One agent, one job: how we scope automation that ships

The first thing customers ask for is usually an assistant that does everything. Answer any question, touch any system, help any team. It is an understandable wish, and it almost never ships.

The agents that reach production look different. They own one process, end to end, with a clear start and a measurable finish. A claims intake agent. A records request agent. A ticket triage agent. Boring names, real outcomes.

Why narrow wins

A narrow agent has three advantages that compound.

You can define done. "Route every incoming ticket with a severity and an owner" is testable. "Help the IT team" is not. When done is defined, you can measure error rates, set an approval threshold, and know when to remove the training wheels.

You can bound the blast radius. An agent that only reads the claims queue and drafts settlement summaries can be reviewed in a week. An agent with broad system access takes a quarter, if it passes at all.

You can prove value fast. One process means one before-and-after number. Hours saved per week. Days of turnaround reduced. That number funds the second agent.

The scoping questions we ask

Every engagement starts with the same working session. We look for a process that scores well on four questions:

• Volume. Does it happen daily, in quantity? Automating a quarterly task rarely pays.
• Structure. Are the inputs consistent enough to declare? Forms, tickets, and filings, yes. Free-form strategy work, no.
• Reversibility. Can a mistake be caught and corrected before it costs money? Draft-then-approve flows score high.
• Ownership. Is there one team that owns the outcome and will actually use the thing?

A process that scores four for four becomes the pilot. Everything else goes on a roadmap, in order.

What this looks like in practice

A specialty insurer came to us wanting "AI for the claims department." Six workflows, four systems, many opinions. The scoping session narrowed it to one: intake. Extract loss details from incoming documents, verify coverage terms, draft a settlement summary, route to an adjuster.

That agent shipped in five weeks. Claims that took days now clear in hours, and the adjusters kept final say on every decision. The second and third agents came later, each scoped the same way, each faster to ship because the governance skeleton already existed.

Start narrow. Ship. Widen. It is slower for the first month and faster every month after.