LiamAI made hours cheap. Liam sells what's left: a brain that remembers.
Expertise and generic advice is now cheap and only a few clicks away. The only scarcity left in consulting is memory: a consultant's brain and judgement that compounds across every engagement.
But it only wakes up when the human is in the room. We designed Liam, an always-on productized advisor living in CGS's clients' Workspace.
Watch Liam work
Problem
Hours used to be the product
Now AI does two partner-hours of work in ninety seconds. Expertise is cheap and generic advice is free. Consulting can't sell on scarcity anymore.
When the consultant leaves, so does the advising brain
A consultant's hard-won judgment gets tapped into only when they are actually there in the room. The second they log off, it all goes with them.
Solution
Meet Liam, CGS's brain, always on
He lives in the client's Google Drive and works like a senior advisor who never logs off. He reads the engagement, surfaces what matters before anyone asks, and writes up the thinking. Copilots give you output.
Liam gives you counsel.
Model
A new revenue line, sold by the seat
CGS used to bill by the hour. Liam is sold by the seat. Each executive gets their own always-on advisor, priced like a subscription, not a timesheet. It runs on pennies of compute and never burns a partner's hours. One engagement becomes a standing product line.
What Liam does
Shows up before you ask
While Alex (the CEO of our simulated client Meridian) slept, CGS published new research. Liam read it overnight and referenced it against everything in the Workspace. By the time Alex opens his laptop, the analysis is already waiting. He works in the background and shows up in the sidebar in files when you need him.
Gives counsel, not output
Fourteen hours before a board vote, Alex has to handle Diana, the board member nobody's cracked. Liam doesn't hand him a deck. He hands him a brief to walk in with. “Diana will open on assumptions. Not numbers. She always does.”
Knows when not to answer
Alex asks something outside Liam's remit. Liam won't answer it. A Governor agent checks every response for confidence, scope, and risk, and this one failed all three. So Liam does what a good associate does. He calls the partner instead. “I don't have grounded data to be confident. I've flagged this for James Chen.”
One click in, one click out
Adding Liam to an engagement works like sharing files with a consultant. Removing him is just as easy. The moment he's out, he reads nothing and learns nothing from that client. What he built till then stays in their Drive. What he learned stays with CGS.
How it works
3 models, 5 agents, 1 Workspace
Gemini Flash watches the Drive and moves files. Fast and cheap. Gemini Pro does the grounded analysis, every claim cited. Claude Opus handles the hard judgment, like the Governor's scoring and escalation. Every task goes to the smallest model that can do it.
At each seat, Liam remembers the engagement. That memory never leaves the client's Drive. Across the client, he learns shapes, not documents. Which arguments win. Where conviction holds. Across every CGS engagement, the patterns compound. CGS IP flows down to every Liam. What each Liam learns flows back up. Never client content, just what works. Every engagement makes the brain bigger.
Bigger vision
The company brain is the table. Liam is one seat.
Y Combinator calls the company brain one of 2026's biggest builds. Build for one phase of a customer journey and you've made a tool. Build for the whole journey, and you're rebuilding how consulting gets delivered. We framed that bigger vision for CGS — and Liam is the first step toward it.
The shared brain — a bidirectional flywheel. Scout, Atlas, and Liam all read from Forge, and all write back: patterns, never artifacts.
InternalWhat I learned
The guardrail is the feature
Making Liam refuse was harder than making him answer.
The hard part wasn't the AI
It was structuring context so the agents reasoned well.
The reframe was the work
We started building a better chatbot. The breakthrough was realizing the product wasn't availability — it was memory.