Contents5 sections
Field notesARTICLE
A Falling Handover Rate Is Not Automatically Good News
Everyone asks what percentage the assistant handles alone. It is the wrong first number, and the direction it moves in tells you nothing until you cut it by topic.
Costa4 min read
Short answer
The share of conversations an assistant closes without a person is not a quality score, and a falling one can mean either that the rules improved or that the escalation triggers were loosened. The readable numbers are handovers cut by topic, time to the first human sentence, how often a customer comes back after a handover, and how many handovers a person answers with something the assistant already had.
Key facts
- Handovers are recorded with the reason that triggered them, so the rate can be cut by topic rather than read as one number.
- Requests that hit a hard limit are counted separately from requests the assistant could not answer.
- The time measured is to the first sentence written by a person, not to the moment the conversation was assigned.
- A conversation that returns to a person twice is flagged, because the second handover has a different cause than the first.
The first question a buyer asks about an assistant is what percentage of conversations it closes on its own. The number is easy to produce, it moves in a satisfying direction, and on its own it says almost nothing.
Two businesses can report the same figure while one of them is well run and the other has quietly stopped escalating complaints.
The number people ask for first
A single handover rate mixes three unrelated populations: requests the assistant answered, requests it could not answer, and requests it is forbidden to answer.
The third group has a floor built by you on purpose. If refunds, disputes and anything about a person's tenancy go to a human by policy, then a share of your traffic will always be handed over, and it should be. Reporting that alongside coverage gaps produces a number that cannot go anywhere useful.
Two ways the same drop happens
Say the rate halves between month one and month two. Two stories fit.
In the first, week one did its job. The price rule got written properly, the parking answer got added, the word rethink got a trigger, and questions that used to reach a person now have documented answers behind them. The drop is concentrated in a handful of topics and the sensitive ones are untouched.
In the second, somebody adjusted a threshold. The assistant now attempts things it used to pass on, the dashboard improved, and nobody noticed until a complaint got answered in a friendly paragraph. The drop in this version is even across all topics, including the ones that have a floor, which is the tell.
Same headline movement, opposite meaning. The only difference visible in the data is where the drop sits.
Four cuts that mean something
By topic. The only cut that separates the two stories above, and the one to build the report around.
Time to the first human sentence. Not time to assignment. A conversation assigned in eight seconds and answered in fifty minutes is a fifty-minute handover as far as the customer is concerned, and assignment latency is the number that is easiest to make look good.
Return rate after a handover. A customer who comes back the next day with the same question was handed to someone who did not close the loop. This one is quiet, and it is where trust actually leaks.
Handovers a person answers from the rules. If the human reply is something the assistant already knew but did not feel entitled to say, that is a coverage gap with a name and a fix. This cut generates the weekly edit list, and it is the most productive report in the whole set.
The rate you should actually want
The one your limits list implies, plus whatever genuinely novel demand arrives that week.
Written out, that is not a target anyone can chase, which is the point. A business that treats the handover rate as a target ends up with an assistant that is confident about refunds, and a business that treats it as a description ends up with better rules.
When somebody is optimising the number
Two symptoms, both easy to check.
Handovers fall in a topic that appears on the limits list. That is not improvement, it is a limit that moved, and it should have moved in the configuration with the owner's knowledge rather than by drift.
And the average time to a person drops while the return rate climbs. Faster handovers into a queue nobody is watching, which is the operational version of the same problem: what actually happens at the moment of the handover matters more than how quickly it was recorded.
Handover is not the assistant failing. It is the assistant doing the second thing it was built to do, and the rate is a map of where your written rules currently stop.
Questions this raises
- What handover rate should we be aiming for?
- Not zero, and not a fixed figure. A business whose limits list includes refunds and complaints has a floor built into it on purpose, and pushing below that floor means somebody moved a limit.
- Our rate dropped by half in month two. Is that good?
- Cut it by topic before deciding. If the drop is in questions with documented answers, the rules got better. If it is spread evenly across every topic including the sensitive ones, somebody widened what the assistant is willing to attempt.
- Is a rising rate always bad?
- It usually means new demand you have not written rules for yet, which is information rather than damage. A rise concentrated in one topic in one week is often a product change nobody told the assistant about.