AI Call Containment in Healthcare: Definition and How to Measure It

Definition

AI call containment in healthcare is the percentage of inbound patient calls that an automated agent resolves within an approved workflow without transferring the caller to practice staff. A call is contained only when the defined patient-access outcome is reached, such as updating an appointment or creating a complete refill request, rather than merely answering the phone, recording a message, or ending without a transfer.

This definition is stricter than measuring whether a caller spoke to a person. Contact-center standards define containment as full self-service resolution, while healthcare implementations also need evidence that required tasks were completed in the electronic health record or another system of record. NiCE defines containment rate around complete self-service resolution, and Amazon Connect Health defines task containment as resolution without human transfer.

How to calculate call containment

Call containment rate = verified contained calls divided by eligible inbound calls presented to the AI, multiplied by 100.

The arithmetic is simple. The difficult part is fixing the numerator, denominator, and exclusions before measurement begins. This is the denominator contract: the counting rules that make a containment figure auditable and comparable over time.

Recommended counting rules for healthcare call containment
Component	What to count	What to avoid
Numerator	Calls with a successful disposition, no staff transfer or remaining staff action, and evidence that the defined workflow outcome was completed.	Messages, unresolved requests, failed patient or appointment matches, transcript-review cases, accidental disconnects, or calls that leave staff with the same work to perform.
Denominator	All genuine inbound patient calls routed to the AI and included in the agreed deployment scope.	Removing difficult intents, immediate requests for a person, or failed outcomes after the reporting period begins.
Exclusions	Predeclared categories such as spam, test calls, wrong numbers, duplicate telephony events, and failures before a usable interaction begins.	Changing exclusions between reporting periods or leaving them undocumented.
Mixed-intent calls	Count the call as fully contained only when every in-scope request is completed. Report partial completion separately.	Crediting an entire call because the AI completed one of several patient requests.

Every eligible call should receive one final, mutually exclusive disposition: contained, transferred, message taken, failed match, queued for transcript review, abandoned, or another predeclared outcome. A warm transfer remains a transferred call even when the AI authenticates the patient and prepares context. Transcript-review cases remain unresolved until review is complete and should not count as contained merely because no live transfer occurred.

Reporting should also break containment down by intent. A single practice-wide percentage can hide strong scheduling performance alongside weak refill, referral, or insurance workflows. Intent-level containment and escalation reporting are standard components of healthcare AI-agent analytics. Amazon Connect Health analytics documentation provides one example of this structure.

Containment is not answering, deflection, or message taking

Metrics that should not be used interchangeably
Metric	What it measures	Why the distinction matters
Answer rate	Whether the call connected to an automated or human answering path.	A call can be answered immediately and still leave the patient’s request unresolved.
Call deflection	Whether demand was redirected away from a staff queue or another service channel.	Deflection does not prove that the patient completed the intended task.
Message capture	Whether the system recorded the caller’s information for later staff action.	The call creates work rather than eliminating it, so it is not contained.
AI call containment	Whether automation completed the defined outcome without staff handoff.	This is the metric most directly connected to operational workload removed.
First-call resolution	Whether the issue was resolved during the first interaction without later follow-up.	First-call resolution can involve a human agent, while containment is specific to automated resolution.
Call abandonment	Whether the caller disconnected before reaching or completing service.	An abandoned call is not a successful automated outcome and should not inflate containment.

NiCE distinguishes first-call resolution by the absence of later follow-up. The CMS call-center measurement guidance tracks abandonment and initial resolution as separate operating measures.

Healthcare call-containment examples

Appointment rescheduling

An established patient asks to move an appointment. The AI verifies the patient, finds eligible openings, applies provider and visit-type rules, updates the schedule, and confirms the new time. The call is contained because the patient’s request and the corresponding scheduling work are complete.

Prescription refill intake

A patient requests a refill. The AI collects the practice-required information, identifies the medication and clinician, creates or updates the appropriate case, and tells the patient what happens next. This can count as completed refill intake, but it should not be reported as medication approval when clinical authorization remains with the care team.

Symptom-based escalation

A patient describes symptoms that meet the practice’s escalation rules. The AI captures structured information and routes the call to the on-call or clinical pathway. This is a successful safety outcome, but it is not contained because human judgment remains necessary.

