Start with the install, not the pitch
Most of the first week was small stuff. The right installer for each computer. A connection that wouldn't hold. One executive asked for help the same day they tried it. Nobody needed a sales pitch. They needed it working.
Give the agents read‑only access to the data
The clinical ops VP wanted clinic-level reporting. The blocker was permission, not the agent. I gave them read-only access to just the reporting data they needed. They can read it. They can't change anything.
Make every number show where it came from
I built a check that any executive agent can run. Every number comes back with its source. If it can't prove one, it says "Not verified." That one habit does more than any prompt I've written.
Specialists, then a lead, then one assistant
Here's how I coach everyone. Give each agent one job. Have the specialists report to a lead. Have the leads report to your assistant agent. Then you only talk to the assistant.
What they run
- Sales: agents for the CRM and sales data tools, one for research, and an assistant agent with daily and weekly tasks in order.
- Clinical Ops: reports on how the clinical team is performing, plus trends and forecasts across our care programs.
- Finance and operations: operating costs, and the work of keeping the company running.
- One team's project-manager agent now tests its own automations before they go live.
A person stays in the loop
The agents do the work. A person reviews it and makes the call. Changes also go through a security review agent before a person signs off.
What's still rough
Some connections break. A tool stops working for one person and not another. I fix those one at a time. And I don't have hours-saved numbers yet. I'd rather get them from the people doing the work than guess.
Is that ahead of the curve?
I'd rather show you dates than say "cutting edge."
- On Nov 9, Anthropic released Claude Fable 5.2 and called it its strongest lead agent, built to direct other models. Same shape as one assistant agent running a few specialists.
- Zapier's survey of 500+ enterprise leaders says 71% have agents in production. Finance teams grew fastest, from 24% to 41%.
- Genpact's survey of 2,000+ executives says 79% of large companies have agents in production. Only 20% use them broadly.
Genpact also found only 15% of companies have fully tied their AI spending to profit. I'm not going to claim we're different until we've measured it.
When we have real numbers, I'll write part two. What would you want to know first?
Sources
- Anthropic, Claude Fable 5.2 and Mythos 5.2 (Nov 9, 2026)
- CNBC coverage of the same release (Nov 9, 2026)
- Zapier, AI agents survey (November 2026)
- Genpact press release on PR Newswire (Nov 12, 2026)