Published, free to read

The Agency AI Infrastructure Stack

Five layers behind consistent AI work in an agency, in the order they have to exist.

Why the order matters

Most agencies never decided to build an AI system. They bought seats, people started working with them, and after a few months something that looks like a system exists.

The entry point is nearly always the same. Someone sets up access and connections, because that is the visible problem. Or someone builds workflows, because that is the interesting one. Both are real layers, and both sit on top of a layer that was never built: one current, agreed version of what is true about the brand being worked on.

So good tools produce output that changes depending on who ran it. The tools are not the reason.

Five layers, from the bottom up. Each one gets three things here: the job it has to do, what breaks in an agency without it, and one question about your own agency. Not a client account. Your own.

The questions are the point. An agency that cannot answer them about the brand it knows best has its answer about everything else.

Layer 1 — Shared context

The job: One current, agreed version of what is true about a brand — positioning, audience, offer, messaging, working rules — for everyone and every AI environment that touches the work.
What breaks without it: Ask four people to describe the same client's positioning and you get four answers, each defensible, none marked as current. The AI output inherits that: copy that reads fine on its own and is wrong against the account, corrected in a chat window instead of in the source, and corrected again next month.
The question about your own agency: What would an AI environment read about your agency before it wrote a line for you, and who last agreed that what it reads is true?

This is the only layer here with nothing to buy. A folder holds whatever you put in it, correct or not. Someone still has to decide what is true, what is verified, and what is missing.

One agency, worked through · constructed example, not a client

Rasterwerk

Rasterwerk is invented for this example. Fourteen people, Hamburg, mostly B2B industrial clients, one founder who still writes the important pitches.

Asked the question above, the founder's first answer was the website. So we looked: the site says Rasterwerk is a "digital agency for ambitious brands". The pitch deck says "positioning and demand generation for industrial B2B". The LinkedIn profile of the second-longest-serving employee says "we build campaigns that convert". None of the three is wrong. No two of them would produce the same piece of writing.

What was missing was not a description. It was a decision about which description holds, written where a person and a machine both find it. Here is what the first page of it looked like:

Constructed sample Rasterwerk — shared context, page 1

Positioning: Rasterwerk positions industrial B2B companies for a buying process with four to nine people in it and a decision that takes months. Not brand work in general. Not performance marketing in general.
Who we are not for: Consumer brands. Companies that want media buying without positioning work first. Verified against the last eleven enquiries: four were turned down for this reason.
The claim we make: That the buying centre can name what the company is for after one page.
Open, marked as open: Whether the automotive supply work counts as a second segment or as three coincidences. Founder decides by end of quarter. Nothing in a proposal until then.

Four lines and one admission. The admission is the part that makes it usable: an AI environment reading this writes around the open question instead of inventing an answer to it, which is what it did before.

If your tools change: one approved current version per brand still has to exist, and people and AI environments have to read the same one.

Layer 2 — Access and environments

The job: Decide who works in which approved environment with which material, and make that decision hold without anyone being asked.
What breaks without it: The best setup in the agency lives on one laptop, and keys get pasted into a chat on a Friday to unblock someone. When a person leaves or an account ends, you find out what access existed by finding out what breaks.
The question about your own agency: Which approved environments hold your own agency's material today, and who decided that list — or did it just accumulate?

If your tools change: access is granted per brand and revoked in one place, and no individual holds credentials the agency cannot see or replace.

Layer 3 — Skills and repeatable work

The job: Turn recurring work into a defined task with a named input, a defined output and an owner, so the result does not depend on who ran it.
What breaks without it: Everyone keeps a private prompt collection. The good one belongs to the person who wrote it, so when they are away the output drops in a way the client notices before you do, and a fix made on Tuesday never reaches the colleagues who need it.
The question about your own agency: Which of your own recurring deliverables — the credentials deck, the case study, the monthly report you send yourselves — could a new colleague produce without asking you first?

If your tools change: recurring work belongs in a shared, versioned definition the team runs from, not in one person's memory.

Layer 4 — Governance

The job: Set who may change the shared material, who reviews it, and who owns the current version.
What breaks without it: A client changes their positioning. One person updates the material they work from, another does not, and weeks later the old version turns up in a deliverable that already passed review. Nobody broke a rule. There was no rule about who owns the change.
The question about your own agency: Who is allowed to change how your agency describes itself, and where would the rest of the team notice that it happened?

If your tools change: someone owns the current version, and the right to change it stays narrower than the right to work with it.

Layer 5 — Automation

The job: Take repetitive preparation off people, once the four layers underneath hold.
What breaks without it: Automation is usually the first thing an agency buys and the first thing that hurts. It does not create the inconsistency. It copies it into more places at higher speed: more output, arriving faster, with the same problems in it.
The question about your own agency: What already runs automatically on your own brand, and which version of your positioning is it running on?

If your tools change: automation is what you add once the foundation holds, and it earns a task only when the task clearly recurs and its output is already good.

Where this leaves you

Layers 2 to 5 are buying and configuration decisions. There is a product for each of them, and when one turns out to be wrong you replace it and keep the layer.

Layer 1 has no product. Nothing you can buy knows what is true about your positioning, what has been verified, and what was only assumed. A folder stores it. A session loads it. Neither of them produces it.

Four of the five layers you can buy. The bottom one you have to build.

You have five questions now. Rasterwerk's answer is on this page. Yours is not, and it is the one that decides whether anything above layer 1 is worth configuring.

Build the context once. Then let your AI work from it.

€9

€9, one-time. No subscription.

Get the BrandOS Builder for €9 →

{price} once, no subscription. The access link arrives by email right after checkout.