AI Call Centre Statistics: The Numbers, and Where They Came From

Research date: 23 September 2026. Last updated: 23 September 2026.

The single most reliable statistic in this category is not from a vendor. The US Bureau of Labor Statistics counts 2,666,000 customer service representatives in 2025 and projects that number to fall 5%, or 141,800 jobs, by 2035 — while still expecting 289,500 openings a year, every one of them from replacing people who leave rather than from growth. BLS attributes the decline to automation directly.

That one data series says more than most of what is published about this market, because it is measured rather than asserted. This page collects the statistics that have a traceable source, shows the arithmetic where a figure is derived, and is explicit about what nobody publishes.

Brilo AI publishes this page and sells AI call handling. Brilo's own published numbers appear in their own section, sourced the same way as everyone else's, and what Brilo does not publish is stated.

Why this page is shorter than the others you will find

Most "call centre statistics" pages are long. They are long because they repeat figures that cannot be traced to anyone who measured anything — a percentage appears in one blog post, gets cited by the next, and after four hops it is common knowledge with no origin.

Every number below names its source and the date it was read. Where a figure is calculated rather than published it is marked derived and the arithmetic is shown, so you can disagree with the assumption instead of the conclusion. Where the honest answer is that nobody publishes a number, the section says so rather than filling the space.

That rules out a great deal of what is usually included. What remains is shorter and it is checkable.

The employment picture

This is the strongest data on the page, because it comes from a statistical agency rather than a company selling something.

Source: US Bureau of Labor Statistics, Occupational Outlook Handbook, Customer Service Representatives, read 23 September 2026.

Two things in that table are routinely misread.

The decline is real and BLS says why. In its own words: "There is expected to be less demand for customer service representatives, especially in retail trade, as their tasks continue to be automated," and "Self-service systems, social media, and mobile applications enable customers to do simple tasks without interacting with a representative." It adds that "Advancements in technology will gradually allow these automated systems to do even more tasks." This is not a vendor's claim about its own product. It is the federal projection.

But 289,500 openings a year is not a growing industry. BLS is explicit that "All of those openings are expected to result from the need to replace workers who transfer to other occupations or exit the labor force." The occupation is shrinking and simultaneously hiring heavily, because turnover is high. Anyone citing the openings figure as evidence the sector is booming has misread it, and anyone citing the 5% decline as collapse has misread it in the other direction.

BLS also notes an offsetting factor worth keeping: "Some companies will continue to use in-house service centers to differentiate themselves from competitors, particularly for complex inquiries." The complex call is the one that stays.

What a human call minute costs — derived

No agency publishes a fully loaded cost per call minute, so this is calculated. The arithmetic:

BLS puts the median customer service representative wage at $21.53 an hour (May 2025). Wages are not the employer's cost. The BLS Employer Costs for Employee Compensation series puts civilian worker compensation at $49.46 per hour worked in June 2026, of which $33.85 is wages and salaries and $15.61 is benefits — so wages are about 68.4% of what an employer actually pays. Sources: BLS Occupational Outlook Handbook and BLS Employer Costs for Employee Compensation (released 9 September 2026), both read 23 September 2026.

Derived: $21.53 ÷ 0.684 = roughly $31.50 per hour, or about $0.52 per minute of paid agent time.

The assumption, stated so you can reject it: this applies the civilian-average benefits share to one occupation's median wage. If customer service roles carry a lower benefits load than the civilian average — plausible, given the mix of part-time and outsourced work — the real figure is lower. Treat $0.52 as an order of magnitude, not a precise number.

And one correction most comparisons skip. That is a minute of paid time, not a minute of talking. Agents are not on calls continuously, so a minute of actual conversation costs more: divide by your occupancy rate. At 70% occupancy, $0.52 becomes roughly $0.75 per talk minute. Use your own occupancy figure — a generic one would be exactly the kind of untraceable statistic this page exists to avoid.

What AI call handling costs

All figures read from each vendor's own pricing page on 22 September 2026.

