Published containment for AI voice agents runs from about a fifth to two thirds of calls, and in regulated firms governance takes far longer than the build. It is for contact center, operations, risk and technology leaders choosing which calls to automate and how to test the case.
This paper establishes that the voice agent itself is the quick part of a regulated deployment. Published results vary by a factor of three, from about 22% to 67% of calls handled without a person, and the call mix explains most of the gap. The work that decides whether an agent stays live is governance: disclosure, identity, vulnerable customers, card data, change control and a replayable record of each call.
What results can a regulated firm expect?
The paper sets out three published deployments, with the caveat that definitions differ and the figures are not directly comparable. NS&I, a UK government savings bank, answers all inbound calls with a voice agent and about 22% are handled solely by the agent, according to its GOV.UK transparency record. PolyAI reports 42% containment for Allstate with no back-end integrations, and 67% for PG&E. The PG&E figure was 6% higher than the legacy phone system it replaced.
The paper advises planning on the low end for broad service lines. It also argues that containment is the wrong target. A call that ends without the request being finished comes back. It proposes measuring completed requests, meaning finished in the system of record in one contact with no return within seven days.
Which calls should be automated first?
The paper gives five tests: volume, structure, what the call triggers in your systems, risk and emotional load. Status and balance queries, outage updates and appointment or payment date changes are usually good first candidates. Complaints, bereavement, fraud disputes, hardship conversations and regulated advice stay with people, where the agent's job is to recognize the call and hand it over well.
Why does governance take longer than the build?
A PolyAI deployment for a large UK bank was, in the vendor's words, built in about six weeks, with governance taking six months. The paper lists ten controls, each needing an owner, a written rule and evidence. It recommends running governance alongside the build, with risk and compliance involved in week one.
It also summarizes the rules in the United States, United Kingdom, Canada and European Union, and describes a Proof of Value that analyses a firm's own call data over three to four weeks with nothing deployed.
Key takeaways
- Published containment in three live deployments runs from about 22% at NS&I to 42% at Allstate and 67% at PG&E, and the figures are not directly comparable.
- A voice agent for a large UK bank was built in about six weeks, and governance took six months, according to PolyAI's own account.
- Containment is the wrong target, because a call that ends without the request finished comes back. ETT Group measures completion within seven days instead.
- Five tests rank candidate calls: volume, structure, what the call triggers in your systems, risk and emotional load.
- Every regime the paper covers expects a firm to show what the agent said and did on a given call, so a replayable record satisfies most of them.
- The Voice Agent Proof of Value analyses a firm's own call data over three to four weeks, with nothing deployed, and ends in a go or no-go.
What is in the full paper
- The phone is still where the hard problems go
- What to expect: the honest range
- Which calls to automate first
- Build takes weeks. Governance takes months.
- The rules that apply to a voice agent
- From the call to the completed request
- Prove it on your own calls before you deploy
- ETT Group and PolyAI
The full paper has the figures, the sources and the working.
Download the PDF
