Deflection counts customers who did not reach a human. Resolution is supposed to count customers whose problem was solved. The two get used interchangeably in vendor reports and pricing pages, and when a vendor bills per resolution, the exact definition decides your invoice. Before you compare prices, pin down what each vendor counts.

Two definitions worth writing down

A conversation is deflected when it ends without a human agent getting involved. That includes the customer who got a correct answer, the customer who got a wrong answer and believed it, and the customer who gave up and went to a competitor's site. Deflection measures only the absence of a handoff.

A conversation is resolved when the customer's actual problem is solved. The only reliable signals are ones the customer gives you: they confirm it, the problem does not come back, or a human reviewer who reads the transcript agrees the answer was correct and complete.

Most automated resolution metrics sit in between. They use proxies such as "the customer did not ask for more help" or "the conversation was not escalated", which are closer to deflection than to resolution. That is not dishonest in itself, since there is no cheap way to confirm every outcome, but it means a reported resolution rate is only as good as the proxy behind it.

How per-resolution pricing ties your bill to success

Under per-resolution pricing, you pay a unit price each time the vendor counts a resolution. The pitch is fair on its face: you pay only when the AI does its job. In practice it has two consequences.

Your cost scales with how well the product works. If you improve your documentation and the AI starts answering twice as many questions, your bill for that line doubles. The savings are real (a human did not handle those conversations), but budgeting becomes harder, because the better the rollout goes, the larger the invoice.

The vendor defines success. Whatever proxy the vendor uses is also what bills you. If silence after an answer counts as a resolution, then a customer who closed the tab in frustration costs the same as one who got what they needed.

Two public examples show how much the definitions vary. Intercom prices its Fin AI Agent per outcome, and its pricing page (checked September 2026) says an outcome is counted when the customer confirms the issue is resolved, when they do not ask for more help after Fin responds, or when Fin completes a workflow, handoffs included. The same page says you are charged once per conversation even if several questions are answered. Zendesk bills AI agents per automated resolution, and its pricing page (checked September 2026) describes that as a customer request resolved by the AI agent without escalation to a human agent. Both definitions are defensible, but for the same traffic they will produce different counts. Check both pages yourself before you sign, because unit prices and definitions change.

How to audit what a vendor counts as a resolution

Ask these questions in writing and get the answers into the contract or order form, not a sales call.

  • What event triggers a billable resolution? A customer confirmation, a period of silence, a closed conversation, a completed workflow?
  • If silence counts, how long is the window, and what happens if the customer comes back after it closes? Is the charge reversed if they reopen the same issue the next day?
  • Do handoffs to a human ever count as a billable outcome? Some definitions include them when an automated workflow performed the handoff.
  • If the customer asks three unrelated questions in one conversation, is that one charge or three?
  • Can you export the list of billed resolutions with conversation IDs, so you can read a sample of transcripts and check them yourself?
  • Is there a monthly minimum or committed volume, and what happens to unused commitment?
  • Can you dispute a charge for a conversation where the AI's answer was wrong?

The export is the one that matters most. With it, you can pull a random sample of billed resolutions each month, label them correct or not, and compute what you actually paid per correct answer. That number, not the list price, is what you compare between vendors.

What we chose instead, and what it costs you

LayBuild does not charge per resolution. Plans are a flat monthly price with a monthly conversation allowance: Starter is ₹1,499$19 a month for 300 conversations, Pro is ₹3,999$49 for 500, and Premium is ₹9,999$125 for 1,200, or 20,000 with your own model provider key (annual billing is lower; see pricing). Our usage count is conversations created in the billing period, whatever their outcome. A conversation covers up to 15 AI replies; each further 15 replies counts as another conversation.

That model has its own trade-off. You pay the same for a conversation the agent answered well as for one it refused or got wrong. If your AI answers only a small share of questions, a flat plan can cost more per useful answer than a per-resolution plan would. The flat model makes the bill predictable and removes the incentive to count generously. It does not make the answers better. If you are on Premium and bring your own LLM provider key, token costs also arrive on that provider's bill.

Whichever model you pick, measure the same thing: cost per correct answer, from a labelled sample of your own conversations.

A metric set to track each month

If you want one internal dashboard that does not flatter the AI, track these per month:

  • Total AI conversations, and the share that ended in a human handoff.
  • The share that ended in the fixed no-answer reply. In LayBuild, when retrieval finds nothing or the answer shares too little vocabulary with the sources, the agent sends "I do not have specific information about that in the knowledge base..." instead of guessing. A rising share points to documentation gaps.
  • Correct, wrong and partial rates from a labelled random sample.
  • Repeat contacts: the same customer asking about the same issue within a few days.
  • CSAT average and the share of ratings at 4 or 5, read with the caveat that most customers never rate.

Deflection rate can stay on the dashboard. Just do not call it resolution, and do not let anyone put it in a savings calculation on its own.

Frequently asked questions

What is the difference between a deflection and a resolution?

A deflection counts conversations where a human support agent was not involved, including users who abandoned the chat out of frustration or accepted an incomplete response. A resolution means the customer's actual problem was solved, verified by a signal the customer gives you: they confirm it, the problem does not come back, or a human reviewer who reads the transcript agrees the answer was correct and complete. CSAT is worth tracking alongside it, but most customers never rate.

How do you calculate resolution rate for AI customer service agents?

Resolution rate is calculated as:

Resolution rate (%) = (confirmed resolved inquiries ÷ total AI-handled conversations) × 100.

To avoid flattering automated metrics, subtract repeat contacts (the same customer asking about the same issue within a few days) and check the result against a labelled random sample of transcripts marked correct, wrong or partial.

What happens if an AI agent doesn't meet resolution rate targets?

Start with your labelled sample and the share of fixed no-answer replies. For each failed question, check whether the documentation was missing, split badly across chunks, or filtered out by the relevance threshold, and fix the content first.

Then look at the conversations that went round in circles and make sure those customers can reach a person rather than staying stuck with the bot. For questions that come up often in many wordings, add Q&A pairs.

Next steps