The best part of an AI agent is the list of things it won't do

TL;DR — The capability list is the easy half of an AI agent. The half that makes it safe to point at real customers is the other list: what it must never do on its own, and exactly how it hands off. We design that list first. Here’s what goes on it, and how to pick the first job worth handing over.

A typical Monday

Take a B2B services firm getting about 40 new inbound conversations a week through its website form and the info@ inbox. A real person answers them between other work. Average first response during the week: around five hours. Anything landing after Friday lunch waits until Monday — call it 60 hours.

Read a month of these conversations and about two-thirds are the same six things: what does it cost roughly, do you cover our area, are you taking new clients, can you handle X, can we book a call, can you send the spec sheet. None of them need judgement. All of them need speed.

That is the shape of a job an agent should own.

The two lists

Before any logic gets written, we write two lists with the client. The first is what the agent handles on its own:

• Qualify an inbound lead against criteria you wrote down — budget band, location, company size, what they actually need
• Answer the repeat questions, from material you approved
• Book meetings against a real calendar with real availability
• Write the outcome back to the CRM the same minute
• Send the follow-up on schedule when someone goes quiet

The second list is what sends the conversation to a person, immediately:

• Anything that falls outside the agreed script
• “Can I talk to a human” — always, no persuasion attempt first
• Low confidence about what the person actually means
• A decision above a value threshold you set — a discount, a refund, a commitment
• Frustration, hostility, or anything that reads like a complaint brewing

The second list is the product. Anyone can demo an agent that answers questions; what makes one safe to point at real customers is that it never improvises above its level. A slow handoff costs you a few minutes. A confident wrong answer to a hot lead costs you the deal — and you find out weeks later, if at all.

The handoff has to be designed, not implied

“Escalates to a human” is not a feature line, it’s a design decision with three parts:

What triggers it. The second list, plus anything the agent can’t classify confidently. When in doubt, it escalates — the default direction of uncertainty is toward a person, never toward a guess.

What travels with it. The full transcript, the qualification answers collected so far, and what the agent already tried. The person picking up should never have to ask the customer to repeat themselves. A cold transfer is a failed handoff.

Who gets paged. A named owner with a real notification, not a shared inbox someone checks twice a day. Escalation that lands nowhere is just a slower way to ignore the customer.

Picking the first job

Not every repetitive task is a good first agent. The ones that work share five traits:

1. High volume, same shape. Dozens of times a week in roughly the same form — that’s what makes both the economics and the reliability work.
2. A clear definition of done. Booked, qualified, answered, synced. If you can’t say what “handled” means in one sentence, the agent can’t either.
3. Low blast radius. A mistake is recoverable — a misrouted follow-up, not a wire transfer, a contract term, or a medical answer.
4. A human is reachable for the tail. The 10-20% that escalates has to land with someone who can actually act on it.
5. Measurable before and after. First-response time, share handled without a person, meetings booked. If you can’t measure it now, you won’t be able to prove it worked.

Good first jobs: inbound triage, lead qualification, the repeat questions, keeping records in sync between tools. Bad first jobs: anything moving money, anything contractual, angry customers, or a process nobody has ever written down — the agent will faithfully automate the mess.

The arithmetic

Back to the 40 conversations. Say the agent fully handles two-thirds — 27 of them. At roughly six minutes of someone’s time each, that’s about 2.5 hours a week handed back. Useful, but not the real win.

The real win is that first response drops from five hours — or a whole weekend — to under a minute, every time, including 11pm on a Sunday. And every outcome lands in the CRM the same minute, qualified and tagged, instead of whenever someone gets around to it. Speed and consistency are what you’re buying; the hours are a bonus.

What the first two weeks look like

No agent should meet all your traffic on day one. We run a limited pilot first — a slice of real conversations — and watch every one. The boundary almost always moves: inward at the start (escalate more than strictly necessary), then outward as the agent proves itself on real phrasing, real edge cases, real typos. Two weeks of that monitoring and tuning is part of the build, not an extra.

By the end of it you have something genuinely boring: a job that just gets done, and a team that only sees the conversations that actually need them.

---

If you have one job that fits the five traits above, that’s a scoping conversation. We build agents that own one workflow end to end — and hand over the moment they should: AI agent development.