The wall I kept hitting
Every architect (or GTM consultant) on my team hits the same ceiling: three accounts, maybe four. The fifth one comes in and something gives: a board report goes out late, a client asks a question nobody prepped for, somebody starts working Sundays. So I'd hire. Then we'd hit the wall again, one headcount higher.
For a long time I thought that was just the shape of the business. Senior people are expensive and slow to train, so growth is a recruiting problem before it's anything else. I priced accordingly and hired accordingly.
Then I actually looked at what my architects were spending their week on, and it wasn't just advising. It was data assembly.
The cost I didn't have a line item for
Every client comes with a standing reporting obligation — board report, pipeline view, attainment tracking — and building it by hand ate a real chunk of a senior person's week, every week, for every client we ran.
And that's before a new client even gets advice. Most of them show up unable to tell me what their own revenue is, because the number in the CRM doesn't match the number in billing and nobody's reconciled them.
"That's kind of where everyone usually hits us up, because they don't even know what their revenue is in Salesforce."
Neither of those is consulting. Both need a senior person, because a junior one can't tell which of two conflicting numbers is the real one. That's the actual ceiling; not talent, not process. My most expensive people were doing work that needed their judgment for reasons that had nothing to do with judgment.
What I built first
I didn't start by automating advice. I started by automating the part that wasn't advice.
On day one of a new client now, a diagnostic agent gets pointed at their connected data and runs the full teardown: funnel conversion by stage, where deals stall, rep coverage, pipeline that's quietly dead. It writes that up against our playbook, in our voice, as a finished diagnostic.
What used to take a senior person eight hours now takes one.
"The consultant is still in the loop, but now they're editing judgment, not just assembling data."
That's the whole thing, right there. I didn't remove the senior person from the engagement. I removed them from the spreadsheet.
What almost got me in trouble
Here's the part I'd tell anyone trying to copy this: don't skip the step before the agent.
If you've got a Claude or ChatGPT license, you can connect it to a CRM this afternoon and ask for your pipeline coverage ratio. You'll get a confident answer instantly, and it'll probably be wrong. Four things are usually missing: the definitions (expansion pipeline and new-business pipeline aren't the same thing, raw AI blends them), identity resolution (is "Acme Inc." and "Acme Co." one company or two), the plan (no targets, no such thing as on track), and memory (what an ICP means at this company, what happened last quarter).
At one company, a wrong guess is an annoyance. Across a book of dozens of clients, it compounds fast.
"If we ran just on raw AI, we would be shipping confident wrong answers to 30 different boards. That is not a business. That's a liability."
So I did the unglamorous part first. Resolved the data, built the semantic layer, and only then let an agent into the equation. Skipping that order is the fast way to be wrong in front of a client.
What I built next
Once the diagnostic agent was working, I pointed the same idea at us. Every call transcript that lands gets read by an agent that pulls out commitments and turns them into scoped tasks. Another watches accounts for signals that something's heating up or cooling off. A third watches our own delivery for where someone's stuck and could use coaching.
"The agency is improving itself while we work."
Running dozens of clients creates its own operational load, and honestly, that load breaks agencies more often than the client work does.
What that capacity actually bought
Before this, an architect could carry three or four accounts. Now it's six to eight — no hiring round, same team.
I want to be precise about what caused that, because it wasn't one agent. It was the diagnostic, the daily digest that replaced a manual performance email, and a data layer that assembled the numbers so nobody had to. The diagnostic is the most visible piece. It's not the whole mechanism.
What I'd tell you to do first
If you're running an agency or a RevOps function and you're staring at the same wall I was staring at, here's the order I'd actually go in:
- Find out what your senior people are spending their week on. Not what you assume — actually look.
- Resolve the data before you automate anything. Definitions, identity, plan, memory. This is the boring, unglamorous, front-loaded work, and skipping it is why most first attempts at this fail.
- Automate the assembly, not the judgment. If you can't say what the senior person is now editing instead of building, you've automated the wrong part.
- Only then ask what the freed-up time should go build next.
The capacity ceiling in a services business gets described as a people problem, and it usually isn't. It's a question of what your expensive people spend their week doing. Move the assembly, and the ceiling moves with it.
Related reading
