OnClarity Insights
If AI costs nearly as much per contact as a person, where is the saving?
A contact-centre decision-maker at a high-volume collections and service operation asked us this while comparing an AI voice proposal with agent salaries. Here is the full answer.
6 min read
The short answer
When AI is priced per conversation or per resolution and still needs people to review, escalate and maintain it, its cost per contact moves toward a person’s. The saving was never the cheaper contact. It is the contact that stops: repeat and preventable contacts removed by fixing their cause, so nobody, human or AI, handles them again.
Why do AI and human costs per contact converge?
Most AI in customer service is now priced per conversation or per resolution. That is a fair model, and it means the AI bill grows with volume, the same way headcount does. Then add what sits around it: people who review what the AI said, escalations that still land with an agent, integration with the systems every answer depends on, and a knowledge base that has to be kept current or the AI answers from last quarter’s policy.
The mix shifts too. Once routine questions go to AI, the contacts left for people are the hard ones: the dispute, the exception, the customer who is already upset. The average human contact gets more expensive even as there are fewer of them.
Analysts expect the gap to keep narrowing. Gartner predicts that by 2030 the cost per resolution for generative AI will be higher than many B2C offshore human agents, citing rising data centre costs, AI vendors moving from subsidised growth to profit, and more complex use cases (Gartner, 26 January 2026).
None of this makes AI the wrong call. It does lower the cost of many simple contacts, and customers who want a fast answer get one. It means the saving was never going to come from the price gap alone.
What does a cost-per-contact business case get wrong?
It treats volume as fixed. The usual case multiplies the contacts you automate by the difference in cost per contact, as if next year’s contacts were a fact of nature.
They are not. A large share is the same problem coming back. SQM Group’s benchmarking research finds that 41% of customers had to contact a company more than once to resolve the same question or problem. That is a share of customers, not of contacts, but the direction is clear: much of what you pay to handle is work you already paid for once.
Automating a repeat problem only makes it cheaper to have it again. Bain put the general version plainly in June 2026: “AI doesn’t fix workflow debt; it locks it in.”
Where does the saving hide?
In three places, and none of them is the price of a contact.
Repeat contacts on the same issue. The same customer, back about the same thing, in any channel. A common example: a ticket marked resolved while a follow-up is still promised, and the customer calling back to check. That is a second contact for one problem. This is repeat cost.
First contacts caused by something you can fix at the source. A bill that confuses, a policy that contradicts the help article, a delivery update that never goes out. Fix the cause and the first contact disappears for every customer after, whoever would have answered it.
What a wrong or partial answer sets off next. The escalation, the credit, the complaint. These rarely sit on the AI line of the budget. They are paid for anyway.
One calculation you can redo with your own numbers
Annual repeat cost = monthly contacts x repeat share x cost per contact x 12
Illustrative inputs: 100,000 contacts a month; 30% of them repeats (an assumption, so use your own); $7 per contact, rounded from ContactBabel’s 2026 US average of $7.20 per inbound call.
100,000 x 0.30 x $7 x 12 = $2,520,000 a year spent on repeat contacts.
Now compare two moves on that one line.
A. Cheaper contacts. Automate the repeats at a 30% lower cost per contact. Saving: 360,000 x $7 x 0.30 = $756,000 a year. All 360,000 contacts still happen.
B. Fewer contacts. Remove the causes of a quarter of the repeats. Saving: 90,000 x $7 = $630,000 in year one. Then fix a quarter of what remains each year.
Illustrative. 100,000 contacts a month, 30% repeats, $7 per contact. Flat volume; fixes hold.
Assumptions: flat volume; fixes that hold; B removes 25% of the remaining repeats each year (90,000, then 67,500, then 50,625 contacts); A’s 30% gap holds for all three years.
In year one, automation is ahead. That is the honest result. But A is a one-time step: it holds only while the price gap holds, and it shrinks if the cost per resolution rises. At a 10% gap, A is worth $252,000 a year. B compounds, because a fixed problem stays fixed and the next fix adds to it. A contact that stops saves its full cost. A cheaper contact saves only the difference.
You do not have to choose. Fix first, then automate what remains, and the AI handles fewer contacts that are not the same problem coming back.
What should you ask instead?
Stop asking what AI costs per contact handled. Ask what it costs per issue fixed: everything you spend on an issue, every contact, every channel, human and AI, divided by the issues that did not come back within your window. Ask your vendor. Ask your own team. If nobody can answer, that is where the saving is hiding.
That is the executive version of this whole answer: stop paying for the same problem twice.
The honest trade-off
Fixing at the source is slower than buying automation. The owners usually sit outside service, in product, billing, policy and operations, and they have their own roadmaps, so expect quarters, not weeks. Some contacts should simply be automated: a password reset or an order status check is better served by a fast answer than by a project. And the saving lands in other teams’ budgets. Service rarely gets the credit unless finance signs a baseline before the fix ships and measures the same window after it. Without that baseline, a fix looks like a seasonal dip.
What to do this week
Pull 90 days of contacts from every channel, with a customer ID you can match across them.
Count repeats: the same customer on the same issue within 7 days.
Price them with finance’s cost per contact, by channel.
Name the top three causes and an owner for each, outside service if that is where the cause sits.
Ask your AI vendor, and your own team, for cost per issue fixed.
Go deeper
What is cost per contact? The definition, what it includes and how to work out your own figure, the one input this calculation depends on.
Questions people ask
AI chatbot vs human agent cost: which is cheaper?
Per contact, AI is usually cheaper on simple, high-volume questions. The gap narrows once you add per-resolution fees, human review, escalations, integration and knowledge upkeep, and once the contacts left for people are the hardest ones. Compare cost per issue fixed, not per contact handled: a contact that stops costs nothing.
How do you reduce cost per resolution?
Need fewer resolutions. Find the issues customers contact you about more than once, fix their cause at the source (the article, the policy, the product or the process), then check that the contacts on that issue stopped. Automating what remains lowers the unit cost; fixing the cause removes the contact altogether.
What cost reduction percentage should you expect from AI customer service?
It depends on how much of your volume is repeat or preventable. Automation alone gives a one-time step down in unit cost, and that step shrinks if per-resolution costs rise. Removing repeat and preventable contacts adds up each year. Build the case on your own repeat cost, against a baseline finance signs.
Sources
Gartner, “Gartner Predicts GenAI Cost Per Resolution for Customer Service Will Exceed Offshore Human Agent Costs by 2030”, press release, 26 January 2026.
SQM Group, “Call Center Customer Experience Research”.
ContactBabel, “The 2026 US Contact Center Decision-Makers’ Guide”.
Bain & Company, “Your AI Budget Is Growing. Your Returns Aren’t. Here’s Why.”, 1 June 2026.
