How often should a business review its voice AI metrics?

A business should review containment rate, cost per call, and automation rate weekly for the first 90 days after launch, then monthly once call volume stabilizes. Balto recommends breaking containment out by intent at each review, since a single blended number hides weak performance on complex call types like complaints.

What is a red flag in voice AI metrics?

A red flag is containment falling below 60 percent on standard support flows, a threshold Hamming's enterprise voice AI guidance treats as a warning line. Another red flag is a low cost per call paired with rising repeat-contact rates, which signals the AI is closing calls without actually solving the problem.

Does a higher automation rate always mean lower operating cost?

No. Automation rate only counts the share of calls handled by AI, and if those calls generate repeat human contact afterward, the business pays twice: once for the automated attempt and again for the live-agent resolution, which raises blended cost per call instead of lowering it.

Which call types should a business automate first with voice AI?

A business should automate high-volume, low-complexity intents first: appointment scheduling, order status, password resets, and balance lookups. IrisAgent reports these use cases reach 65 to 85 percent containment in production, well above the 45 to 60 percent industry average across all call types combined.

What Are the Key Metrics for Voice AI Success? Automation Rate, Containment Rate, and Cost per Call Explained

A glossary explainer on the three metrics that determine whether enterprise voice AI is actually working: automation rate, containment rate, and cost per call, with 2026 benchmark ranges and formulas.

By Mohammad-Ali AbidiSpeed to lead and call automation ROI7 min readAugust 12, 2026

This article was created with AI assistance.

Voice AI metrics are the operational measurements, chiefly automation rate, containment rate, and cost per call, used to determine whether an AI voice deployment resolves calls end to end, at what cost, and at what quality compared with live-agent handling.

What are the key metrics for voice AI success? Automation rate, containment rate, and cost per call are the three that matter, showing whether a voice AI system resolves calls end to end, costs less per interaction than a live agent, and scales past 50 percent containment without losing service quality or compliance control.

What is containment rate and how is it calculated for voice AI?

Containment rate is the percentage of inbound calls a voice AI system resolves completely without transferring to a human agent. It is calculated as contained calls divided by total calls, multiplied by 100, and mature enterprise deployments typically report rates between 50 and 70 percent.

Balto, whose KPI framework is widely cited in enterprise voice AI guides, defines containment as "the percentage of calls fully resolved end-to-end by voice AI without escalation to a human agent," and recommends tracking it by intent because a blended average can hide weak performance on complex call types like complaints. A dental group routing after-hours scheduling calls, for example, might see 80 percent containment on booking requests but near-zero containment on billing disputes, a gap the blended number would never surface. Tracking by intent turns one KPI into a diagnostic tool rather than a scoreboard number.

What is the difference between automation rate and containment rate?

Automation rate measures the share of all customer interactions or workflows completed by automation, while containment rate measures only whether a call ended without human escalation. The two overlap but are not identical: some vendors use them interchangeably, and others reserve automation rate for broader process work beyond phone calls.

A real estate brokerage's automation rate might include automated SMS follow-up and CRM data entry alongside phone answering, while its containment rate applies strictly to inbound calls that finish without a transfer. Because the terms are used inconsistently across vendors, teams should write down their own definition of automation rate before reporting it internally or to a board, and specify whether it includes partial automation like call routing or only full resolution. Without that definition, two deployments reporting "80 percent automation" may not be comparable at all.

How is cost per call calculated and what does it include?

Cost per call is the fully loaded cost to handle one interaction, calculated as total operating cost divided by total calls handled. For voice AI, this typically includes platform fees, model inference, voice minutes, and infrastructure costs, not just wages, which is why AI-handled calls can run under $1 while human-handled calls average $5 to $8.

CloudTalk's KPI guidance notes that voice AI cost per call should include platform fees, API costs, voice minutes, model tuning, and infrastructure, since ignoring those line items understates true cost. Balto's contact-center benchmarking splits the number further: contained calls that the AI resolves alone cost roughly $0.30 to $0.50, while escalated calls that still reach a human run $2.40 to $10, because the business pays for both the AI attempt and the live-agent resolution. For a deeper breakdown of these unit economics, see What Is the Cost per Call for Voice AI? Enterprise Economics Report. Agxntsix structures its voice AI engagements around a 60-day ROI commitment, a positioning statement about how the practice runs its engagements rather than a guaranteed dollar figure for any specific business.

What are the benchmark ranges for containment rate in enterprise voice AI?

Enterprise containment benchmarks span roughly 41 to 85 percent depending on the source and call type, with mature, well-scoped deployments generally landing between 50 and 70 percent. Narrow, highly automatable intents such as account balance lookups reach 88 to 94 percent, while complaints and escalations often fall below 12 percent.

A 2026 Deloitte Digital survey cited by Pathors found an average containment rate of 41 percent across industries, a figure well below the 60 to 80 percent range that mature, well-scoped programs report elsewhere in vendor benchmarking.

