Blue line drawing of a balance scale weighing a single coin against a stack of trays on a grid background

Support

AI customer service: deflection rate vs. resolution rate

Aug 14, 2026 · 7 min read

Updated Oct 31, 2026

UNISTAFF.AIUNISTAFF.AIUNISTAFF.AIUNISTAFF.AIUNISTAFF.AIUNISTAFF.AIUNISTAFF.AIUNISTAFF.AI

Short answer

What is the difference between deflection rate and resolution rate?

Deflection rate is the share of customer contacts closed without a human agent. Resolution rate is the share where the problem was actually fixed, so the customer did not have to come back. A conversation can be deflected and unresolved at the same time, which is why deflection alone can rise while service gets worse.

Deflection is easy to count and easy to inflate: a customer who gives up in frustration counts the same as one who got the answer. Resolution is harder to count and closer to what the customer wanted. Here are the formulas, the gap between them, and the numbers worth putting on a dashboard.

The formulas, side by side

MetricFormulaWhat it misses
Deflection rateContacts closed without a human agent ÷ all contactsWhether the customer got what they needed
Containment rateConversations that ended inside the bot ÷ conversations the bot startedCustomers who left and came back by phone or email
Automated resolution rateAI-handled contacts with no repeat contact on the same issue within 7 days ÷ AI-handled contactsProblems the customer stopped reporting without a fix
First-contact resolutionIssues fixed on the first contact, AI or human ÷ all issuesEffort: how hard the customer had to work
Repeat-contact rateCustomers contacting again about the same issue within 7 days ÷ all customers who contactedNothing on its own; it is the check on the other four
Five support metrics. Vendors define them differently, so ask any vendor for the exact formula behind a quoted rate.

Containment and deflection are often used as synonyms. The useful distinction: containment is measured inside one channel (did the conversation stay with the bot), deflection across the whole support operation (did the contact avoid a person anywhere).

The same AI conversations counted two waysTop row, counted as deflected: answered and fixed; customer gave up; wrong answer and ticket closed. Bottom row, counted as resolved: answered and fixed; handed over and fixed on the first human touch; no repeat contact within seven days.Counted as deflectedAnswered,fixedCustomergave upWrong answer,closedCounted as resolvedAnswered,fixedHanded over,fixed on firsthuman touchNo repeatcontact in7 daysSame conversations, two counts.
The same conversations counted two ways. Deflection counts everything the bot closed; resolution counts only what was fixed.

Why deflection alone misleads

  • It rises when a person becomes harder to reach: a hidden “talk to a human” option lifts deflection and lowers satisfaction at the same time.
  • A wrong answer that closes the conversation counts as a success.
  • It says nothing about the next week: the customer who comes back by phone sits in another report.
  • It rewards short conversations, while a good handover often takes one more message.

The four numbers to put on the dashboard

  1. Satisfaction on AI-handled conversations: a thumbs-up or thumbs-down on each one, read as a share once enough are rated.
  2. Automated resolution rate: AI-handled contacts with no repeat contact on the same issue within seven days.
  3. Escalations finished on the first human touch: the handover carried enough context that the person did not have to ask again.
  4. Repeat-contact rate within seven days: the honest check on the first three.

A lower deflection rate with high satisfaction and few repeat contacts beats a higher one with customers coming back. Track deflection too, for staffing, but read it next to these four.

When a handover counts as a success

Some conversations should reach a person, and a metric that punishes those handovers trains the wrong behaviour. The Unistaff AI employee hands over on the word HUMAN, a complaint, an emergency, a legal, tax or medical question, or a customer ready to pay. A missed handover demotes it automatically; a handover it makes never does.

The weekly review is built on the same idea. It shows results against the trade targets and three conversations to read — one rated down, one handed over, one random. Demotion is automatic when the thumbs-down share reaches 30% once five replies are rated, or after three rule hits in seven days.

Resolution needs access to the record

An AI that reads your documents can answer questions. Resolving often means acting: checking an order, changing a booking, issuing a refund. That needs a connection to the system that holds the record, plus rules on what the AI may change without a person.

Where Unistaff stands today: pre-sale questions are Live through the AI inbound sales manager on web chat and email. The support desk is On request: it drafts answers to repeat questions from your own docs, and a person sends each one. No helpdesk connector ships today, and the platform has no payment path, so refunds stay with your team.

Training for resolution, week by week

What you correct is what the AI learns. At Unistaff a thumbs-down with a comment becomes a lesson, and a wrong number becomes a proposed fix to the knowledge folder. A lesson applies only after a re-sit of the related exam cases without regression, so fixing one answer cannot quietly break another.

Questions people ask

The share of customer contacts closed without a human agent over a period: contacts handled by self-service or AI divided by all contacts. It measures workload kept away from people, and it counts a customer who gave up the same way as one who got the answer.

Deflection counts contacts that never reached a person. Resolution counts problems that were fixed, usually checked by no repeat contact on the same issue within seven days. A conversation can be deflected and unresolved at once.

The number says little on its own, because it rises when a person becomes harder to reach. Read it next to satisfaction on AI-handled conversations and the repeat-contact rate within seven days; a lower deflection rate with fewer repeat contacts is the better result.

Containment is measured inside one channel: the conversation ended with the bot. Deflection is measured across the support operation: the contact never reached a person anywhere. A contained chat followed by a phone call is contained but not deflected.

Track four numbers: satisfaction on AI-handled conversations, the share of AI-handled contacts with no repeat contact in seven days, the share of escalations finished on the first human touch, and the repeat-contact rate. The last one checks the first three.

It can answer questions from your documents. Acting on an order, a booking or a refund needs access to the system holding the record. At Unistaff the support desk is On request, a person sends each answer, and no helpdesk connector ships today.

Where to go next

Free 30-day pilot

Send your website. We reply with a pilot plan.

A person reads it and answers within one business day: which job description fits, what the pilot covers on your web chat and email, and what we need from you.

Or talk to a live AI sales manager first →

No card. How we store it: privacy policy.

Book a demo