“AI doesn’t really work. What do we do with that? Because people say, ‘Well, over time it’ll get better.’ All right — but you’re building a company now.”
That’s Jason Cohen. WP Engine, Smart Bear, two billion-dollar companies. If you run a WordPress agency, you’ve probably worked on top of his infrastructure.
And honestly? He’s right.
Open a fresh AI chat. Paste a URL. Explain who the client is. Remind it what you did last week. Get back something 80% right. Spend an hour cleaning up the 20%.
That’s not AI doing great work. That’s you doing great work — after the AI pretended to.
Same Claude. Same ChatGPT. Same Cursor. Same models every agency on earth uses. And yet — some agencies are pulling real 10x gains out of them. Others are stuck at “saved me twenty minutes, then I rewrote it anyway.”
Same tool. Different outcome. So what’s the difference?
Jason again:
“If AI is supposed to be so revolutionary, how come it’s only increasing my performance by 20%? It’s just not worth the hassle.”
That’s the whole game.
A 20% AI isn’t worth learning. A 10x AI rewrites how you operate. And the gap between the two is almost never the model.
It’s the context you give it.
What “context engineering” actually means
It’s a fancy term for something very unfancy. Context engineering is making sure your AI tools know the things they need to know before you ask them anything.
Their voice. Your standards. The history. The full picture.
Think of it like a new hire. The one who nodded through every briefing and then, on day one, wrote a proposal in a tone nobody at your agency would use, to a client they half-invented, quoting a service you haven’t offered since 2022.
That’s what generic AI does every time you open a fresh chat. Confident. Sounds right. Often catastrophically wrong. And you spend the next hour cleaning it up.
Congratulations. You just automated the easy part and kept the hard parts for yourself.
Why “just write better prompts” isn’t the answer
For a while, the fix was prompt engineering. Write cleverer instructions. Build a shared prompt library. Hand the team a one-pager of magic incantations.
It worked. Kind of. For single, one-off questions.
But agency work isn’t one-off. You’re juggling dozens of clients, each with their own voice, history, sites, preferences. No prompt carries all of that. You’d be writing novels before every question.
And you’re not paid to write novels.
So the real question stopped being how do I phrase this better? and started being how do I make sure the AI already knows the things it needs to know?
That’s context engineering. And it’s the difference between AI as a vaguely-useful 2x assistant and the 10x multiplier Jason is actually talking about.
What actually changes when the AI has the picture
You ask: “What’s the status of Mike’s sites?”
Generic AI shrugs. Context-aware AI answers: three sites, two up to date, one with a WooCommerce vulnerability flagged last Thursday, a backup from this morning, 99.98% uptime, last client call was about page speed, here’s where that landed.
You ask: “Draft a quick check-in for Sarah before our call.”
Generic AI invents a client. Context-aware AI pulls her last three tickets, the plugin update she was nervous about, the redesign brief you started in February, and the fact that she prefers warm, direct emails under four lines.
You ask: “Write the monthly report for Acme.”
Generic AI produces something vaguely plausible. Context-aware AI writes in your voice, from your data, at the exact level of technical depth Acme actually wants — because you told it once, and it stuck.
That’s not a small gap. That’s the whole difference between AI tried and AI delivered.
What a real agency brain actually unlocks
Every serious agency already has this context. It’s in senior developers’ heads. Scattered across Slack threads, Notion pages, client folders on Dropbox. In the muscle memory of whoever’s been on the team longest.
A real agency brain is all of that — clients, sites, standards, voice, processes — structured once, in one place, so any AI tool can use it.
When that’s in place, four things start to happen:
- Any AI tool becomes agency-aware. One MCP endpoint — the open standard for connecting AI tools to your data — and Claude, Cursor, ChatGPT, and Claude Code all answer from the same source of truth. No more pasting the briefing three times.
- Routine work collapses. The monthly report that used to take two hours takes ten minutes. Status summaries become conversations. Onboarding docs write themselves.
- Operations stay traceable. Every site update, plugin install, and client change is logged — who did it, when, on which site. A real operational record, not a shrug when a client asks what changed.
- Your best processes become repeatable. The way your senior dev audits a new client site. The way your PM writes a kickoff email. The way your agency handles a migration. Captured once, runs every time.
That last one is where the real gap opens up. In a world where every agency has the same AI models, the ones that win are the ones whose expertise actually travels — out of people’s heads, into something every team member and every AI tool can use.
Try this before Wednesday
You don’t need Dollie — or any platform — to start doing this.
Open an AI tool you can actually talk to out loud — ChatGPT voice, Claude voice, whatever your team already uses. Talking beats typing. You’ll go deeper, the answers come out more honestly, and the whole thing takes fifteen minutes instead of an afternoon.
Then paste this in:
“Hi. I’m an agency owner. I run an agency called [YOUR AGENCY NAME]. Your job is to help me build a repeatable agency brain I can reuse across any AI tool. Do an in-depth interview with me. Ask me questions until you genuinely understand how we work, how we communicate with clients, what we stand for, what makes us different, and what our standards are. When we’re done, turn my answers into a file called
agency.mdthat I can reuse anywhere.”
Now actually answer. Properly. The way you’d brief a senior hire on day one. How you scope projects. How you write to clients. What a good kickoff looks like, and a bad one. The principles you won’t bend on. The stuff your team just knows without anyone saying it.
Do the exercise again for one of your best clients. Same structure — ask for the interview, get an acme.md out the other end.
Then open a fresh chat, drop both files in, and ask the same AI to draft Acme’s monthly report. Or write a kickoff email for a new project. Or audit one of their sites.
The difference will stop you in your tracks.
That’s an agency brain — the manual, hand-built version. A couple of markdown files, a couple of hours of interviews, and every AI tool your team uses gets dramatically sharper overnight.
Here’s the thing: the file you just built is a sneak peek at something bigger. At Web Agency Summit we’re announcing Agency.md — an initiative to turn that exact idea into a real, open standard, so your agency has a discoverable identity across the whole ecosystem of AI tools and hosts built for agencies. What you build by hand this week becomes a portable file the whole industry can read.
And we’ll show you how Dollie does it automatically — keeping that brain alive across every AI tool your team uses, for every client, without anyone rebuilding the context by hand. WAS attendees get a small early access window: a Dollie spot with all Pro features included, free during early access.
Either way — start this week. The tool question comes second.
