Agency  IP is  spread across inboxes, calls, slack threads and somebody's laptop in another timezone. Of course the relationships still sit with people, as does the creativity and the discernment.

Sitting on top of the new AI reality we’re all in, for the first time some of that knowledge and judgement can stay accessible inside the business, wherever anyone happens to be working. Not replacing anyone, just making the organisation a little smarter every day. I’m not talking Slack, Notion or the company intranet – though they are a part of this.

Love it or hate it, AI is the opportunity. Connected in the right way it can be a big unlock.

Like most people I have been sliding down the razor blade of AI, legs akimbo – always learning, looking for the promised unlock.  

Where we find ourselves now is developing our AI-Brain, and around here his name is Andre.

People in AI circles have a few names for the same idea, so you'll hear Context Brain, Obsidian Brain, Second Brain, or simply a structured knowledge repository. They are all describing much the same thing, which is a way of organising everything your business knows so that your team plus AI can use it intelligently. We prefer AI-Brain because that is what it feels like to work with.

It isn't another model, and it isn't the latest and greatest app. It is the accumulated knowledge, methods and memory of the business, organised so that every new task begins with years of experience already in the room. That one shift has unlocked more value than anything else we have done. We now run much of the business in cahoots with our AI mate Andre. 

The funny thing is how ordinary an AI Brain looks. A few folders, mostly plain text files, arranged so that the model loads exactly what it needs before a conversation even begins. The least interesting part of the build turns out to be the part carrying almost all of the weight.

Looking back, we travelled through three fairly distinct stages.

The first is prompts. Learning to ask better questions does matter, and the journey teaches you a lot about how these models actually behave. You get some useful answers, some average ones, and the occasional response that genuinely surprises you. Then you carry on working much as you always have.

The second is context, where you start teaching the machine about your business. Your clients, your frameworks, your methodology, your language, your tone, your brand. The answers improve straight away, because half the briefing has already been done before you have typed a word. Most people think they have nailed it at this point. I did too, for a while.

The third is harnesses, and this is where the AI-Brain becomes more than isolated pots of information. The brain and the harness together "decide" which knowledge is relevant, when a specialist capability should appear, and how work actually flows through the organisation. That was the breakthrough, and I owe the nudge that got me there to Henry Stanley at Attn: Seeker, who put the idea of an AI brain in front of me in the first place.

For months before that I assumed the problem was me. I built specialist AI after specialist AI, one for research, one for finance, one for buyer identification, one for design. Individually they were good. Collectively they were chaos, because each one lived in its own little world, and every time I needed one I had to remember it existed, find it, and work out whether it was even the right tool. There was no linkage between them. It felt like owning a workshop where every tool was locked in a different building.

The AI brain fixed the knowledge and the harness enabled it. Now the right capability turns up when it is needed, carrying exactly the right context with it, and I no longer have to think about where it lives.

Which is where the scattered-knowledge problem from the top of this piece actually gets solved. We work across Barcelona, Sydney, Los Angeles, the UK and Dublin, which used to mean either everybody sat in on every call or somebody got briefed badly afterwards. Now one person covers the call and the rest of us pick the account up from the record rather than from a handover. Remote working stops depending on how generous people are with their time, and starts depending on whether the business writes things down.

The next lesson was a good deal simpler than the time it took to learn it. Don't start by building skills, however enticing that may be. Build an AI brain first. Once you have this the skills are so much more relevant to the business, wired into it - can be evolved easily. Then the game turns to ‘skill-stacking’. And the promise of AI starts to unlock, at pace.

That also changes who should be building it. This isn't a job for an AI specialist, it is a job for the people who actually know the work. I have spent a surprising number of hours teaching Andre what we value, how we make decisions, and what good looks like, and every correction makes the next interaction better. That is an an ongoing process. It is less like installing software and more like mentoring a new colleague.

Except that Andre rarely forgets anything.

It’s the same argument behind the money the holding groups are spending. Publicis has committed 300 million euros to CoreAI and Havas 400 million to Converged.AI, and none of them has exclusive access to better models, because they all buy from the same handful of providers as everyone else. What they are paying for is the ability to make decades of accumulated data and method available instantly and consistently, right across the group. The value isn't the model. The value is the brain.

For an independent agency that brain looks completely different. It isn't billions of customer records, it is your discernment and your perspective. The questions you always ask, the standards you refuse to compromise, and the way your team approaches a problem before anyone else knows where to begin. Alongside all of that sits the ordinary operational detail of the business, which is suddenly always to hand.

By definition, LLMs know what everyone knows. Most agencies are valuable for the opposite reason, because they cut through the shit to what everyone else misses. They break new ground.

Which is why every AI-Brain will be different, and why it should be. Different clients, different capabilities, different culture, different beliefs, different guardrails. The opportunity isn't to build the same AI as everyone else. It is to build one that thinks like you.

So what does that actually look like day to day? The agencies I see doing this properly:

• Don't do timesheets, because the brain works out what a job cost better than most people manage from memory, and then asks a human to confirm it.

• Don't run client sat or NPS, because they have built signals that detect a problem long before a survey will.

• Don't spend hours prepping proposals. They spend the time on the thinking, and riff with the brain on past projects and research to build something more compelling.

• Don't use PowerPoint anymore. It's HTML and at a pinch, PDF, which is waaaaay better.

• Don't write status reports, timesheet analyses or billing reports, because the admin runs itself.

• Don't take meeting notes, because every transcript is available immediately and drives a dozen other bespoke applications

• Do focus on leading indicators rather than trailing ones.

• Do mash up the whole knowledge repository using MCP  and connected services to create new skills

• Do skill stack into a build roadmap

• Do apply their creativity to the things that actually matter.

• Do focus on human connection rather than admin.

For decades we have said an agency's assets walk out of the door every evening, and I still believe that. But something new is happening alongside it, which is that for the first time agencies have a practical way of capturing how they think in a useful repeatable way, rather than in assets that atrophy on a server somewhere. That doesn't replace people. It makes the organisation itself smarter.

This has the power to change not just how agencies work, but also what they're actually worth.

Happy to share our learning on this. Hit us up for a chat!