By Hakim Team8 min read

Voice agent containment rates: what 40-70% actually means

Containment rate is the metric every vendor quotes and few buyers know how to interrogate. How the number gets inflated, the dialect variable most quotes don't disclose, and what to measure yourself before it drives a staffing decision.

ThesisProduct

"Containment rate" is the metric every voice-agent vendor quotes and almost no buyer knows how to interrogate. The number you'll see in 2026 sales decks is usually "40-70% containment on narrow domains": true, and also close to meaningless without knowing what "narrow domain" means, how containment is measured, and what happens to the 30-60% of calls that aren't contained. This is a practical guide to what containment rate actually measures, how vendors game it, and what questions to ask before you use it to justify a headcount decision.

What containment rate is supposed to measure

Containment rate is the percentage of calls a voice agent resolves without transferring to a human. A support line handling 10,000 calls a month at 60% containment means 6,000 calls never reached a human agent. The economic case for voice agents is built almost entirely on this number: if containment holds, you need fewer human agents for the same call volume, or the same number of agents can handle more volume as it grows.

The three ways containment gets inflated

Narrowing the domain until it's not representative. "40-70% containment" is true for a genuinely narrow domain (checking an order status, resetting a password) and false for a broad one (open-ended technical support, anything with emotional stakes like a billing dispute). Vendors quote the number from their best-case domain; ask specifically what domain the quoted number is measured against, and whether it matches your actual call mix.

Counting a transfer-avoided call as contained even when it shouldn't have been. Some measurement methodologies count a call as "contained" if the agent didn't explicitly transfer, even if the caller hung up frustrated without getting help. A call where the caller gives up isn't the same as a call the agent resolved, but both look identical in a naive "transferred vs. not transferred" count. Ask whether the vendor's containment number excludes hang-ups, and if not, what the hang-up rate is separately.

Measuring on a curated test set instead of live traffic. A containment number measured against a QA team's scripted test calls (which tend toward clean audio, clear intent, standard phrasing) will be higher than the same agent's containment on live traffic (background noise, code-switching, ambiguous requests, callers who've already tried self-service and are frustrated). Ask whether the quoted number is from production traffic or an evaluation set, and over what time window.

The dialect variable that most containment numbers don't disclose

For Arabic voice agents specifically, containment rate is highly sensitive to dialect match between the caller and the agent's STT/TTS, and this is the variable most vendor-quoted numbers don't break out. An agent tuned and measured against MSA or Egyptian Arabic callers will show a materially lower containment rate against Gulf or Levantine callers if the underlying STT mis-transcribes dialect-specific vocabulary or code-switched segments, because a misheard intent routes to the wrong response (or a fallback "I didn't understand" that triggers a transfer). If a vendor's containment number doesn't specify which dialect(s) it was measured against, ask; if you serve multiple dialects, ask for the number broken out per dialect rather than a blended average that hides the worst-performing segment.

What to measure instead of trusting the vendor's number

  1. Run your own pilot on live traffic, not a vendor demo, for at least 2-4 weeks before committing to a containment-based headcount decision. Vendor demos are, reasonably, tuned to look good; your actual call mix has edge cases a demo won't surface.
  2. Segment containment by call type and by dialect if you serve more than one. A blended 55% containment number can be masking 80% on password resets and 25% on billing disputes, very different implications for how much human staffing you can actually reduce.
  3. Track hang-up rate as a separate metric from containment. A call that ends in a hang-up should count against the agent, not toward it, and the two numbers together tell you whether low containment is a routing problem (agent knows it can't help and transfers appropriately) or a quality problem (caller gives up because the agent isn't understanding them).
  4. Measure containment on a rolling window, not a point-in-time snapshot. Containment on a newly-tuned agent in its first week is often optimistic (the easy cases got tested first); measure over a full month of real volume including whatever seasonal or time-of-day variation your call center actually sees.
  5. Ask what "resolved" means for calls the agent handled: a call the agent didn't transfer but that generated a callback three days later because the original issue wasn't actually fixed isn't a resolved call, even though it counted as contained at the time.

Where containment numbers are honest inputs, not just marketing

None of this means containment rate is a bad metric: it's the single most useful number for the actual business decision (how much human capacity do you need), which is why it gets quoted everywhere. The point is that a single blended percentage from a vendor's best-case demo is a starting hypothesis to validate, not a number to plan a staffing reduction around. The teams that get the ROI they expected from voice agents are the ones that measured their own containment on their own live traffic, segmented by the call types and dialects that actually matter to their business, before making the staffing call.

If you're evaluating Hakim for a support-line deployment and want help structuring a pilot that measures containment honestly (segmented by dialect, on live traffic, over a real time window rather than a demo), our enterprise team has run this measurement exercise with several GCC and EU customers and can walk through the methodology before you commit to anything.

Read next