Lexicost
AI voice agents / Costs, terms and calculators

Containment rate and resolution rate

Checked against its sources on

Containment rate is the share of calls an AI voice agent ends without a transfer to a person. Resolution rate is the share where the caller's problem was solved. A contained call can still fail: the caller may hang up unhelped or ring back. Some glossaries merge the two; a platform's documented success figure follows the criteria it is given.

The formula

containment rate = contained ÷ answeredtransfer rate = transferred ÷ answeredresolution rate = resolved ÷ answereddeflection rate = sent elsewhere ÷ answeredrepeat-contact rate = came back ÷ contained
  • answered: calls the AI agent picked up; the base for every rate except repeat contact
  • contained: calls that never reached a person, hang-ups included; the rest are transferred, so the transfer rate is 100% minus containment
  • transferred: calls the AI handed to a person, also called hand-offs or escalations
  • resolved: calls where the caller's need was met, shown where possible outside the call: the booking exists, the payment went through
  • sent elsewhere: calls the AI ended by pointing the caller to a web page, a text message or another channel; these count as contained too, need met or not
  • came back: callers who contacted the business again about the same need within a set window, such as 7 days

Where the contact-centre sources disagree

NiCE, which sells contact-centre software, defines containment rate simply as "the share of interactions the virtual agent handles without escalating to a human" and keeps it apart from resolution. Its quadrant puts the caller who "hung up to call again tomorrow" under "Contained and abandoned", a call "counted as success by naive containment". Its stricter count is "Honest containment = contained AND verified-resolved AND no repeat contact within the window".

Cognigy's glossary, published under the NiCE Cognigy name, puts resolution inside the word: containment rate is "the percentage of customer interactions fully resolved by an AI Agent without requiring escalation to a human agent." Five9's FAQ defines self-service containment in nearly the same words: "the percentage of customer interactions that are fully resolved by an automated system without requiring a transfer to a human agent." Under those two wordings, which require an interaction to be "fully resolved", a caller who hangs up on the AI unhelped is not contained at all.

We searched Amazon Connect's metric definitions page for "containment" on 14 November 2026 and found no metric by that name. It counts hand-offs instead, as AI Handoff Rate: "(AI handoffs / AI involved contacts) * 100.0". Resolution gets its own metric, Goal Success Rate, for sessions where the Orchestration AI Agent "successfully resolved customer issues."

Deflection is a different move. TELUS Digital's glossary describes call deflection as a strategy to redirect inquiries "away from traditional voice-based channels".

The same 100 calls, counted five ways

In this invented example, taken from no real product, an AI agent answers 100 calls in a week and hands 22 to a person. Of the 78 that never reach a person, 46 get what they rang for and do not call again within 7 days. Eleven get an answer and end the call politely, then ring back within 7 days about the same thing. Nine are told to book on the website instead, and 12 hang up partway through; none of those 21 calls back within the 7 days.

Containment is 78 ÷ 100, or 78%, and the transfer rate is the other 22%. Deflection is 9 ÷ 100, or 9%, and repeat contact on contained calls is 11 ÷ 78, about 14%. The AI's resolution rate is 46 ÷ 100, or 46%, 32 points below containment. It leaves out calls people solve after a transfer.

A success figure whose criteria come down to "the call ended without a transfer" is containment: 78%. A judge reading only transcripts passes the 46 plus the 11 polite endings with an answer given, blind to those callers ringing back days later: 57%. If the criteria also accept pointing the caller to the website, the 9 pass too: 66%. Only a check outside the call, against the booking system or the following week's calls, brings the figure down to 46%.

How each platform decides that a call succeeded

Retell AI has two places to set success criteria. AI QA: The customer writes the criteria (AI QA conditions and metrics); by default a call must meet every one. On the AI QA metrics page, Call Resolution Rate is the percentage of calls resolved according to the customer's resolution criteria. Post-call analysis: The customer writes the criteria; call_successful (Post Call Extraction) records whether a call met them.

Vapi sets a default in its analysisPlan, though its call analysis page recommends structured outputs, where a structured output or a scorecard does the evaluating. Success evaluation: by default a model judges success against objectives inferred from the system prompt. The judge gets the call's transcript and the AI participant's system prompt.

ElevenLabs allows a third result. The customer writes the criteria; each is checked against the transcript as success, failure or unknown. Whether unknowns stay in the count changes the success share: in an invented example for one criterion, 50 successes, 30 failures and 20 unknowns give 50% of all 100 calls, or 62.5% of the 80 that were decided.

Synthflow: The customer writes the criteria for what makes a successful call (custom evaluations). Its analytics dashboard reports a success rate for each assistant, and how that rate is counted is not publicly listed on the pages we read.