Sources: aws.amazon.com/connect/pricing, retellai.com/pricing, zendesk.com/pricing, genesys.com/pricing, talkdesk.com/pricing, five9.com/pricing, goodcall.com/pricing, brilo.ai/pricing, synthflow.ai/pricing — all read 22 September 2026.

The comparison the two sections above allow

A human talk minute costs roughly $0.52–$0.75 on the derivation above. Published AI rates run $0.038 to $0.31 per minute.

That gap is the entire commercial case for this category, and it is also where most published analysis stops. Three things keep it honest:

The AI minute is not the whole cost. Integration build, overage, concurrency and telephony sit on top, and none of them appear in a per-minute rate.

Only resolved calls count. A call the AI cannot finish costs you the AI minute and the agent minute. At 50% containment the effective saving is far smaller than the rate difference implies, which is why cost per resolved call is the only figure that compares the two honestly.

Nobody publishes their containment rate. Not one vendor in the table above publishes what proportion of calls its system actually resolves without a human. That is the single number a buyer most needs and the one the category does not supply.

What free tiers actually contain

Sources: rasa.com/pricing, brilo.ai/pricing, vapi.ai/pricing, retellai.com/pricing, read 22 September 2026.

Converted to minutes at each vendor's own published rates, $5 on Vapi buys roughly 40–60 minutes and $10 on Retell buys 32–140 minutes, the range depending entirely on configuration. Credits are quoted in currency, which tells you nothing until you convert them.

The performance numbers that are published

Latency is the one operational metric vendors do publish, because it flatters them.

• ~75ms — ElevenLabs Flash v2.5 speech synthesis

• ~280ms — ElevenLabs v3 Conversational, the more expressive model

• Sub-600ms — Vapi's advertised whole round trip

• 10 milliseconds — Deepgram's default endpointing threshold

Sources: elevenlabs.io/docs/models, docs.vapi.ai, developers.deepgram.com/docs/endpointing, read 23 September 2026.

Both ElevenLabs figures are marked "excluding application & network latency", which is to say they are laboratory numbers. Treat every published latency figure as a floor.

What nobody publishes

A statistics page is more useful for naming the gaps than for filling them. As of 23 September 2026, across the vendors above:

No containment or resolution rate. Not one publishes the proportion of calls its system resolves unaided.

No accuracy or error rate. Nothing on how often the system gets an answer wrong, which is the risk that matters most in regulated work.

No speech recognition accuracy by accent or in noise. Language lists say what is supported, never how well.

No escalation quality measure. Whether the human receives context or starts from nothing is not a published metric anywhere.

Three vendors publish no price at all — Dialpad, PolyAI and Cresta, verified 22 September 2026.

If you see any of those numbers quoted as an industry average, ask who measured it and on whose calls. In this category the answer is usually nobody.

How to read a statistic from this market

Ask who benefits. A containment rate published by the vendor whose containment it describes is marketing until a methodology is attached.

Follow it to a primary source. If the trail ends at another blog post, it is not a statistic.

Check the date. Pricing in this category moved sharply in 2026 — Synthflow's entry point is now $30,000 a year. A figure from eighteen months ago may describe a product that no longer exists at that price.

Prefer measured to projected. BLS counting 2,666,000 jobs is a different kind of claim from anyone's forecast of how many interactions AI will handle by some future year.

Brilo's published numbers

Brilo publishes a free tier at 10 minutes a month with one agent and one workspace, Pro at $149 a month for 600 minutes with overage at $0.16, and Growth at $499 a month for 2,500 minutes at $0.14, with a 10% annual discount. Source: brilo.ai/pricing, read 22 September 2026.

What Brilo does not publish, held to the same standard as the section above: no containment rate, no accuracy rate, and no recognition accuracy by accent. On the metrics that matter most to a buyer, Brilo discloses no more than the vendors it is listed beside. That is worth stating plainly on a page about the reliability of this market's statistics.