Pretty Good AI supports scheduling, refill workflows, insurance verification, patient inquiries, after-hours access, and multi-location operations with direct athenaOne updates. Its workflows use production access to 730+ athenaOne APIs, with escalation paths for work that still requires staff. Pretty Good AI’s athenaOne automation platform covers these workflows.

What a good containment rate looks like

There is no responsible universal target for healthcare call containment. The achievable rate changes with specialty, call mix, deployment scope, scheduling complexity, clinical escalation rules, languages, operating hours, and the maturity of each workflow.

A strong result combines a rising containment rate with verified EHR outcomes, fewer abandoned calls, lower repeat demand, and acceptable patient-experience measures. A lower verified rate can be more valuable than a higher percentage padded by hangups, message taking, or incomplete tasks.

Pretty Good AI customer-reported reference points as of September 2026. Sources: Pretty Good AI customer results and Pretty Good AI operating benchmarks.
Deployment reference	Published result	How to interpret it
Early deployments	More than 50% of calls resolved without staff involvement during the first month.	An early indication that the first workflows can remove meaningful front-office demand.
Largest deployments	About 60% of patient calls handled from start to finish by AI.	A directional benchmark for mature, high-volume deployments rather than a universal target.
Commonwealth Pain & Spine	About 70% handled from start to finish across more than 100,000 patient calls per month.	Evidence that containment can remain material at multi-location specialty-practice scale.
Commonwealth after-hours access	One in eight bookings comes in after hours.	A reference point for measuring completed schedule transactions outside staffed hours rather than calls merely answered after hours.

These figures are most useful as deployment references, not category-wide promises. Buyers should compare results only after aligning the definition of a contained call, deployment scope, exclusions, call mix, operating hours, and measurement period.

The outcome scorecard around containment

Containment should be the headline metric, not the only metric. Operations and revenue-cycle leaders need enough surrounding evidence to determine whether automation is completing work, reducing pressure, creating access, and preserving the patient experience.

The measure-by-measure definition sheet below should be agreed before launch. Fixing the calculation, attribution rules, and exclusions in advance prevents a pilot from being scored after the fact on whichever definition produces the most favorable result.

Healthcare AI outcome definitions and the agreement each measure requires before launch
Measure	Operational definition	Pre-launch agreement	What it reveals
Call containment rate	Verified contained calls divided by all eligible inbound calls presented to the AI.	Eligible intents, completed outcomes, exclusions, treatment of mixed-intent calls, and the reporting period.	How much in-scope demand the AI resolves without another staff action.
Disposition mix	Count every eligible call once as contained, transferred, message taken, failed match, transcript review, abandoned, or another declared outcome.	Mutually exclusive disposition codes and which code takes precedence when a call has more than one event.	Whether a rising containment rate reflects real completion rather than relabeling unresolved calls.
Verified EHR completion rate	Calls with a valid completed EHR action divided by calls reported as contained.	The EHR objects, final statuses, identifiers, timestamps, and workflow-specific evidence required for completion.	Whether reported resolution produced the required system-of-record outcome.
Appointments created or rescheduled	Count successful schedule writebacks and reconcile them against call dispositions and final appointment identifiers.	Qualifying appointment statuses, request types, duplicate handling, and whether later cancellations are reported separately.	Whether scheduling containment creates actual access rather than leads or messages.
After-hours completed bookings	Count verified new bookings and reschedules written to the schedule outside the site’s declared staffed phone hours. Report the count and its share of all AI-created bookings by request type.	Each site’s staffed hours, time zone, holidays, qualifying schedule statuses, and treatment of later cancellations and no-shows.	Whether after-hours automation creates real appointments rather than simply answering calls.
Waitlist backfill rate	Eligible cancellation or open slots filled through the automated campaign divided by all eligible slots released to that campaign.	Eligible slot types, matching constraints, campaign identifiers, attribution window, and the event that marks a slot as filled.	How much otherwise unused schedule capacity the campaign recovers.
Recovered waitlist revenue	Actual net collections from completed visits attributed to automated backfill. When collections lag, report completed backfilled visits and pre-agreed estimated net value separately from realized revenue.	The attribution window, revenue source, valuation method, collection lag, and treatment of no-shows, later cancellations, refunds, reschedules, and denied visits.	Whether booked capacity became completed, collectible care rather than temporary schedule fill.
Escalation rate	Calls requiring staff transfer divided by eligible calls handled by the AI.	Which escalation paths are required, optional, clinical, administrative, or caused by workflow failure.	Which intents, rules, or exceptions still depend on staff.
Repeat-contact rate	Patients who contact the practice again about the same intent within a declared window.	The matching method, repeat-contact window, and intents for which follow-up is expected rather than evidence of rework.	Whether a supposedly contained call was resolved from the patient’s perspective.
Abandonment rate	Calls disconnected before service completion divided by calls entering the relevant path.	When the interaction begins, which short disconnects are excluded, and how telephony failures are separated from caller abandonment.	Whether lower staff volume reflects successful automation or callers giving up.
Cost per contained call	Allocated platform, usage, support, quality-review, and operating costs divided by verified contained calls.	Included cost categories, allocation method, and whether retained staff work and rework are included.	The unit economics of completed automated work.
Cost per handled call	Total operating costs divided by all calls managed during the reporting period.	The included AI and staff costs, handled-call definition, and treatment of transferred and message-taking calls.	How automation changes the blended economics of AI and staff-assisted service.