Bland's own definition of a successful call is not publicly listed on the pages we read. Its docs describe Dispositions, questions the customer writes to be asked of every call, such as "Did the caller book?", and a call status, completed, that covers calls answered by a person or by voicemail.

What a price per resolved call depends on

This site's records hold no price per resolved call for these five platforms. Fin, an AI agent priced per outcome, counts a resolution as one kind of outcome and defines it in its own help centre: a resolution is counted when, following Fin's last answer, the customer "either confirms the answer was satisfactory (confirmed resolution), or exits the conversation without requesting further assistance (assumed resolution)". In the example above, a fee per successful call would be owed on 78, 66, 57 or 46 of the 100 calls, depending on the contract's definition.

Often confused with

  • First contact resolution (FCR). The share of issues solved on the first contact, so any repeat contact counts against it; Amazon Connect's case version counts a closed case as resolved at first contact only when it had exactly one contact.
  • Success rate. A platform's count of calls that met the criteria set for it, which can match containment, resolution or neither.

Vendors cited on this page

Bland AI, ElevenLabs Agents (ElevenAgents), Retell AI, Synthflow AI, Vapi. Each profile lists everything that vendor publishes, with its sources. All of them are side by side in the comparison.

Sources

Checked against its sources on 14 Nov 2026. Vendor figures on this page open their source when clicked. Lexicost has no commercial relationship with any vendor named on this site.

Cost per call: AI agent or a personWorks out an AI call's cost from a vendor's own published rate and a person's from Bureau of Labor Statistics wage and benefit figures, finds the finish rate at which the AI stops saving money, and puts side by side how vendors bill a call after the AI hands it to a person.AI voice agent cost per minuteAdds up the five parts of a minute from each vendor's own published rates, with a box for any part a vendor does not price, and shows which parts each vendor's headline price leaves out.
Source for every vendor figure on this page (8)
  1. Retell AI, how the vendor defines a resolved or contained call: The customer writes the criteria (AI QA conditions and metrics); by default a call must meet every one (AI QA cohorts; weighted scoring can change how the criteria combine).You define this using two kinds of criteria: AI-evaluated conditions and performance metrics. By default, a call counts as successful only when it meets every condition and every metric. Enable weighted scoring to change how these criteria combine. docs.retellai.com/ai-qa/define-resolution-criteria Read 14 Nov 2026.
  2. Retell AI, how the vendor defines a resolved or contained call: Call Resolution Rate is the percentage of calls resolved according to the customer's resolution criteria (AI QA metrics).Call Resolution Rate: Percentage of calls that were successfully resolved according to your defined resolution criteria. docs.retellai.com/ai-qa/terminologies Read 14 Nov 2026.
  3. Retell AI, how the vendor defines a resolved or contained call: The customer writes the criteria; call_successful (Post Call Extraction) records whether a call met them (Built-in Post Call Extraction field; the page says its prompt can be customized).call_successful — whether the call met the success criteria you define. docs.retellai.com/features/post-call-analysis-overview Read 14 Nov 2026.
  4. Vapi, how the vendor defines a resolved or contained call: Success evaluation: by default a model judges success against objectives inferred from the system prompt (analysisPlan success evaluation; the prompt and rubric can be customized, and the page recommends structured outputs or a scorecard as the equivalent).Used to determine if the call was successful, stored in call.analysis.successEvaluation. Default prompt: You are an expert call evaluator. You will be given a transcript of a call and the system prompt of the AI participant. Determine if the call was successful based on the objectives inferred from the system prompt. docs.vapi.ai/assistants/call-analysis Read 14 Nov 2026.
  5. ElevenLabs Agents (ElevenAgents), how the vendor defines a resolved or contained call: The customer writes the criteria; each is checked against the transcript as success, failure or unknown (Evaluation criteria set per agent, up to 30; the Analytics dashboard shows success, failure and unknown rates per criterion).Success evaluation allows you to define custom goals and success metrics for your conversations. Each criterion is evaluated against the conversation transcript and returns a result of success, failure, or unknown, along with a detailed rationale. elevenlabs.io/docs/eleven-agents/customization/agent-analysis/success-evaluation Read 14 Nov 2026.
  6. Synthflow AI, how the vendor defines a resolved or contained call: The customer writes the criteria for what makes a successful call (custom evaluations) (Custom evaluation actions attached to an agent).Custom evaluations allow you to define and track your own criteria for what makes a successful call or conversation. docs.synthflow.ai/configure-custom-evaluations Read 14 Nov 2026.
  7. Synthflow AI, how the vendor defines a resolved or contained call: Not publicly listed. Read 14 Nov 2026.
  8. Bland AI, how the vendor defines a resolved or contained call: Not publicly listed. Read 14 Nov 2026.