Most articles comparing AI with human support staff end the same way: AI is cheaper, works 24/7, never quits, so replace the team. It's a tidy story. It's also not what's happening in the companies actually doing this.

In a December 2025 survey, Gartner found that only 20% of customer service leaders reported AI-driven headcount reduction. The rest are using it differently — to absorb growth, cover nights and weekends, and move people onto the conversations that need judgment.

So the useful question isn't "AI or people?" It's "what is an hour of my team's time worth, and how many of those hours can an AI agent hand back?" That's a question you can answer with arithmetic.

01 · Side oneWhat a support hour really costs.

Start with the fully loaded cost of one person — salary plus benefits and payroll taxes, the number finance actually pays, not the figure on the offer letter.

Using US public data as a worked example: the Bureau of Labor Statistics puts median pay for customer service representatives at $44,770 a year, and its employer-cost survey shows benefits adding about 29.7% on top of wages. That makes a loaded cost of roughly $58,000 a year, or about $28 an hour across 2,080 working hours.

Your number will differ — a support agent in Kochi or Dubai costs something very different from one in Ohio. Use your own payroll figure. The method is what matters.

$44,770Median annual pay, US customer service reps (BLS)
29.7%Benefits and payroll taxes on top of wages (BLS ECEC)
~$28/hrResulting loaded cost per working hour

02 · Side twoHow many hours an agent hands back.

Three inputs decide this:

  1. Monthly conversation volume across every channel.
  2. Handle time per conversation. Chat and messaging benchmarks cluster around 6–7 minutes, but agents handle two or three threads at once, so the paid time per conversation is lower.
  3. The share the AI resolves end to end — no human touch at all.

The third number is where most vendor maths falls apart. Published vendor figures often claim 60–80% automation. Public case studies we reviewed for our ROI calculator land between 26% and 56%. Gartner found only 14% of customer service issues are fully resolved in self-service. A sensible starting assumption for an untuned agent is around 30%, rising as you fill knowledge gaps.

Worked example
3,000 conversations a month × 30% resolved by AI × 6.5 minutes each ÷ 2.5 handled at once ≈ 39 hours a month returned. At ~$28 an hour, that's about $1,090 of team time — roughly a fifth of one full-time person.

That is a modest number, and that's the point. At low volume, an AI agent doesn't replace anyone. It buys back a slice of every person's week. At higher volume or higher automation, the hours grow quickly — which is why the calculation is worth doing on your own data rather than trusting anyone's headline.

03 · The part spreadsheets missWhere the real savings hide.

Payroll hours are the easiest thing to count and rarely the biggest effect. Three others often matter more:

  • Out-of-hours coverage. Covering nights and weekends with people means shifts, premiums, or a second location. An AI agent answers at 3 a.m. at the same cost as 3 p.m. For businesses where enquiries arrive after work — car dealerships, real estate, retail — this is often where the value concentrates.
  • Absorbing growth without hiring. If volume doubles next year, the question isn't who you let go. It's whether you need to hire three more people or one.
  • Hiring and turnover. Support roles have high churn, and every replacement means recruiting and weeks of ramp-up. An agent's knowledge doesn't walk out the door.

There's a revenue side too: the enquiry answered in two minutes at night instead of the next morning is sometimes a sale that would otherwise have gone to a competitor. That one is harder to measure honestly, so put your own value on it rather than borrowing someone else's multiplier.

04 · The honest limitsWhat AI is still bad at.

The comparison only holds if you're clear about what an agent shouldn't do:

  • Judgment calls with money attached — goodwill refunds, exceptions, disputes.
  • Emotionally loaded conversations — a complaint about a bereavement, a frightened customer, a public escalation.
  • Anything it hasn't been taught. An AI agent is only as good as the knowledge base and tools behind it. Without them it will either refuse or, worse, improvise.

That's why the best setups don't choose between AI and people. The agent takes the routine, gathers context on the rest, and hands off cleanly — so each person spends their day on the conversations that need a person.

05 · The verdictSo which saves more?

Neither, on its own. A human team without automation pays people to answer the same five questions all day. An AI agent without humans frustrates exactly the customers you most need to keep.

The combination saves the most: AI for volume and coverage, people for judgment. Price it honestly — your volume, your costs, a conservative automation rate — and decide from there. If the numbers don't work at your volume today, they probably will when you grow.