.

One: it retrieves

This is the boring one and it is the one that changes your week.

Right now, when someone asks what we agreed with a client about their Q3 budget, the answer involves a person remembering roughly when the conversation happened, searching email, failing, searching Slack, finding a partial version, and then asking someone else. Fifteen minutes, two people, and a fifty percent chance the answer is wrong.

With a brain, that question is answered from the record, with the source attached, in the time it takes to type it. Not because retrieval is glamorous, but because your agency asks itself questions like that forty or fifty times a day and has never counted the cost.

The thing that makes this work is coverage. A brain connected only to Google Drive is a search box. A brain connected to meetings, email, chat, project management, and your platform data can answer questions that span all of them, which is where real questions live. Most of those connections now run on the open standard for connecting tools to one system, which is worth knowing about if your developer asks. Our post on what actually connects and what each source contributes goes through the stack one system at a time.

There is a second-order effect worth naming. Once retrieval is reliable, people start asking questions they had stopped asking. Nobody chases a fact that takes fifteen minutes to confirm, so they guess instead, and the guess goes into the client email. When confirming takes ten seconds, the guessing stops. We noticed this about two months in and it was not on anyone’s list of expected benefits.

Two: it drafts against real history

Every AI tool drafts. The difference is what it drafts from.

A generic tool drafts from the prompt you gave it. A brain drafts from the prompt plus a year of this client’s previous reports, plus the two calls where they told you what they care about, plus the recommendation they rejected in April, plus the tone they responded well to.

The practical effect is that the first draft lands closer. Not perfect. Closer. In our own agency the difference showed up as fewer rounds rather than faster rounds, which is a better outcome and a harder one to notice on a timesheet.

This applies to monthly reports, content briefs, proposals, meeting follow ups, and internal handovers. It applies badly to anything genuinely new, where there is no history to draft against, and you should expect a brain to be no better than a good generic tool on a first-time-ever piece of work.

One warning on this. A brain drafting from history will reproduce whatever is in the history, including your bad habits. If your reports have always buried the uncomfortable number on page four, the drafts will bury it on page four. Automating a process locks in its current shape, which is an argument for fixing the process before you automate it rather than after.

Three: it watches

Humans check things on a schedule. Systems check things continuously.

Rankings, uptime, Core Web Vitals, budget pacing, a page that lost its canonical tag, a competitor that just published forty pieces. A brain plus monitoring means the gap between something changing and someone knowing about it goes from a fortnight to about a day.

The catch, and it is a real one, is that watching produces alerts and alerts produce noise. We got this wrong early on. Tuned too sensitively, we generated enough false positives that people stopped reading. The fix was not clever: fewer, higher thresholds, and a human reading anything that goes to a client. Watch too much and you have built an expensive way to be ignored.

Four: it answers questions about your own business

This one surprises people, so I will be specific.

Which clients have we quoted video work for in the last two years, and what happened. Which of our proposals converted, and what did the winners have in common. How long does onboarding actually take, measured from first email to first deliverable, across our last thirty clients. What did we tell this client we would do in January that we have not done.

Every agency owner has wanted the answer to at least one of those and given up because getting it meant a week of someone’s time. When the history is in one place and queryable, those become questions you ask in the ordinary course of a Tuesday. That changes what you can manage, which is a different kind of value from saving hours.

There is also a quiet governance benefit here that nobody sells on. When the record is queryable, “we never agreed to that” becomes a checkable claim rather than a standoff between two people’s memories. That has resolved two scope disputes for us in the last year, both in about four minutes, both in a way that would previously have cost a difficult call and some goodwill.

The four things it cannot do

Now the part that matters more.

It cannot exercise judgement

A brain can tell you organic traffic fell eighteen percent, isolate it to two page templates, and note that the drop began the week of a deployment. It cannot tell you whether to escalate to the client’s CTO, absorb the fix inside the retainer, or use it as the moment to have the wider conversation about their technical debt.

That is not a limitation that goes away with a better model. The Harvard and BCG field experiment on where AI helps and where it quietly hurts found people were meaningfully better on tasks inside the model’s capability and measurably worse just outside it, without being able to tell which was which. Knowing where you are is the human part. It is a decision with commercial and relational consequences, made by someone who will answer for it. The whole design of a serious rewire assumes a person is standing there.

It cannot hold the relationship

Clients stay with agencies because of people. That has been true for as long as agencies have existed and none of this changes it.

A brain can make sure you never forget a commitment, never miss a follow up, and never ask a client to repeat something they already told you. Those things make the relationship easier to hold. They do not hold it. If your retention problem is that nobody at your agency has a real relationship with the client, automation will make you efficiently forgettable.

It cannot be accountable

When something goes wrong on an account, a client wants a person to explain it, own it, and say what happens next. “The system flagged it” is not an answer, and an agency that starts giving that answer has begun quietly transferring risk onto its clients.

This is also why every publishing action in our own setup requires a named approval. Not because the drafts are bad. Because accountability has to attach to a person, and the only reliable way to make that true is to require a person before anything leaves the building.

It cannot automate a process you have not defined

This is the one that stops most projects, and it is the least technological problem on the list.

Most agency processes are not processes. They are habits, performed slightly differently by four people, with the real logic living in whoever has done it longest. You cannot automate that, and the attempt produces something that is confidently wrong in ways nobody predicted.

So the first phase of any rewire is not building. It is writing down what actually happens, which is usually the first time anyone has. Plenty of agencies find that this step alone was worth the money, and a few find they did not need the rest of it. We would rather you found that out in week two than month five. Our post on what the audit actually finds covers what that fortnight turns up.

How to work out whether this is worth it for you

Three tests. They are honest ones, which means some of you should fail them.

Do you generate a lot of context and lose it? If your agency runs on recorded calls, long email threads and a decade of client history, a brain has something to work with. If you are three people with two clients and everything genuinely fits in one head, it does not.

Are your expensive hours spent on retrieval and assembly, or on judgement? Time the last monthly report honestly. If most of it was thinking, automation saves you very little. If most of it was gathering, formatting, chasing and reformatting, that is the part that goes.

Who would own it? This is the test people skip. A brain with no owner degrades. Someone has to notice when an agent starts producing slightly worse work, and that someone needs the time and the standing to stop it. If you cannot name the person today, the project will drift, and it will drift at exactly the point where everyone has started trusting it.

Do you have processes, or habits? If nothing is written down, you are looking at a documentation project with an automation project attached. That is fine, and it is still worth doing, but price it as what it is.

What we would say if you asked us directly

The honest summary is that a brain moves work, it does not remove people. The hours come off the parts of the job nobody enjoys and nobody bills confidently for. What you do with the freed capacity is a business decision that has nothing to do with technology, and it is the decision that determines whether any of this shows up in your accounts.

We covered that side of it separately, because it deserves its own argument rather than a paragraph here.