Every AI support vendor pitches the same number: deflection rate. "Our AI deflects 60% of tickets." "We reduced inbound volume by 45%." The slide deck has a chart going up and to the right. The CFO nods.

Take a step back and think from first principles, why does nobody in the room asks what happened to the customer?

What deflection actually measures

A ticket is "deflected" when a customer interacts with the AI and does not subsequently open a ticket with a human agent. That is the entire definition. The customer asked a question, the AI responded, and the customer went away.

There are three reasons a customer goes away after an AI interaction. They got the right answer and their problem is solved. They got a wrong answer and do not realize it yet. Or they got a useless answer and gave up. Deflection counts all three the same.

A customer who asks "can I get a refund?" and receives an incorrect "you are not eligible" will not open a follow-up ticket. They will accept the answer, feel frustrated, and eventually churn. That interaction shows up as a successful deflection. The metric improved. The customer outcome got worse.

This is not a theoretical concern. When we audit RAG-based support deployments, we consistently find that 15-25% of "deflected" tickets were deflected with incorrect or incomplete answers. The customer left, but their problem was not solved.

Why vendors love this metric

Deflection is easy to measure and easy to inflate. You do not need to verify whether the AI's answer was correct. You do not need to track whether the customer's problem was actually resolved. You just count: did the customer come back? If not, success.

This creates a perverse incentive. An AI that gives confident, authoritative wrong answers will have a higher deflection rate than an AI that honestly says "I'm not sure, let me connect you with a human agent." The first AI looks better on paper. The second AI produces better customer outcomes.

Vendors also conflate deflection with resolution in their marketing. "60% of tickets resolved by AI" often means "60% of tickets where the AI responded and the customer did not follow up." Those are different statements with very different implications for customer experience.

What resolution actually requires

Resolution means the customer's problem is fixed. A refund was processed. An account was updated. A billing error was corrected. An order status was confirmed against live data. The customer can verify that the outcome is real.

Measuring resolution is harder than measuring deflection. You need to confirm the action was taken, check that the data is correct, and ideally validate that the customer is satisfied. This means connecting your AI to your backend systems so it can take real actions, not just generate text about what it would hypothetically do.

This is exactly why deflection became the industry standard. It is the metric you can report without building the infrastructure to verify outcomes.

The cost of optimizing for the wrong metric

Teams that optimize for deflection make predictable decisions. They deploy the AI on easy, high-volume queries (password resets, business hours, shipping timelines) where the deflection rate will be high. They avoid deploying it on complex queries (refund eligibility, billing disputes, account changes) where the AI might escalate to a human and hurt the deflection number.

The result is an AI that handles the tickets your help center could already answer and avoids the tickets that actually cost your team time. Your deflection rate looks great. Your support costs barely move. The hard tickets, the ones that take 15 minutes per interaction and require three system lookups, still go to humans every time.

This is why so many support teams plateau at 30-40% automation despite deploying AI. The AI is optimized to deflect easy tickets, and easy tickets were never the bottleneck.

How we think about it

At Fini, we charge per resolution. A resolution means the customer's problem was actually fixed: a refund processed, an account updated, a query answered with verified data from a live system. If the AI escalates to a human, we do not charge for it. If the AI gives an answer but the customer comes back with the same problem, that is not a resolution.

This pricing model only works if the AI can actually resolve tickets, which requires the ability to take actions against real systems, not just generate responses from documents. It also forces us to care about accuracy in a way that deflection-optimized vendors do not have to. A wrong answer that deflects a ticket would cost us a customer. A correct answer that resolves a ticket earns revenue.

Our production deployments resolve 70-85% of tickets end-to-end. That is a lower number than the 90%+ deflection rates some vendors report, and it represents dramatically more value. Every one of those resolutions is a ticket where the customer's problem is confirmed fixed, not a ticket where the customer stopped asking.

The metric you should ask for

When evaluating an AI support vendor, ask for resolution rate, not deflection rate. Specifically:

Confirmed resolution rate. What percentage of AI-handled interactions resulted in a verified outcome (action taken, data confirmed, customer satisfied)? If the vendor cannot answer this, they are measuring deflection and calling it resolution.

Accuracy on policy-dependent queries. What percentage of answers involving business rules, eligibility checks, or calculations were factually correct? This is where RAG-based deflection engines break down. Informational accuracy is easy. Policy accuracy is where the architecture matters.

Escalation rate with context. When the AI escalates, does the human agent receive full conversation context and a summary of what the AI already checked? A clean escalation is a feature. A blind handoff where the customer repeats everything is a failure.

Re-contact rate. Of the tickets the AI handled, what percentage of customers came back within 48 hours with the same issue? This is the simplest way to catch deflection being reported as resolution. If 20% of "resolved" tickets generate a follow-up, the real resolution rate is 20% lower than the number on the dashboard.

The shift

The AI support industry spent 2024 and 2025 selling deflection. Buyers are starting to notice that their deflection rates went up and their customer satisfaction did not follow. The next generation of buyers will demand resolution metrics: confirmed outcomes, verified accuracy, real actions taken.

We built Fini for that buyer. Resolution-based architecture, resolution-based accuracy, resolution-based pricing. The metric we optimize for is the one that matters to your customer.