In PolyAI case studies and NS&I's GOV.UK record, voice agents in regulated service lines contain 21% to 42% of calls. That is a good result. What matters more is what happens to the rest, and to the callers who ring back.
Most voice agent business cases lead with containment: the share of calls that end without reaching a person. It's easy to measure, it rises when things are working, and it's the wrong number to run the service on.
We put it this way on our own voice service page: "A contained call that produces a repeat contact tomorrow has resolved nothing."
What containment counts
A call is contained when it ends without a person. That includes the caller who got exactly what they needed. It also includes the caller who gave up, hung up and rang back an hour later, or tried the website, or wrote in to complain. A dashboard can't tell those apart unless you build it to.
So a rising containment rate can mean the agent is getting better. It can also mean the route to a person has got harder to find. Anyone who has run an old-style IVR has seen this: the menu "contains" the call, and the caller rings back and presses zero.
Callers notice. Gartner found 87% of customers say it is "essential for companies to provide an option to reach a human agent when using genAI" (August 2026). Make that option hard to find and containment goes up while trust goes down.
There is a quieter version too. An agent that answers confidently but wrongly contains the call perfectly. The caller hangs up happy, acts on the wrong answer, and the cost turns up weeks later as a complaint.
What good actually looks like
Headline numbers on vendor sites tend to come from narrow, structured calls: reservations, simple routing, outage updates. In broad regulated service lines, the published figures are lower:
- Quicken (US personal finance): 21% call containment (PolyAI case study, July 2025)
- NS&I (UK government-backed savings): "approximately 22.23% of calls are contained" (GOV.UK record, October 2025, the one independent figure here)
- Simplyhealth (UK health insurer): 25% of calls resolved (PolyAI case study, September 2023)
- UniCredit Zagrebačka banka (Croatia): 27% call automation (PolyAI case study, May 2026)
- Howard Brown Health (Chicago health centre): 30% call containment against a 20% target (PolyAI case study, April 2025)
- Allstate (US insurer): 42% containment with no back-end API integrations (PolyAI blog, July 2026)
At the structured end, a PolyAI case study reports 67% containment for PG&E across its utility lines, which include outage and billing calls that follow a clear pattern. PolyAI adds that this is 6% higher than PG&E's legacy IVR, which is the baseline that matters (PolyAI case study, May 2026).
Two things stand out. First, the 21% to 42% in those case studies and records is a good result in these sectors. NS&I's GOV.UK record says 100% of its inbound customer calls are now answered by its agent, which is also available around the clock. The same record says the agent contains about a fifth of calls and passes others to people by intent, sentiment or topic. At the more than 100,000 customer interactions a month that the record reports, that is a lot of routine calls people no longer have to take.
Second, look at the labels. Containment, calls resolved, call automation. Even within one vendor's case studies, the measure changes from customer to customer. Before you compare any two numbers, ask how each one was defined.
It grows, if you tune it
Quicken's containment went "from 5% at launch to 21% a year later", according to PolyAI. PG&E set a return-on-investment target of 3.5% containment and, PolyAI reports, contained 15% on its first day. A first-month figure tells you where tuning starts. It says very little about where it ends.
That tuning only helps if you're tuning towards the right thing. Tune for containment alone and you'll get containment, whether or not callers got what they needed.
What to measure instead
Keep containment on the dashboard. Just don't put it at the top. This is the scorecard I'd run a voice agent on, in this order.
- Resolution. The share of calls where the caller's need was met end to end, including the work after the call. A balance read out is resolved. A promised callback that never happens isn't.
- Repeat contact within 7 days. Same caller, same reason, any channel. This is the check on every other number. If containment rises and repeat contact rises with it, you've moved the work rather than removed it.
- Handover reasons. Why each call went to a person: the caller asked, vulnerability, low confidence, out of scope, a system failure. Each reason has a different fix. Check the handover too. Did the context arrive, or did the caller repeat their security checks?
- Satisfaction on agent-handled calls. Measured separately from the human queue, so a good agent can't hide a struggling team, or the reverse.
- Accuracy against policy. Sample calls and check what the agent said against your approved documents, claim by claim. An agent can sound confident and be wrong, and in a regulated firm a confident wrong answer is a conduct issue.
- Cost per resolved call. Platform cost, people and after-call work, divided by the calls actually resolved. PolyAI's published pricing charges for ongoing use by the minute, but your finance team will want the cost per outcome. PolyAI reports one unnamed EU bank putting it at "£2 per resolved call for the AI vs £4.48 for their agents" (PolyAI blog, July 2026). For context, ContactBabel puts the average UK inbound call at £6.17 (2026).
Put the six on one page and review them monthly with operations, compliance and finance in the room. Each of them owns one of these numbers, and each will spot a different problem.
Set the target from your own baseline
None of these numbers means much without a baseline. Measure the call type before the agent goes live: resolution, repeat contact, satisfaction, cost per call. Then set targets against your own figures. A vendor's best customer is a useful reference, but it isn't your starting point.
If you are writing a voice agent business case for 2027, try one change. Replace the containment target with a resolution target and a ceiling on repeat contact, and see what it does to the numbers.
That is also how ETT scores a voice agent Proof of Value: one call type, measured against your own baseline. There is more on how we design and measure voice agents on our conversational AI and voice automation page.
Sources
- ETT conversational AI and voice automation service page
- PolyAI case studies: Quicken (July 2025), Simplyhealth (September 2023), UniCredit Zagrebačka banka (May 2026), Howard Brown Health (April 2025), PG&E (May 2026)
- PolyAI blog, "Voice AI compliance in financial services", 28 July 2026
- GOV.UK Algorithmic Transparency Record, NS&I (30 October 2025)
- Gartner, 4 August 2026
- ContactBabel, UK Contact Centre Decision-Makers' Guide 2026
- PolyAI pricing page, poly.ai/pricing (October 2026)