AI vs human customer service cost

The vendor comparison usually shows you the price of a platform against the price of a seat and stops there. That is the least interesting part of the arithmetic. Below is the full cost, including the two line items that decide most of these deals and almost never appear in the pitch.

The comparison nobody shows you

What AI does not remove

• Supervision. Someone still reviews escalations, tunes intent coverage and handles the tickets the model gets wrong.
• The residual queue. At 50% deflection you still need humans for the other half — the half that is hardest.
• Repeat contacts. A confident wrong answer generates a second ticket. One badly automated answer can cost two human touches.
• Knowledge debt. Automation only knows what is written down. If your FAQs live in three heads, it automates nothing.

What it genuinely does

• Kills the queue, not the work. Median first-response time collapses even when headcount is unchanged.
• Absorbs the 2am ticket. Out-of-hours coverage you were paying premium for is nearly free.
• Handles the boring 60%. Order status, password resets, refund status — the tickets that consume senior agents.
• Scaling is linear in cost. 10x volume does not mean 10x headcount if coverage caps out.

Cost per ticket, by volume

E-commerce, 9.5 min average handle time, $26/hour fully loaded, 50% automation coverage at $0.65 per resolution. “AI cost per ticket” is the platform fee spread over the tickets it actually resolved — the number that matters, not the list price.

Tickets / mo	Agents	Human / ticket	AI / ticket	AI ÷ Human	Absolute $ / mo	Verdict
500	1-1	$8.29	$0.65	0.08x	-$162	Do not buy
1,000	1-1	$4.14	$0.65	0.16x	-$325	Do not buy
2,000	1-2	$4.14	$0.65	0.16x	+$3,493	Automation pays for itself
5,000	3-5	$2.49	$0.65	0.26x	+$2,518	Automation pays for itself
10,000	5-9	$2.49	$0.65	0.26x	+$9,179	Automation pays for itself
25,000	13-22	$2.32	$0.65	0.28x	+$20,876	Automation pays for itself
50,000	25-43	$2.32	$0.65	0.28x	+$41,751	Automation pays for itself
100,000	49-85	$2.32	$0.65	0.28x	+$83,502	Automation pays for itself

The result that surprises people. Read the two middle columns together and the ratio gets worse as you scale — from 0.08x to 0.28x. AI does not become relatively cheaper with volume. It becomes relatively less advantageous.

The reason is the human column. Cost per human-handled ticket falls as volume rises — $8.29 at 500 tickets/month, because at that size your one agent is idle a lot of the time, down to $2.32 at 100,000, where agents are fully utilised. AI priced per resolution stays flat at $0.65. Scale improves your human efficiency faster than it improves your automated one.

So decide on absolute dollars, never on the ratio. The ratio narrows while the absolute number swings from -$162/month to $83,502/month. A buyer who negotiates on “cost per ticket is 4x cheaper!” is optimising the one column that gets worse for them, and will be disappointed by month three.

The cliff at the low end is real though, and it is a cliff rather than a curve: below roughly 1,500–2,000 tickets a month there is no headcount to remove, so a per-resolution fee buys you nothing at all.

The four numbers to get right before you sign anything

1. Your real average handle time, measured over two weeks. Not estimated, not from a survey. At 7 vs 12 minutes, your agent clears 936.146 vs 537.786 tickets a month. That is a 8 vs $ 4 difference in cost per ticket before you buy anything.
2. Your deflection ceiling, not your ambition. 48%–62% depending on mix. If a pilot reports 80%, ask what was excluded.
3. Whether the saving is actually cashed. Headcount is inelastic. Savings accrue as people leave, over 12–18 months, not the week you switch it on. Model the ramp.
4. What happens at 30% lower volume. Every table on this page assumes volume is flat. If it is not, your payback months are the number that matters and it is the one every vendor skips.

Run your own numbers on the calculator — it will show you the same cliff, or tell you not to buy at all.