The Handoff Is the Product

Every agent demo shows the happy path. Customer asks a normal question, agent answers well, everyone nods.

Nobody demos the other thing. The customer asks something slightly off, or the agent's confidence quietly drops, or the request is technically in scope but obviously needs a person.

What happens in that moment is the product. It's also the last thing anyone designs.

The 20 percent everybody remembers

Say your agent handles 80 percent of requests cleanly. Good number. Most teams would take it.

Now think about who's left. The 80 percent barely register that the interaction happened, because it went fine. The 20 percent are the people with the unusual problem, the urgent problem, or the problem that already went wrong once.

They're the ones who'll describe your company to somebody else.

Your reputation gets set almost entirely by the cases your agent couldn't handle. So the design question is what those people experience.

3 ways handoffs get built badly

The dead end. The agent says it can't help and offers a link to a contact form. Your customer starts over, from zero, in a different channel, after already spending 4 minutes explaining. Most common design, worst outcome.

The silent loop. The agent doesn't know it's stuck. It keeps answering, plausibly and confidently, while the customer keeps rephrasing. Nothing escalates because nothing registered as a failure. Worse than the dead end, because you never find out it happened.

The cold transfer. The agent hands off to a human who receives a ticket saying "customer needs help." No transcript, no context, no idea what's been tried. The customer explains it again. Now everyone's annoyed and you've added a step instead of removing one.

All 3 are engineering decisions. A better model fixes none of them.

What a good handoff does

The agent knows when it's out of its depth. Imperfectly, but it has a threshold, and crossing it triggers something. Uncertainty has to be a signal in your system rather than a vibe.

The transition gets said out loud. "I'm not going to get this right, let me bring in somebody who will." Customers are fine with that. They're much less fine discovering it after 10 minutes.

Context travels. The human gets the full conversation, what the agent tried, and what it thinks the problem is. The customer never repeats themselves. This one thing does more for the experience than any accuracy improvement you can buy.

The human can see and undo what the agent did. If it took actions, they're visible and reversible.

And it feeds back. Every escalation is a labeled example of where your system's edge actually sits. Most teams throw these away, which is a shame, because it's the highest-quality data you'll ever get about your own product and it arrives free.

[ROGER: a handoff you built or fixed. What the failure mode was, and what changed after.]

Why it keeps getting skipped

The happy path is what gets demoed, and the demo is what gets funded.

The escalation path has no wow moment. It doesn't show well in a board deck. It looks like plumbing, so it gets pushed to "phase 2," which arrives after your customers have already formed their opinion.

The deployments reaching production this year are hybrid by design, agents on the routine cases and humans on the edges. The ones stuck in pilot are usually still trying to engineer the edges out of existence.

You're building a system where an agent handles what it's good at and hands off cleanly when it doesn't. The handing off is most of the product.