What is a good containment rate for an AI phone agent?

Almost nobody writes about this metric, which is odd, because it is the one number that tells you whether an AI phone deployment is working.

That second sentence is the whole post. The arithmetic is trivial and the definition is where all the disagreement lives.

What containment rate should a practice expect?

For a single location primary care or family medicine practice on a real EHR integration, 60 to 75 percent at steady state is a good outcome. Above 80 percent, ask hard questions about the definition. Below 50 percent after three months, something is misconfigured.

Here is the one measured data point I can publish. At Synergy Medical, a family medicine practice in Michigan running athenahealth, MedPhone handled 2,848 inbound calls between June and mid August 2026. Containment at launch was 53 percent, measured from the call log. The current figure is an estimated 65 to 75 percent.

I am labeling the second number as an estimate rather than rounding it up into a headline, because the honest position is that the launch figure is measured and the current range is derived. Every vendor in this category quotes the steady state number. Very few tell you where they started.

Why does launch containment understate steady state?

Four reasons, and none of them are the model getting cleverer.

• Transfer rules start conservative. A sensible deployment routes more to staff in week one than it needs to, then tightens as the transcripts show what the agent handles cleanly. Every one of those cautious transfers is an uncontained call
• Practice specific vocabulary has not been learned yet. Provider name pronunciations, local place names, the way your patients actually describe your locations. These are configuration, and they get fixed in the first fortnight
• The long tail of call types has not been seen. Week one shows you the common calls. Week five shows you the unusual ones, and each one that gets handled adds to containment
• Availability rules are usually wrong at first. Appointment types, durations, which providers take new patients, slots held back for triage. Getting these right converts a category of transfer into a category of booking

The practical consequence is that the evaluation point should be week six, agreed before you start. A practice that judges the deployment in week one is judging the configuration backlog rather than the product.

How do vendors inflate the number?

Four definition games. All four are technically defensible and all four produce a number that does not mean what a practice owner thinks it means.

Ask all four in one go and you will learn more in ninety seconds than in an hour of demo. A vendor whose number survives all four is quoting something real.

What drives containment improvement?

In rough order of how much movement each one produces.

1. Transfer rule tuning. The largest single lever, and it is pure configuration. Tightening rules that were deliberately cautious at launch converts transfers into contained calls
2. Provider availability accuracy. If the agent cannot see a slot it cannot offer it, and every wrongly hidden slot becomes a transfer for something that should have been a booking
3. Call type coverage. Adding a supported type, for example a specific pre visit question that your patients ask constantly, moves the number visibly
4. Pronunciation and vocabulary. Small individually, and collectively worth several points in a practice with unusual provider names or a distinctive local patient population
5. Patient familiarity. Repeat callers who know how to interact with it get through faster and abandon less. This one takes months and is the only item on the list you cannot accelerate

Notice that four of five are configuration. That is the thing to take away. Containment is largely bought with attention during the first six weeks rather than with a better model.

What containment is realistic by specialty?

These are expectation ranges rather than measured benchmarks, and I want to be explicit about that. They come from reasoning about call mix, not from a dataset across specialties, and the only measured figures I hold are from family medicine. Treat them as a starting hypothesis to check against a vendor with real data in your specialty.

The pattern is simple enough to state without the table: containment tracks how much of your call volume is routine scheduling. A practice whose phone is mostly bookings will land high. A practice whose phone is mostly clinical questions will land low, and should, because those calls belong with a person.

What containment does not tell you

Three limits, because a metric quoted without its limits becomes a target and then stops being useful.

It says nothing about quality. A deployment can contain 80 percent of calls and produce patients who found the experience frustrating. Read transcripts alongside the number, particularly the contained ones, because those are the calls nobody complained about at the time.

It says nothing about revenue. Containment counts calls resolved, not patients booked. A practice can raise containment by answering more routine questions while its new patient booking rate stays flat, which is an operational win and not a commercial one.

It is not comparable between practices without the definition attached. Two vendors quoting 70 percent may be measuring completely different things, and two practices with the same vendor will differ by call mix alone. It is a metric to track against your own baseline, not a league table.

The number that does track revenue

Questions people ask

What is a good containment rate for an AI phone agent?

For a single location primary care practice, 60 to 75 percent at steady state is a good result. Above 80 percent, check the definition being used. Below 50 percent after three months, something in the configuration is wrong rather than something in the product.

How is containment rate calculated?

Calls resolved without the caller reaching a person, divided by all inbound calls the agent answered. The denominator is where vendors differ. Insist on all inbound calls rather than only the ones the agent attempted.

Does a voicemail count as a contained call?

It should not. A message that a staff member has to action tomorrow is work moved, not work removed. Some vendors count it anyway, which is worth ten to twenty points on the headline figure.

How long does it take to reach steady state containment?

Six to ten weeks in the deployments I have seen, with most of the movement in the first four. At Synergy Medical it started at 53 percent, measured, and now sits at an estimated 65 to 75 percent.

Is a higher containment rate always better?

No. Containment can be raised by making transfer rules more aggressive, which means callers who wanted a person had to work harder to get one. If containment goes up and patient complaints go up with it, the number is being gamed against you by your own configuration.

What should we track alongside containment?

Transfer reasons, so you know what the remaining share consists of. Failed calls, meaning anything that ended without resolution or transfer. And your original abandonment rate, because the point was never containment for its own sake, it was answering calls you were previously missing.