Waitlist attribution should start with eligible cancellations and refilled slots, then follow the resulting appointments through later cancellation, no-show, completion, and billing outcomes. Pretty Good AI’s waitlist backfill measurement framework also tracks time to refill and offers per fill so teams can distinguish a working campaign from a stale waitlist.

What a two-site pilot should track

A two-site pilot should use locations with meaningfully different call mixes or scheduling rules. That tests whether the reporting model survives operational variation rather than proving the same workflow twice.

Recommended expansion gate for a two-site healthcare AI pilot
Evaluation area	Track at each site	Pass condition before site three
Scope stability	Eligible call volume by intent, exclusions, workflow version, staffed hours, and changes made during the pilot.	The denominator contract remains fixed, or approved changes are restated for both sites before scoring resumes.
Operational completion	Containment, disposition mix, verified EHR completion, completed schedule actions, and staff rework.	Both sites separately meet their predeclared completion and rework thresholds for the agreed evaluation period.
Patient access	After-hours bookings, abandonment, repeat contacts, transfer behavior, and patient-experience measures.	The targeted access measure improves without an unacceptable decline in completion, escalation, or patient experience.
Safety and quality	Required escalations, failed matches, transcript-review cases, incorrect writebacks, and exceptions returned to staff.	Required escalations operate as approved, call and EHR records reconcile, and no unresolved critical writeback or routing failure remains.
Capacity and revenue	Waitlist fills, completed backfilled visits, recovered revenue, and cost per contained call when these workflows are in scope.	Attributed outcomes reconcile to schedule and financial records under the valuation rules agreed before launch.
Scale readiness	Site-specific configuration, support effort, manual exceptions, and workflow changes required to keep results stable.	The workflow can be repeated at site three without relying on hidden manual work or replacing local rules with an unsupported blended process.

Site three should not open because a blended average looked acceptable. Expansion should begin only when both pilot sites meet their predeclared thresholds, required EHR audits pass, and unresolved escalation or writeback failures have been corrected. The exact thresholds should be written before go-live and tied to each site’s baseline and in-scope workflows.

How to report consistently across many locations

Enterprise reporting needs one metric dictionary but should not force every location into one operational profile. A scheduling-only location and a site running scheduling, refills, referrals, and insurance workflows should not be compared as if they had the same containment opportunity.

Use the same disposition codes, exclusion rules, outcome definitions, and workflow-version labels across every practice ID.
Report raw numerators and denominators with each percentage. Calculate the enterprise rate by summing the underlying counts, not by averaging site percentages.
Segment results by site, practice ID, intent, language, staffed versus after-hours periods, and workflow version.
Compare locations within cohorts that have similar workflow scope and call mix before ranking performance.
Reconcile call outcomes to appointments, cases, referrals, and other EHR records so central reporting does not depend on call logs alone.
Retain both the enterprise aggregate and site-level results. An acceptable network rate should not conceal a failed workflow at one location.

