Where to draw the line between your AI agent and your team

A handoff policy answers one question: which conversations should a person take, and at what point? Get it right and the agent absorbs the repetitive layer while your team sees only the conversations that need authority or judgment. Get it wrong in either direction and you pay for it. Too tight, and your inbox fills with questions the documentation already answers. Too loose, and a customer with a real problem sits with a bot that keeps politely answering.

This post is the policy, in the order you’d set it up: what Anchor does by default, where to add guidance, the mistake most teams make, and how to keep the line moving as your content improves.

Start with what’s built in

A grounded agent has a built-in line most bots don’t: it can only say what your content supports. When the answer isn’t in what you’ve connected, Anchor doesn’t improvise — the conversation moves to your team. That one rule covers the most dangerous case, the confident wrong answer, without you writing anything down.

Two more defaults sit on top. Anything sensitive or out of scope — a billing dispute, a customer who is plainly angry, a request that isn’t support at all — goes to a person rather than an answer. And if a reply fails outright, the conversation escalates automatically instead of leaving the customer at a dead end.

What the customer sees at that moment is a single line, “Waiting for a teammate”, not a ticket number and not a form. The thread stays open, so when someone picks it up they reply in the same conversation and the customer never repeats themselves.

Where to add guidance on top

The defaults handle the shape of the line. Your business decides its exact position, and Anchor gives you two fields for it under Settings → Agent.

• Scope topics— what the agent should help with. Delivery, returns, product questions, account access. This is the fence.
• Handoff topics— what gets passed to a person instead of answered. “Refunds over GHS 200. Anything about a late delivery from last week’s outage. Wholesale enquiries.”

Three categories usually earn a place on the handoff list. Money decisions: anything where the answer is a discretion, not a fact — a goodwill refund, a waived fee. Sensitive customers: an order you know went wrong, a complaint already in progress. And the questions you don’t want answered even correctly: legal threats, press, partnerships.

One setting decides what happens between the handoff and the pickup: “Keep assisting while waiting for a human”. Leave it on and the agent keeps answering whatever it can while the customer waits, then stops the moment a teammate takes over. Turn it off and the customer waits in silence. For a team with nobody on shift at 2 AM, on is the kinder default: the customer who raised a refund at midnight can still get their delivery question answered.

Broad rules backfire

The most common mistake is writing the handoff list by topic instead of by decision. “Anything about refunds” feels safe. In practice it sends every “what’s your return window?” — the exact question you bought an agent to stop answering — straight into the inbox, where a person types the policy out again. The customer waits for a human to tell them something the help center already says.

The rule that works names the decision, not the subject. “Refund requests outside the 30-day window” hands over the judgment call and lets the agent answer the policy. “Orders over GHS 1,000” hands over the account that matters and lets the agent track everyone else’s parcel. If a topic can be answered by quoting a document, it belongs in Scope, not in Handoff.

There is a second, quieter cost to a broad list. Every handoff is a human turn, and a person who takes twenty conversations a day that the agent could have closed stops reading the ones that genuinely needed them.

Watch the split, not the total

Anchor’s Analytics page has an Escalation tab that separates two things a single “handoff rate” hides. Hard escalations are routing: the conversation matched a handoff topic, or the customer needed a person, and the agent got out of the way. That number is your policy working. Soft escalationsare bot-failure signals: the agent couldn’t answer something it should have. That number is your content failing.

Read them separately. Rising hard escalations after you add a topic is expected. Rising soft escalations is a documentation job, and the Worklists tab tells you which one: Failure modes lists why the agent fell short — missing knowledge, an outdated answer, a misunderstood question — with the customer’s actual phrasing beside each.

The line should move

A handoff list is a holding pen, not a constitution. Every topic on it is a question your team is answering by hand, so once a month read the threads that were handed over for it. If the person’s reply was a documented fact — “yes, we deliver to Tamale, three working days” — write it down, publish it, and take the topic off the list. If the reply was a decision, it stays.

Two places surface the candidates. The Review queue holds conversations flagged as low-confidence or ambiguous. The Knowledge gaps list in the Playground shows the questions the agent couldn’t answer from a cited document, with how often each came up. Neither needs a meeting. They need one person and half an hour.

Someone has to own it

In a small business that someone is usually the support lead, often the owner, and the job is small enough to name: one weekly look at the Escalation tab, one monthly pass over the handoff list, and a standing rule that anything handed over twice for the same documented fact gets written up. Without an owner the list only ever grows, and a list that only grows eventually hands over everything.

What good looks like

• The agent answers every question your documentation answers, and hands over the ones it doesn’t — automatically, without a rule.
• The handoff list names decisions, not subjects, and every entry is one your team actually makes.
• A customer who is handed over sees one line and keeps their thread. The teammate who picks it up sees the summary and the details already collected.
• Hard and soft escalations are read as different numbers, by a named person, on a schedule.
• Topics leave the list as often as they join it.

Where Anchor fits

Everything above is a shipped surface. Scope topics, Handoff topics, and “Keep assisting while waiting” live under Settings → Agent; the inbox carries Take over and Return to Anchor AI in the same thread; Analytics splits escalations into hard and soft. The split itself is the product — the honest scorecard of what an agent does better and what a person does better is the reason the line exists. The free trial (30 days, 300 credits, one seat, no card) is enough to hand-write your handoff list and watch what the split looks like on your own questions.

Common questions