Measuring the ROI of customer support AI.

By COGScontrol Team · July 8, 2026

To measure customer support AI ROI, divide the fully loaded cost of the support AI by genuine resolutions — contacts closed without a human and without the customer returning — and compare that cost per resolution to a human-resolved contact at equal quality. The denominator is resolutions, never contacts or deflections, because deflection without resolution moves a cost rather than removing it.

Support is one of the most heavily automated functions in the enterprise — McKinsey found service-operations optimization was the most commonly adopted AI use case four years running — which makes it the function where ROI is most often overstated. The headline results are genuinely large: Klarna reported its AI assistant handled two-thirds of its customer service chats in the first month, doing the equivalent work of 700 full-time agents with CSAT on par with humans and resolution times falling from 11 minutes to under 2. Those are real numbers. They are also exactly the kind of numbers that hide a measurement question.

Why is deflection not resolution?

The fastest way to inflate support-AI ROI is to count deflections as if they were resolutions. A deflection means a contact didn't reach an agent. A resolution means the customer's problem was actually solved. The gap between them is where the cost hides: a bot that deflects a contact the customer immediately re-opens has created a re-contact; one that deflects a contact the customer abandons in frustration has created a churn risk and a CSAT hit. Both score a deflection. Neither is a saving.

Klarna's own arc is the most honest illustration available. After the celebrated rollout, the company moved back toward keeping human agents available for complex cases. That is not a story about AI failing — it is a story about what you can only see if you measure resolution quality alongside cost. Without both numbers, you cannot tell the contacts where the bot pays back from the contacts where it quietly costs more. With both, you route accordingly.

What is the cost-per-resolution formula — and which two qualifiers matter?

Two qualifiers carry the whole result. Fully loaded: the AI cost is not just the per-conversation model fee but the retrieval/vector store, the orchestration, the failed and escalated attempts, and the human review of those escalations — the model bill alone understates the truth, the same way ServiceNow's CFO could only state AI was under 10% of cost to serve because the company measured everything around it. Genuine resolutions: closed contacts that did not re-contact, measured at matched CSAT. Strip either qualifier and the cost per resolution flatters itself. The underlying unit-cost mechanics are covered in AI unit economics.

How do you map support metrics to the right denominator?

What does cost per resolution look like in practice?

The numbers below are illustrative — a hypothetical company, not customer data. Call it Solstice: a consumer subscription business handling 100,000 support contacts a month, with a support AI whose fully loaded cost — tokens, retrieval, orchestration, failed and escalated attempts, and the human review of those escalations — is $30,000 a month.

The helpdesk dashboard says the bot handled 60,000 contacts: a 60% deflection rate. Measured honestly, those 60,000 split three ways: 45,000 genuine resolutions (closed, no re-contact within seven days, CSAT matched to human-handled contacts), 9,000 re-contacts that landed back in the human queue, and 6,000 abandons — customers who gave up.

• Cost per resolution. $30,000 ÷ 45,000 genuine resolutions = $0.67. The deflection version — $30,000 ÷ 60,000 = $0.50 — flatters the number by a quarter.
• ROI. At a loaded human cost of $9.00 per resolution, the AI displaced $405,000 of agent work: ($405,000 − $30,000) ÷ $30,000 = 12.5×.
• The cost that moved. The deflection headline would have claimed 60,000 resolutions — $540,000 displaced. The $135,000 gap is not savings: 9,000 contacts were handled twice, and 6,000 customers are a churn line CSAT will surface next quarter.

Notice what honest measurement did to a genuinely good result: it survived. The defensible ROI is still 12.5× — smaller than the deflection version, but a number finance can take to the renewal — and the re-contact split now shows exactly which contact types to route back to humans.

Why won't a support dashboard or a cost tool tell you this?

Your helpdesk reports deflection and CSAT but not the fully loaded AI cost behind each resolution; your cloud-cost or FinOps tool reports the AI spend but has no concept of a resolution to divide it by. Neither can produce a cost per resolution on its own, and stitching them together by hand survives until the second channel or the second model ships — the distinction between spend tooling and value tooling is the subject of FinOps vs AI Value Management.

COGScontrol joins the two sides. It attributes the fully loaded cost of the support AI — tokens, retrieval, orchestration, escalation handling — into one ledger, divides it by genuine resolutions imported from your helpdesk, holds CSAT and re-contact rate in the same view, and reconciles cost to provider invoices every 24 hours. It is the operating loop from how to measure AI ROI and internal AI ROI, pointed at support. When someone claims the bot saved a fortune, you can confirm or correct it with a cost per resolution and a quality trend rather than a deflection headline. See the features or start free; pricing is a fixed subscription, never a percentage of your AI spend.