Source	Containment Benchmark	Context
CloudTalk	50%, 70%	Mature deployments, well-scoped intents
Hamming	70%, 80% (warning below 60%)	Standard customer-support flows
IrisAgent	45%, 60% average; 70%, 85% order status	Tier-1-eligible calls
Pathors (2026 Deloitte Digital survey)	41% average; 88%, 94% account balance; 5%, 12% complaints	Across industries
UIRIX	55%, 85%	Varies by use-case complexity
Twig	60%, 75% containment	Leading deployments

What are the benchmark ranges for cost per call in voice AI?

Cost per call for AI-handled interactions generally runs between $0.10 and $1 per call, compared with $5 to $8 for a fully loaded human agent call in most cited industry benchmarks. Contained AI calls that never reach a human typically cost far less than calls that escalate partway through the conversation.

A 2026 benchmark citing ContactBabel data puts an AI-handled call on a modern voice platform at roughly $0.10 to $0.30, against about $7.16 for a human phone interaction, a gap wide enough to change staffing math even at modest volume.

Source	AI-Handled Cost	Human-Handled Cost
Balto	$0.30, $0.50 (contained)	$2.40, $10 (escalated)
AdAI News (IBM-sourced)	$0.50, $1 per interaction	$5, $8 per interaction
StealthAgents (Forrester-based)	$0.05, $0.18 per minute	Not published
Hamming	$0.01, $0.25 per minute	$5, $8 per call
ContactBabel benchmark	$0.10, $0.30 per call	$7.16 per call

These ranges also explain why raw containment percentages alone cannot tell a finance team whether voice AI is working: the cost per call side of the ledger determines whether a 60 percent containment rate is a bargain or an expensive way to route calls.

How do automation rate, containment rate, and cost per call interact operationally?

These three metrics interact as a system: automation rate shows how much volume the AI takes on, containment rate shows how much of that volume it finishes without help, and cost per call shows what that performance costs per interaction. A high automation rate paired with low containment usually means the AI is routing calls, not resolving them.

Three failure patterns show up repeatedly in vendor guidance. If automation rate is high but containment is low, the system is acting as a routing layer rather than a resolution layer, and agents still absorb most of the work. If containment is high but cost per call is also high, the deployment is likely over-engineered for routine intents, running expensive infrastructure where a simpler flow would do. And if cost per call drops while containment or compliance quality falls, the business is cutting spend at the expense of service or control, a trade that surfaces later as repeat calls or complaints. Agxntsix, a member of the Claude Partner Network, builds intent-level containment and cost dashboards into its voice AI deployments so operators see all three numbers side by side, by call type, instead of one blended average.

What compliance considerations relate to voice AI metrics?

Compliance for voice AI metrics means every contained or escalated call leaves a traceable audit record, not just a completion count. Regulated workflows under the TCPA, the National Do Not Call registry, and HIPAA where healthcare communication is involved require documented consent, escalation reasons, and post-call review evidence alongside containment and cost data.

A call that the AI "contains" but mishandles, misidentifying a consent status or skipping a required disclosure, is a compliance failure even though it shows up as a win in the containment number. Enterprise guides increasingly treat the audit trail itself as a core voice AI metric, tracked alongside containment and cost, precisely because a regulator or auditor will ask for the handoff reason and the transcript, not the percentage. Businesses running outbound or healthcare-adjacent voice AI should confirm specific TCPA, DNC, and HIPAA obligations with counsel before scaling volume, since the stakes of a compliance miss are higher than the stakes of a slow rollout.

What are the common pitfalls when interpreting containment rate?

The most common pitfall is treating a high containment rate as automatic proof of success, even when customers are stuck in automation without resolution. Leading deployments report containment rates of 60 to 75 percent, but Twig's research found only 50 to 65 percent of those calls are CSAT-validated as fully resolved, a gap containment alone never reveals.

A caller who hangs up frustrated after three failed attempts to reach a human still counts as "contained" under a strict formula, which is why vendor guides recommend pairing containment with first-call resolution, handoff quality, and repeat contact rate rather than reporting it alone. A voice AI program is healthy when containment rises and the underlying issue is actually solved, not just when the call ends without a transfer.

What use cases yield the highest containment rates in voice AI?

Routine, high-volume, low-complexity call types yield the highest containment rates in voice AI, typically 65 to 94 percent depending on the intent. Account balance and status inquiries, order status checks, appointment scheduling, and password resets are the strongest performers, while complaints and escalations rarely exceed low double digits.

IrisAgent reports order status containment at 70 to 85 percent and appointment booking at 65 to 80 percent, while Pathors reports account-balance and status inquiries reaching 88 to 94 percent containment against only 5 to 12 percent for complaints and escalations. A private aviation charter desk illustrates the pattern well: automating flight-status checks and quote requests can push containment well past 70 percent, while charter negotiations and irregular-operations calls stay routed to a live team, where judgment and pricing flexibility still matter.