Cost Per Resolution: The Right Way to Evaluate Customer Support AI

When a vendor tells you their customer support AI "saves 40% on support costs," they're usually cherry-picking metrics. One vendor counts only successful first-contact resolutions (easy tickets) and ignores escalations. Another vendor's 40% savings assumes you lay off 40% of your support team immediately, not accounting for the ramp time or the cost of handling complexity with human agents.

The only honest way to evaluate customer support AI is cost per resolution: the fully loaded cost of resolving one customer support request, including API, embeddings, human review, integrations, and observability. This metric is difficult to cherry-pick because it's anchored in business outcomes, not in inputs or throughput.

Why Vendors Cherry-Pick

A customer support AI vendor wants to show their product in the best possible light. So they optimize for whichever metric makes them look good:

Vendor A measures "API cost per interaction" ($0.03) and ignores the embedding cost ($0.08), human review ($0.25), and observability ($0.03). They quote $0.03 and let you do the math (poorly).

Vendor B measures "cost per auto-resolved ticket" ($0.15, excluding escalations) and ignores the fact that 30% of tickets still escalate to human agents, so their true cost is $0.15 * 0.70 + (human cost) * 0.30 = higher than they claim.

Vendor C benchmarks against your current manual process, but uses optimistic assumptions about your current support team's productivity or their new AI's success rate.

Cost per resolution cuts through this. It says: "Show me the fully loaded cost of every ticket that reached a resolution, whether auto-resolved or human-escalated." There's nowhere to hide.

How to Calculate It Correctly

To calculate cost per resolution:

1. Define what counts as resolved. A ticket is resolved when the customer's issue is addressed—they either get their question answered, their problem fixed, or they're escalated to the right team and the issue is tracked. Do not count partial resolutions or hand-offs as resolutions.

2. Measure all cost categories. Include API cost, embeddings, retries, integrations, human review time (if any), and observability. If human review is part of your process, include it at actual labor cost (not an estimated allocation). Don't leave anything off because "it's negligible."

3. Count total resolutions. Over a month, track how many tickets the AI agent handled, how many of those it auto-resolved, how many it escalated to humans, and how many escalations the humans resolved.

4. Calculate cost per total resolution. Sum all costs (AI and human) and divide by total resolved tickets. If you auto-resolved 700 tickets (at $0.32 each) and escalated 300 to humans (at $0.80 each to resolve), your blended cost per resolution is: (700 * $0.32 + 300 * $0.80) / 1000 = $0.464 per resolution.

5. Compare to your baseline. Your current support team costs X dollars per ticket resolved (including salaries, benefits, systems, training). If X is $1.50 and your AI blended cost is $0.464, you're saving $1.036 per ticket, or about 69%.

This is harder than cherry-picking metrics, but it's the only way to know if the vendor is actually saving you money.

The Escalation Rate Is Hidden

The most common place vendors hide costs is escalation rate. An AI vendor might claim their agent resolves tickets at $0.35 per resolution, but that assumes an auto-resolution rate of 85%. If the actual auto-resolution rate in your environment is 60% (because your customers ask more complex questions or have higher expectations), the true cost per resolution is higher. The vendor's $0.35 applies only to the 85% that auto-resolved; the other 60% that escalate to humans cost significantly more to handle.

A good vendor should publish their escalation rate and the cost per escalated ticket so you can calculate your own blended cost. If they won't, assume they're not confident in the number.

Apples-to-Apples Comparison

When comparing two vendors:

1. Use the same definition of "resolution" (ticket closed, customer satisfied, escalation tracked).
2. Demand the same cost breakdown (API, human review, integrations, observability).
3. Ask for the cost per resolution in your use case, not in their best case. They've benchmarked against their ideal customer; you need to know your number.
4. Ask for the escalation rate and the cost per escalated ticket.
5. Ask for quality metrics alongside cost: customer satisfaction score, CSAT, repeat-ticket rate, first-contact resolution rate.

If Vendor A quotes $0.99 per resolution (Intercom Fin's reported benchmark) and Vendor B quotes $0.47 per resolution, don't assume Vendor B is cheaper. Ask: What's your escalation rate? What's included in your cost stack? What's the human review percentage? What's your customer satisfaction score? Once you have the full picture, you can make a real comparison.

Many companies discover that a vendor with a higher per-resolution cost actually delivers better value because their auto-resolution rate is higher, their escalation cost is lower, their repeat-ticket rate is lower, or their customer satisfaction metrics are better. Cost per resolution is a prerequisite for the conversation, not the entirety of it.

The best way to think about this: a vendor's job is to minimize (cost per resolution × (repeat-ticket rate + escalation rate)) while maximizing customer satisfaction. They might quote a low cost per resolution, but if their repeat-ticket rate is 20% (meaning you have to handle the same issue twice), their true cost per unique problem solved is much higher.

A Note on First-Contact Resolution

One related metric vendors love is "first-contact resolution" (FCR)—the percentage of tickets resolved on first interaction without escalation. Higher FCR is better for customer experience, but it's not the same as cost per resolution. A vendor can have high FCR on easy tickets (high-touch but cheaper to resolve) and avoid hard tickets entirely, artificially boosting their FCR while missing the cost picture.

Even worse, a vendor might achieve high FCR by offering solutions that satisfy the customer initially but don't actually solve the underlying problem, leading to high repeat-ticket rates later. FCR looks good on the dashboard; cost per resolution exposed the truth.

The right metric is cost per resolution—whatever that resolution entails, whether it's a one-message auto-response or a multi-turn conversation ending in human handoff. That metric is honest. Add in repeat-ticket rate and you're getting at the true economic picture.

Why This Matters for Your Margin

Every 1% improvement in cost per resolution across your entire support volume flows directly to gross margin. If you handle 1 million tickets per year at an average $0.50 per resolution, a 5% improvement to $0.475 per resolution saves $25,000 per year. For a support organization with $2M/year in costs, that's 1.25% gross margin improvement.

This is why vendors obsess over their cost per resolution metrics and why you should obsess over measurement. The stakes are real.

For more detail on how to structure cost per resolution across different customer segments or channels, see the pillar article on AI cost attribution.