On this page
Facts checked . The products compared here change often; we re-check every 30 days.
An agency owner spends a weekend in March building the thing everyone has been talking about: an AI brain. Every client gets a folder. The call transcripts go in, and so do the brand guides, the pricing sheets and the notes from the last few quarterly reviews. On Monday she shows the team, and it's good. It answers questions about Fernhill Outdoor (a client we made up for this guide) better than half the people in the room.
The following January she checks who has used it. One person has opened it since the summer, and it's her. Kevin, who does Fernhill's SEO, still pastes last quarter's brief into a fresh chat. And the brain still lists Dana as the account lead. Dana left in August.
None of that means she built it badly. She built it for one user, and it has one.
What is an agency AI brain?
An agency AI brain is a shared store of what the agency knows about itself and each client, set up so the AI tools the team uses can read it. It goes by other names: a company brain, a second brain, a knowledge base for AI.
Most are home built. The common versions, simplest first:
- a project per client inside Claude or ChatGPT
- a folder per client that a coding tool reads, with an index and a running changelog fed by call transcripts
- a repository where the agency's processes live as reusable instructions
- a scheduled job that reads the drive, the chat channels and the call notes, and keeps each client's profile current
The instinct is right. AI without client context gives you the internet's average answer, and nearly everyone has the AI now. In a 2026 survey of nearly 200 US agency leaders by Forrester and the 4A's, 87% of agencies use generative AI. Most treat it as a cost of doing business (61%), and few sell it as a line of business (6%). When every agency has the same tools, what your tools know about your clients is the part that's yours.
Why does it only work for the person who built it?
The builder fills the gaps from memory. She knows the pricing sheet in the Fernhill folder is the old one, so she corrects for it without thinking. Kevin doesn't know. The AI gives him the old price with total confidence, and after the second time he stops asking it.
It lives where the builder works. If the brain sits in a developer tool or a repository, using it means opening something the rest of the team has never opened. People keep the habit that works, which is a fresh chat and a paste. A tool that adds a step loses to the habit. A tool that takes over a chore people already do gets used.
Nobody owns the fixes. When Kevin spots that Dana is still listed, one of two things happens. He can't change it, or doesn't know he's allowed to, so the error stays and the next person meets it. Or he can, because anyone can ask the AI to edit a file, and now changes go in that nobody reviewed. Either way the person who would catch it is the builder, who is usually the owner and the busiest person at the agency.
The AI's own summaries drift. Many brains save space by having the AI condense transcripts and notes into a running memory. Each pass loses a little, and now and then it flips something. Marco at Fernhill approves the premium tier for the spring launch only. The summary keeps "premium tier approved." The transcript is still there, but the AI reads the summary first, and nobody reads the summary, because not having to read things was the point. Three months later that line is in a deck.
Five questions to ask of any agency AI brain
They apply whether you built it, bought it, or are about to do either.
- Who has used it this month, other than the person who built it? If you can't see usage, ask around. If the answer is one person, you have a personal tool. That's fine, as long as you know it.
- Who keeps each client's record right? You want a name for each client. "Everyone" means no one. On my calls with agency owners, the ones who have watched a build die ask this first.
- How does a wrong fact get fixed, and can anyone on the account start it? Whoever notices should be able to flag it in the moment. A named person should say yes or no before it changes for everyone.
- What is the AI treating as true: checked facts, or its own summaries? Summaries are fine as notes. The facts that work gets built on (prices, offers, names, what the client approved) need a source and a date.
- What happens when the team changes AI tools? It will. Plain files in a folder move easily. A brain that lives inside one tool's projects and memory has to be rebuilt, and every rebuild is a chance for the team to go back to pasting.
Start smaller than a brain
The usual plan is to build the whole brain and then roll it out. It tends to go better the other way around: get a few people working from the same small, checked record, in the same way, and grow from there.
- One client. Pick one with real activity and more than one person on it.
- One page. The facts people re-explain most: who's who, what the client sells and for how much, how they sound, what's been approved and what's off limits. Each line says where it came from and when someone last confirmed it.
- One owner. A named person who keeps that page right. Usually the account lead, not the agency owner.
- One weekly job. Something the team already does every week, like the client status update, done from the page. The test is whether it replaces the digging.
- One sitting. Half an hour beside one team member, on their screen, with their client. Have them correct one wrong fact while you're there. That person becomes the one the rest of the team asks.
Then add the second client.
If you've already built one, you don't need to start over. Do the same thing inside it. Pick one client folder, put a source and a date on the lines that matter, name its owner, and have changes reviewed before they land (if it's a repository, a pull request does this). Then sit with one person on that account.
