Agency Knowledge Base: How to Build One for Client Work
An agency knowledge base is where your team keeps what it knows about each client and about how the agency runs. For agencies, it should be split by client, with access limited per client, and written so people and AI assistants can both use it. Without that split, client details can end up in the wrong drafts.
That’s the problem a single company doesn’t have: you hold several brands’ confidential plans at once, and the people on each account change.
What should an agency knowledge base include?
Two layers: what’s true for the whole agency, and what’s true for one client.
| Layer | What goes in it | Who reads it |
|---|---|---|
| Agency-wide | Processes, QA checklists, templates, pricing rules, tool how-tos | Everyone |
| Per client | Brand profile, audience, voice, banned words, approvals, past campaigns, contacts | Only the people on that account |
| Per project | Briefs, research, drafts, feedback rounds | The project team |
The per-client layer is where most of the value sits. It’s also the part that tends to live in a few people’s heads and in private AI chats.
How should you structure each client’s space?
Give every client the same skeleton, so anyone on the team knows where to look. Say your agency has two clients, Juniper Lane and Harbor & Pine. Each gets its own space with the same notes:
- Profile. One page: who the client is, what they sell, the audience, the voice in a few lines, words to avoid.
- Voice examples. A few approved pieces, with a line on why each worked.
- Decisions log. Dated entries, for example “2026-09-14: no discount language in subject lines.”
- Approvals. Who signs off on what, and how fast they usually reply.
- Campaign history. What ran, what the result was, what to avoid repeating.
- Contacts. Roles, not just names.
The profile does the most work. When a writer, or their AI, starts on a Juniper Lane task, reading that page first heads off the usual voice mistakes. For the voice part specifically, see how to make ChatGPT remember your brand voice, which works the same way for client brands.
How do you keep client information separate?
Separate spaces, not separate folders in one space. The difference matters once AI is involved. If every client’s notes sit in one searchable pile, an AI asked to find “launch email ideas” can pull a Harbor & Pine idea into a Juniper Lane draft. That’s embarrassing at best, and a confidentiality breach at worst.
Three rules keep clients apart:
- One space per client, with its own search, so a query only ever sees that client’s notes.
- An access list per client. Only the people on the account can open it. A freelancer hired for one client can’t browse the others.
- Client-specific AI connections where possible. If a tool lets you connect an AI to just one client’s space, use that for client work.
The same principle for solo creators running several brands is in one vault per brand and managing multiple brands with AI.
How do you onboard a new client into the knowledge base?
Turn the kickoff into notes the same week. A simple sequence:
- After the kickoff call, write the profile from your notes and the client’s brand guide.
- Add the approved examples they sent, each as its own note.
- Log the first decisions: tone, channels, things they never want said.
- Share the profile with the client contact to confirm. Fix what they flag.
Ask your AI to draft the profile from the call notes, then edit it yourself. The draft saves time; the edit is where the accuracy comes from.
How do you onboard new staff with it?
A new account manager should be able to read one client’s space in a single sitting and then write a passable first draft. That’s a good test of whether the space is complete. If they keep asking the same questions, the answers belong in a note.
Give new people viewer access in their first week. They read and learn without changing anything, then move to full access once they’re on the account.
How does the knowledge base work with AI tools?
If your staff draft with AI assistants, they probably don’t all use the same one. The knowledge base should work with each of them, not just one vendor’s chat.
| Approach | Works with | Risk |
|---|---|---|
| Paste the client profile into each chat | Any AI | Out of date copies, forgotten updates |
| One shared project per client in one AI | That AI only | Staff on other tools are left out |
| Client spaces reachable over MCP | Any AI that supports MCP | Each person connects once |
With MCP, each person connects their own assistant to the client spaces they have access to. The AI reads the profile, searches past decisions, and can save new notes back, so a decision made in one person’s chat is there for the next writer. The shared knowledge base for teams post covers the general setup.
What happens when a client leaves?
Export the client’s space and remove everyone’s access to it. Keep the export in case they come back or you need the record. Don’t leave old client notes searchable in the active base: they’re noise for the team and a risk if an AI pulls from them.
Team changes need the same care. When someone leaves the agency, remove their access and check that their notes stay with the agency, not with their personal account.
FAQ
What’s the difference between an agency knowledge base and a team wiki?
A wiki is usually one shared space. An agency base needs a hard split per client, with separate access, so client work stays confidential.
How detailed should a client profile be?
One page. Long brand books still matter, but the profile is the summary everyone reads first. Link the full documents from it.
Should freelancers get access?
Yes, to the clients they work on and nothing else. An access list per client makes that simple.
Can AI keep the knowledge base up to date?
It can draft and save notes when you ask, which keeps the base current. A person should still own each client’s profile and check changes to it.
Where Valta fits
Valta’s Team plan gives an agency up to 20 shared team vaults, so each client can have one. Every vault can have its own access list, and each has a profile (_vault.md) that every AI reads first. Members get roles (owner, admin, member, viewer), and an activity log shows which person saved what from which AI. Each vault has its own connector URL that opens only that vault, so a client-specific AI setup can’t reach other clients. Staff keep whichever MCP-capable assistant they prefer; Valta runs no AI itself. Any vault exports as a zip of Markdown.
Team is billed per seat, minimum three, with prices on the pricing section. For context on why agencies keep a base like this at all, see a second brain for your business, or create a free account to try the structure first.