Ask any agency that has "tried AI" and the story is remarkably consistent. It isn't a story about a bad tool. It's a story about a missing layer, and the fix costs less than the plateau.
A familiar pattern
Week one: someone on the team (often the most junior person, with curiosity and a deadline) uses ChatGPT to draft a press release. A draft in minutes instead of the better part of an hour. Screenshots go in the team chat. Someone says "we should do this for everything."
Week three: the drafts come back generic. The tool writes like a competent stranger, because that is what it is. It doesn't know that one client bans the word "innovative", that another insists every quote comes from the CFO, or that your Flemish briefs need a different register from your Dutch ones. So people start pasting in background documents by hand. Every task. Every time.
Week six: nobody has opened it in days. Not because it broke. Because feeding it context manually cost more than it saved. Usage dies quietly. The subscription lingers on the company card. "AI" gets filed under tried it, didn't stick.
This pattern repeats at PR and influencer agencies across the Benelux. The tool varies. The plateau doesn't.
The root cause is amnesia, not the model
Each conversation with a general-purpose AI tool starts from zero. It has read most of the public internet, but it knows nothing about your agency: not your clients' tone rules, not the reporting format your biggest account expects, not which creator soured a client relationship last year and must never appear on a longlist again.
Your team carries that knowledge in their heads and in a thousand scattered places: old decks, email threads, a senior colleague's memory. A new hire takes months to absorb it. An AI tool never absorbs it at all, because nobody gives it anywhere to live.
So the output is generic for a boringly logical reason: generic input, generic output. The model is fine. The context is missing.
The tool was never the problem. The blank slate was.
Why a better prompt won't save you
The usual response is to patch the symptom. Someone builds a prompt library. Someone runs a lunch-and-learn on writing better prompts. Both help for a fortnight, then decay for the same reason the first attempt did: they push the burden of context onto whoever is typing.
A prompt library is static. Client rules aren't. The moment a brand updates its messaging or a new account lands, every saved prompt is quietly out of date, and nobody's job is to notice. You've recreated the problem one layer up.
The failure is structural, so the fix has to be structural too.
The structural fix: give the AI a maintained memory
The way past the plateau is to stop re-explaining your agency to the AI and build it a memory instead.
Concretely, that means a structured, written knowledge layer: think of it as an internal wiki with discipline. One page per client: tone rules, banned words, approval chain, reporting standards. One page per workflow: how a creator longlist gets built, what a wrap-up report must contain. Plain text, owned by the team, edited like any other document.
We call this the agency's AIOS, its AI operating system: a written knowledge layer both your team and your AI work from. The AI connects to it directly (the plumbing for this is MCP, an open standard that lets an AI read and act on your existing tools, no copy-pasting involved). When someone updates the client page on Tuesday morning, the AI knows it by Tuesday morning. No retraining, no IT project.
The difference in output is immediate. Ask a blank-slate AI for a creator brief and you get something plausible and useless. Ask an AI that can read your client's page (tone, deliverables, past creators, the sensitivity from last spring), and you get a draft your team edits rather than rewrites.
Point it at the drudgework, not the craft
What you automate matters as much as how. The wins are not in the strategy deck or the client dinner. They sit in the operational grind between them: briefing packs, contract chasing, content logging, screenshot collection, report assembly. Senior people at Benelux PR firms consistently tell us this runs to roughly 10 hours of manual work per creator per campaign: around 300 hours for a 30-creator campaign, and about 1,500 hours a year at five campaigns. We've broken down where those hours hide in The 10-hours-per-creator problem.
That is the work an AI with a proper memory does well, because it is context-heavy and judgement-light. The judgement (which creators fit the brand, what the story angle is, how to handle a nervous client) stays exactly where it is. The point is not fewer people. It's the same people, minus the part of the job nobody joined an agency to do.
The side effect nobody prices in
A maintained memory pays a second dividend that has nothing to do with AI.
Agency knowledge famously walks out the door. When an account manager leaves, their successor inherits an inbox, not an understanding. With a maintained knowledge layer, the client's quirks, history and standards are written down in one place, because the AI needed them to be. Staff churn stops deleting institutional memory.
Onboarding flips the same way. A new hire's first week can start with "read the client pages" instead of three months of absorbing folklore by osmosis. And every improvement compounds: each rule someone writes down makes both the humans and the AI sharper from that day on.
Which raises the real question: whose job is it to keep this memory alive? It can't be nobody's: that is how prompt libraries die. It probably shouldn't be IT's either. We've laid out the options in Who should own your agency's AI stack?.
Where to start
If an AI experiment has already stalled at your agency, treat that as useful data rather than failure: it tells you the missing ingredient is context, and context is something you can build. Start deliberately small. Pick one client and one recurring workflow (the monthly report, say, or the creator brief) and write down, on a single page, everything a competent freelancer would need to do that job well.
That page is the seed of your agency's memory. Connect it to an AI and the step that took an afternoon takes minutes, and a first pilot of that kind is a three-to-four-week build, not a year-long programme.
The test of whether you've fixed the structural problem is simple: the output gets specific, and nobody ever has to paste in the background again.