The bigger builds go on top once people trust what's underneath: the automations, the scheduled reports, anything that runs without a person watching. Automating on top of a record nobody checks produces wrong work faster.
If that sounds too small, the Agency Management Institute's Agency Core 2026 study of 579 agency leaders reached the same advice for the agencies under the most pressure: "It needs a smaller first step."
Should you build it or buy it?
Building is a reasonable choice, and plenty of agencies have someone who enjoys it. A repository already gives you a lot: a history of every change, and a review step if you use it. What it doesn't give you is the upkeep. Someone has to keep every client current, clear the changes waiting for review, get contractors and non-technical people onto the same record, and carry it across the next tool change. If you have someone who will own that every week, build. If it would fall to the owner, weigh that honestly.
Storystak, which publishes this guide, is one bought version: a client context layer for agencies. It keeps one approved record of facts for each client, called a Client Stak, which works as that client's AI knowledge base, and connects it to the Claude or ChatGPT plan your team already uses. Here it is against the five questions.
- Use. The Usage page shows how often each client's record gets read and what the team searched for and couldn't find, without anyone reading anyone's chats.
- Who keeps it right. You name the owner. Storystak holds every change until a person approves it, and shows who asked for it, why, the old line and the new one, and who approved it. Approving is a yes or a no on a change someone else wrote up, and it can be shared across account leads. That's the governance: a named person signs off on what the AI is allowed to treat as true.
- Fixing a wrong fact. Anyone assigned to the client can file a correction from the conversation they're in. It waits for approval, and then it's right for everyone.
- Checked facts. The numbers a client's work leans on (canonical stats) each carry a source and the date someone last confirmed them. One that nobody has confirmed in 120 days gets flagged.
- Tool changes. The record sits outside any one AI tool, so it comes with you when the team switches. If you leave Storystak, you can ask for an export as one file at no fee.
Each person is assigned to the clients they work on and reads only those, contractors included.
What Storystak doesn't do
It doesn't get your team using it. A person does that. We set up each Client Stak with you and work through a real task with someone on your team, because that's what we've seen work. Early on we left owners to run it alone, and it went quiet, for the same reasons as the brain at the top of this page. So now the first working session is with the person who will use it.
It doesn't check the AI's answers. It governs what the AI reads about a client. A person still reviews the work.
It doesn't name an owner for each client, and it doesn't schedule a review of the whole record. The flag covers canonical stats only. Naming the owner is yours to do.
It isn't an AI assistant, and it doesn't replace what you've built. Your prompts, your processes and your automations stay yours.
What does Storystak cost?
Month to month, Storystak is $100 a month for each active Client Stak, plus a $1,500 one-time setup that we are waiving for founding agencies. By the year, it is $75 a month for each active Client Stak, with at least five Client Staks prepaid for the year at signing ($4,500), setup included and the price locked for 24 months. For 20 active clients, that's $1,500 a month by the year. You pay only for the clients you switch on, so month to month you can start with one. Seats are unlimited. You keep paying for your AI tool, because Storystak doesn't replace it.
See it on one of your clients. Take the demo tour on a fictional agency, or book a call and we'll build a Stak from what you already have and connect it to the AI your team uses.
Frequently asked questions
What is an agency AI brain?
It's where an agency keeps what it knows about each client so its AI tools can read it. Most are home built, from a project per client up to a repository with scheduled jobs that keep it current.
Why doesn't my team use the AI brain we built?
Usually for one of four reasons: the builder fills its gaps from memory and nobody else can, it lives in a tool the team doesn't open, nobody owns the fixes, or the AI's summaries have drifted and people stopped trusting the answers.
What should go into an agency AI brain first?
For one client, the facts people re-explain most: who's who, what the client sells and for how much, how they sound, what's been approved and what's off limits. Give each line a source and a date, and give the page a named owner.
Is a Claude or ChatGPT project per client enough?
It's a good start, and for one person it can be all you need. Ask it the five questions above. We compare the options in Is Claude Team enough for marketing agencies?
Should we get the brain right before we automate anything?
Yes. An automation is only as right as the record it reads. Get a few people working from the same checked facts first, then build on top.
What is a client context layer?
A client context layer sits between what an agency knows about a client and the AI tools its team uses. It is one approved record per client that every AI reads, with a source on each fact and a person approving each change. It does the job of an AI knowledge base, with rules about who can change it and who can read it. Storystak's version is called a Client Stak. An AI brain one person built is the same idea without the approval step, the sources or a way in for the rest of the team.


