Most conversations about AI agents focus on what the agent can do alone. We spend more time on the moment it stops.

Every agent will meet a case it shouldn’t handle: a refund above its limit, a customer who is upset, a document that doesn’t look right. What happens next decides whether the people around it trust it. Get the handoff wrong and your team ends up redoing the agent’s work, plus an extra step. Get it right and the agent makes the person faster at the part only they can do.

A bad handoff is a forwarded email

The worst handoffs we’ve seen look like this: a ticket lands in a queue with a note that says “escalated by AI”. The person opens it, reads the whole thread, looks up the customer, checks the order, and works out what the agent already tried. They’ve done all the work the agent was meant to save them, and they’re annoyed about it.

What a good handoff carries

When a Wardlow agent hands off, the person receives four things in one place:

• The original request, exactly as it arrived.

• What the agent checked, with the data it found and where.

• What it would do, as a draft the person can approve, edit or reject.

• Why it stopped: the specific limit or rule, in plain words. “Refund of $640 is above the $500 limit” is useful. “Low confidence” isn’t.

If the draft is right, approving it takes one click, and the agent finishes the job: issues the refund, replies to the customer and logs the decision.

It arrives where people already are

Handoffs go to team chat or the Wardlow queue, to a named person or a rota, not to a shared inbox nobody owns. Each case type has a response target, and if nobody picks a case up within 30 minutes it goes to a backup. A handoff that sits unanswered is worse than no agent at all.

The number we watch

We measure how long it takes a person to decide a handed-off case. Across our customers in 2026, the median is six minutes, from the moment the case arrives to the decision.

We watch it because it tells us whether the handoff is doing its job. If the number creeps up, the agent is passing on too little context or stopping on cases that don’t need a person. If it’s very low on a case type, the limit might be too cautious, and we talk to the customer about raising it. Either way, the person stays the one who decides.

What this means for limits

Teams often start with tight limits and loosen them as they see the agent’s work. That’s the right instinct, and the handoff is what makes it safe. When the person deciding can see exactly what the agent would have done, they can tell whether a limit is protecting them or just slowing them down. Most of our customers raise at least one limit in their first quarter, based on what the Run Log shows them, not on what we tell them.

An agent that never asks for help isn’t confident. It’s badly configured. We design for the moment it asks.