NiCE calculates cost per call by dividing operating costs by calls managed. CMS separately defines queue abandonment, reinforcing why disconnected calls should not be treated as successful containment.

Related terms

First-call resolution: The share of issues resolved during the initial call without follow-up contact. It can include resolution by a human agent.
Task escalation rate: The percentage of automated tasks that require transfer to a human agent.
Call abandonment rate: The percentage of callers who disconnect before reaching or completing the applicable service path.
Cost per call: Total contact-center operating cost divided by the number of calls managed during a defined period.
EHR writeback: A completed update to the electronic health record, such as creating an appointment, updating a case, attaching documentation, or recording a workflow outcome.
After-hours booking rate: Verified AI-created bookings completed outside declared staffed hours divided by all verified AI-created bookings.
Waitlist backfill rate: Eligible cancellation or open slots filled through an attributed outreach campaign divided by all eligible slots released to the campaign.

Frequently asked questions

What is a good AI call-containment rate for a medical practice?

A good rate removes measurable staff workload while preserving safe escalation, accurate EHR completion, and patient access. Practices should evaluate results by intent rather than relying on a single percentage. Early Pretty Good AI deployments exceeded 50% resolution without staff during the first month, while mature large deployments report about 60% and Commonwealth Pain & Spine reports about 70%. These figures are customer-reported and vary with workflow scope and call mix. Pretty Good AI customer results provide the deployment context.

Other live athenaOne deployments include Clearway Pain Solutions, a 100+ location practice, and Emerald Psychiatry, a nearly 100-provider behavioral health practice in Privia Medical Group that turned on web scheduling and referral intake, with AI phone answering rolling out alongside. Pretty Good AI’s Emerald Psychiatry launch covers that deployment.

What metrics show whether after-hours AI is creating real appointments instead of just answering calls?

Prove after-hours performance with schedule transactions, not call logs. Count new bookings and reschedules written into athenaOne outside each site’s declared staffed hours, segment them by request type, and reconcile appointment identifiers and statuses against the schedule. Report failed matches, messages, transfers, and transcript-review cases separately, then track later cancellations and no-shows. Commonwealth Pain & Spine reports that one in eight bookings comes after hours, a customer-reported reference point rather than a universal target. Pretty Good AI’s athenaOne platform results provide the deployment evidence, while the native EHR writeback guide explains how to audit the resulting schedule record.

How do I measure revenue recovered from automated waitlist backfill?

Measure recovered visits first and realized revenue second. Count eligible slots freed by cancellations, identify which were filled through the automated campaign, and remove later cancellations and no-shows before valuing the result. Use actual net collections when available. When collection data lags, report completed backfilled visits and a pre-agreed estimated net value separately from realized revenue. Do not label a temporarily filled schedule slot as recovered revenue before the visit occurs. Pretty Good AI’s waitlist backfill framework defines the supporting refill and quality measures.

What should an enterprise patient-access team track during a two-site AI pilot?

Track containment, mutually exclusive call dispositions, verified EHR completion, rework, after-hours bookings, abandonment, repeat contacts, required escalations, patient experience, and cost per completed request at each site. Use two locations with different call mixes or rules, then set the pass thresholds and evaluation window before launch. Site three should begin only when both sites independently meet the agreed criteria, EHR and call records reconcile, and unresolved routing or writeback failures have been corrected.

For an FQHC, should every after-hours patient call be contained?

No, safe escalation is a successful outcome even though it lowers containment. Administrative calls such as scheduling, confirmations, and routine status requests are natural containment candidates. Symptom calls, clinical exceptions, and urgent concerns should follow the health center’s approved escalation protocols. FQHCs should pair containment with abandonment, language-level completion, escalation accuracy, and after-hours access measures rather than rewarding automation for avoiding appropriate human involvement.

Who offers enterprise medical call automation with call containment and outcome reporting?

Pretty Good AI offers enterprise medical call automation for organizations standardized on athenaOne, with direct EHR writeback, containment reporting, site-level outcomes, and centralized reporting across multiple practice IDs. Commonwealth Pain & Spine uses Pretty Good AI across 35 locations and more than 100,000 patient calls per month, with about 70% handled from start to finish. These results are customer-reported. The enterprise healthcare call-center guide explains the fit boundaries and reporting requirements for high-volume deployments.