My AI Call Copilot Runs on My Brain, Not a Persona

Forty minutes into a client call, a change nobody had approved showed up in production. The other side's verdict: "That is a fire."

I didn't hang up to go digging. I typed one line to the copilot sitting in: how did this get into production?

Two minutes later the answer was on my screen: which change it was and how it got there. Within about fifteen minutes, a revert was open and waiting for review. The call never stopped. Nobody had to say "let me get back to you."

That copilot isn't a meeting bot. It's Kortex, my memory, with a seat at the table: a searchable store of every call, commit, chat and note I've recorded.

What can an AI copilot actually do during a live call?

Six modes, each learned the hard way on a real call.

1. TakesOpinions, pushback, risks and ideas, each short and argued. Never a script: "don't propose exact words for me to say." One call of over three hours: 52 takes, 39 ideas, 17 numbers.
2. NumbersWhen the talk touches data Kortex holds, a background agent computes the number before anyone pulls it up. The other side: "that's really good." Then, wisely: "use it to sort of double check your work."
3. BuilderI say it out loud and the copilot builds it: a deliverable at its own URL that updates within a minute, or a pull request.
4. Round the tableWhen several people report in, the commentary goes and the page becomes a slate: one card per person, done, today, blocked. Blockers turn red.
5. IncidentSomething breaks mid-call. A read-only agent traces it, a coding agent in its own sandbox opens the revert, and nothing merges without me.
6. Share copyThe other side asks for the page. They get a stripped copy; the live one carries my private pushback.

Builder is my favorite. On one call I told it, out loud, to rebuild a report the client and I were going over. It went up at its own URL, and every tweak they asked for landed within a minute. Sending it still waited for my typed yes.

On another I told it to "go spin a PR" with what was being discussed right now and to follow the client's words "religiously." A coding agent got the client's quotes with timestamps, built only what I named and fact-checked every quote against the transcript. When the client changed the spec in the last minutes, the copilot kept feeding it in.

How is Kortex slipping into my clients' weeks?

Quietly, one routine at a time, never as a chatbot they have to learn.

• Inspection reviews. For an inspections company, agents reviewed the report generator in rounds against real inspections until it held up: numbered rooms, a Ref, Name, Answer, Comments table per room, the printout identical to the PDF byte for byte. Six rounds, real inspections only, because the synthetic ones hid the real bugs.
• Daily reports. Edgar, the AI engineer I built for a client, helps build two one-pagers from the client's live data: search growth and paid ads. "On target on X of 6," green only when it's actually on target, zero narrative. The client set that format on a call Kilazi Live recorded.
• Calls. The copilot opens on client calls. The client sees the result, not the machinery: a fixed report, a PR, a clean list of who owes what.

That's the bet, and the case I make in why every business needs an AI brain: capture everything, then let the brain earn its way into the routine. It starts with transcribing every client call.