# Talking is not solving Voice platforms advertise containment. Containment measures whether a call ended without a human. It says nothing about whether the caller's problem went away. - Published: 2026-09-19 - Type: Perspectives - Topics: voice, agents, governance - Canonical: https://hellofig.io/blog/talking-is-not-solving **In short.** Containment is the number voice AI vendors quote and it is the wrong one. A 70 percent containment rate can sit on top of a 40 percent resolution rate, and the difference arrives as callbacks two days later. What moves resolution is not the voice model but whether the agent is grounded in the company's real knowledge, can take the action the caller wants, and knows when to hand off. Measure first-contact resolution with no callback, cost per resolved contact, and the quality of the escalation, by intent class rather than in aggregate. --- Every voice AI vendor has a number on its website and it is almost always containment: the share of calls the agent handled without transferring to a person. Reported figures run from 60 percent to above 90 percent, and the best of them are real. Containment is also the most flattering metric available, because it measures whether a call ended without a human, not whether the caller got what they needed. A call where the agent confidently gave a wrong answer is contained. A call where the caller gave up and hung up is contained. A call that ends with "I've noted that, someone will follow up" is contained, and then generates a second call two days later that counts as a fresh contact. The gap is measurable and it is wide. Published analysis of enterprise deployments this year puts contained resolution at 10 to 25 percent in year one, rising to 70 to 90 percent by year three, with containment and resolution typically 15 to 30 points apart on early deployments. A platform showing 70 percent containment can be resolving 40 percent of what it touches. A companion piece maps [where voice actually works](/blog/where-voice-actually-works). This one is about how to tell whether yours is working, which turns out to be the same question we have been asking about every other kind of agent. ## Three ways a contained call is not a resolved one **The agent answered from the wrong source.** Asked about a coverage rule, a return window or a fee, a general model will produce a fluent, plausible answer drawn from everything it has read rather than from your policy. The caller accepts it, because it sounded authoritative, and finds out later it was wrong. Containment counts this as a success twice: once when the call ends, and again when the corrective call is logged as a new contact. **The agent could talk but could not act.** This is the most common failure in practice. The caller wants the order changed, the payment applied, the appointment moved. The agent can only describe how to do those things, so the call ends with instructions. The measured outcome is containment. The actual outcome is that the customer now has homework. **The caller gave up.** Abandonment inside an automated interaction is contained by definition, since no human was involved. Deployments that do not track hang-up and repeat-contact rates alongside containment cannot tell this apart from success, and the incentive runs the wrong way.
A funnel of one hundred inbound calls. Seventy are contained with no human involved and thirty go to a person. Of the seventy contained, forty are actually resolved and thirty are not, splitting into three cases: answered fluently but wrongly from a general model, talked but did not act, and the caller gave up. All three return as callbacks logged as fresh contacts.
Illustrative, on the containment and resolution gap reported across early deployments. The thirty in the middle are the ones the dashboard counts as wins.
## What actually moves resolution The instinct when a voice agent underperforms is to change the voice model. It is almost never the voice model. **Grounding in the company's own knowledge.** The single biggest determinant of whether an answer is right is whether the agent is reading your policies, your product rules and this customer's actual record, rather than generating from a general model's priors. This is the [ontology and decision model](/blog/what-an-ontology-is-for) problem in an acoustic costume: the agent has to know what objects exist, what is true of this one, and what the rules are. **The ability to act, inside a boundary.** An agent that can look up the order, apply the credit, move the appointment and update the record resolves things. An agent with read-only access describes them. And because the actions that matter are often irreversible, the ability to act has to come with the [action boundary](/blog/getting-the-ciso-to-yes) that decides what the agent may do alone and what needs a person, enforced in the system rather than in the prompt. **Knowing when to stop.** A good agent escalates early and hands over completely, with the caller's history, what was tried and what it believes the problem is, so the person does not start from zero. A bad agent escalates late, after the caller has repeated themselves three times, which is worse than not having answered at all. Escalation is not failure. Escalation with no context is. Those three are the same three things that make a text agent work. Voice raises the stakes because a caller on a phone cannot see what the agent is doing, cannot check its work, and cannot easily go back. ## The numbers to run instead Replace containment with a small set, and read all of them by intent class rather than in aggregate, because an average across password resets and billing disputes describes nothing that exists. **First-contact resolution, no callback.** The share of contained contacts where the customer did not come back about the same thing within a defined window, usually seven or fourteen days. This is the honest version of containment and it is the number the business actually cares about. **Cost per resolved contact.** Not cost per minute, which rewards talking less, and not cost per call, which rewards ending calls. This is the voice version of the cost per accepted outcome we argued for in [why evals matter](/blog/why-evals-matter), and it makes a voice agent comparable to the alternatives, including a person. **Escalation quality.** What fraction of transfers arrive with the context attached, and how often the person has to ask the caller to repeat what they already said. A high escalation rate with high escalation quality is a functioning system. A low escalation rate with poor quality is a trap. **Abandonment and repeat-contact rate.** The honesty check on containment. If both rise while containment rises, containment is measuring people giving up. **Latency at the tail.** Vendors publish best-case turn latency. Ask for the 95th percentile, under interruption, on a real telephone connection, because that is the number your callers experience. ## Prove it before it answers a real call Everything we have argued about evaluation applies here, with two additions specific to the medium. Build the case set from real intents, drawn from actual call reasons in proportion, and grade against what a good agent in your company would do, not against a generic rubric. Then add the adversarial half deliberately: background noise, overlapping speech, strong accents, a caller reading out a policy number, someone interrupting mid-sentence to correct a digit, someone who wants a human immediately. Those conditions are where transcription accuracy collapses, and a set that omits them will tell you the agent is ready when it is not. Score the path as well as the outcome. An agent that reached the right answer by reading a record it should not have touched has failed, and an outcome-only score will call it a success. Then run it beside the current process for a cycle on live traffic, comparing resolution and callbacks rather than opinions, which is the only way to learn whether the agent is better than what you already do. ## What this means for buying Pricing shapes behaviour, so it is worth looking at what a contract is rewarding. Per-minute pricing rewards time on the line. Per-call pricing rewards ending calls. Outcome-based pricing, which is now appearing in this category, aligns the vendor with the thing you want, and it is only as good as the definition of the outcome and the record that proves it. Whatever the model, the thing to insist on is that the contract defines a resolution the same way your evaluation does, and that both sides can audit it from the same trail. A vendor that will agree to be measured on resolutions rather than conversations is telling you something about its confidence. ## How we think about it Fig's [Voice Agent](/agents/voice-agent) is built to end with a record rather than a conversation: the slot booked, the record updated, the confirmation sent, with transcripts and outcomes landing in the systems they belong to and [approval gates](/agents/agent-governance) on anything irreversible. The same evaluation harness that governs an agent's work in a channel or on the desktop governs it on the phone, because it is the same agent, and the standard is the same standard: finished work, measured against a baseline, at a cost you can defend. A caller does not want a pleasant conversation with an automated system. They want the thing done, and to not have to call again. ## Key takeaways - Containment measures whether a call ended without a human. It is compatible with wrong answers, unactioned requests, and callers giving up. - Contained resolution starts far below containment on new deployments, and the gap comes back as callbacks. - What moves resolution is grounding, the ability to act inside a boundary, and clean escalation. Not the voice model. - Measure first-contact resolution with no callback, cost per resolved contact, escalation quality, abandonment, and tail latency, by intent class. - Buy on resolutions, and make sure the contract defines one the same way your evaluations do. The test of a voice agent is not how well it talks. It is how rarely the same person has to call back.