How accurate is an AI phone agent at booking jobs?
Accuracy depends far less on the model than on how much ambiguity you left in the booking. Agents are reliable at capturing what the caller said and writing it into a defined field. They are unreliable at the judgment steps: choosing the right job type from a messy taxonomy, deciding whether a symptom is urgent, and resolving which of three similar customer records is the caller. Fix those upstream and the booking gets accurate.
Booking is four decisions, not one
Every booked job is really a chain: who is this, what do they need, when can we go, and what do we tell them. A voice agent can be strong on one link and weak on another, so a single accuracy number hides where the problem is.
- Identity. Existing customer or new? Which service address? This is where duplicate records get created.
- Classification. Mapping plain speech to your job type list. This is where most errors actually live.
- Scheduling. Choosing a slot the board can serve, with the right skill and the right window.
- Commitment. What the caller was told would happen. Errors here are the ones customers remember.
Your job type taxonomy is usually the real problem
Most field service systems accumulate job types over years. Near-duplicates, seasonal one-offs, types that only one dispatcher understands, types that differ by business unit but share a name. A human dispatcher navigates that by memory and context. A model has to pick from the list you gave it.
Cleaning and collapsing that list before deployment does more for booking accuracy than any amount of prompt tuning. Where the list cannot be collapsed, the fix is to have the agent book a safe general type and let dispatch reclassify — a known, cheap correction instead of a confident wrong answer. The same taxonomy discipline shows up in field service integrations generally.
Measure it against corrections, not transcripts
The honest measure of booking accuracy is not whether the transcript reads well. It is how often a human had to touch the job afterward: reclassify it, fix the address, move the slot, call the customer back to re-ask something.
Instrument that from day one. Every dispatcher edit within a window of an agent-created job is a data point. The edit reasons cluster fast, and the clusters tell you exactly which of the four decisions is failing. Call analysis on the same conversations tells you why.
Where to accept a human
Some bookings should never be automated regardless of how good the agent gets: multi-unit properties, commercial accounts with contract terms, warranty work, anything involving a prior job that went badly. These are low volume and high consequence — the worst possible profile for automation. Route them out by rule and spend the automation budget on the routine majority.
Topics: booking accuracy · job types · intake · data quality
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.