Sign in Book a demo
Glossary
Glossary Metrics Updated Oct 14, 2026

What Is AI Deflection? Definition, Formula & Pitfalls

AI deflection is the share of customer support questions an AI agent or self-service tool handles without a ticket reaching a human. Teams track it as ticket deflection rate, a core metric for judging AI support.

In short
  • AI deflection is the share of support questions AI handles without a ticket reaching your team.
  • Formula: (AI conversations with no ticket ÷ total AI conversations) × 100.
  • A high rate can hide customers who gave up. Pair it with resolution rate and CSAT.
  • CareMaster raised self-service resolution from 30% to 76% with Brainfish, with no new support hires.

The catch: deflection counts tickets avoided, not problems solved, and customers notice when a bot stands between them and help. In a 2026 Gartner survey of 3,566 B2B and B2C customers, 87% said companies using GenAI for service must offer a way to reach a person.

Most customers aren't willing to give a bad bot a second chance, either. Only 27% would try a chatbot again after a negative experience, according to a follow-up Gartner analysis of the same survey.

Resolution changes the picture. After embedding Brainfish inside its product, healthcare admin platform CareMaster raised its self-service resolution rate from 30% to 76%. It absorbed 23% user growth without adding Customer Care staff.

How do you calculate AI deflection rate?

AI deflection rate = (AI conversations that didn't become a ticket ÷ total AI conversations) × 100

Say your AI agent handles 4,000 conversations in a month and 1,800 of them never turn into a ticket. Your deflection rate is 45%.

Teams define the inputs differently. Some include help center searches, and some subtract customers who came back days later with the same issue. Pick one definition and stick with it.

What's the difference between AI deflection and AI resolution?

Deflection tells you a ticket didn't happen. Resolution tells you the customer's problem actually got solved.

Metric What it counts What it can miss Best used for
AI deflection rate Conversations that never became a ticket Customers who gave up or went elsewhere Measuring load taken off your team
AI resolution rate Conversations where AI solved the issue on its own Whether the customer was happy with the answer Measuring whether AI actually works
Containment rate Conversations that stayed in the AI channel Whether anything was solved at all Spotting where handoffs happen
CSAT on AI conversations How customers rated the experience Volume and cost impact Checking that deflection isn't hurting CX

If deflection climbs while CSAT drops, customers may be getting stuck, not helped.

Why can a high deflection rate be misleading?

Because the easiest way to deflect a ticket is to make reaching a person hard. A bot that loops and keeps asking customers to rephrase can post a strong number while trust quietly drops.

The better goal is resolution: answer well enough that most people never need a ticket, and let the ones who do reach a person easily. CareMaster's COO describes what that changes for a support team:

"Brainfish freed our Customer Care team from handling routine queries so they could focus on solving the complex challenges our community faces that truly required human expertise." Elisha Anderton, COO, CareMaster

Where does knowledge-based deflection stop working?

Help articles handle "how do I" questions well, but support teams report they deflect only about 20 to 30% of volume.

The rest needs someone to look inside a system, like a double charge, a cancellation, or a customer locked out after changing phones. No article can close those tickets, so an AI that can only answer questions has to hand them to your team. Closing them takes an AI that checks the account, applies your policy, and takes the action.

How do you improve AI deflection without hurting the customer experience?

  1. Start with your highest-volume questions. Fix those first.
  2. Keep your knowledge current. Stale docs produce confident, wrong answers.
  3. Test with your own team first. Have support staff flag weak answers before customers see them.
  4. Make the human path obvious. Escalate when the AI isn't confident, and pass the full context along.
  5. Automate the actions, not just the answers. Refunds, cancellations, and account access can't be solved with an article alone.
  6. Review what the AI couldn't handle. Each miss is a content gap or a workflow to build.

How does Brainfish approach AI deflection?

Brainfish makes AI support agents for B2B CX teams that need accurate answers. They're designed to resolve tickets end to end, and every answer is grounded in your real docs and tickets, with its source cited. They work across chat, email, your help center, developer docs, Slack, and Microsoft Teams.

For requests that need action, Ballet by Brainfish runs workflows that check systems like Stripe or your CRM, apply your policy, and take the action. Anything over a limit you set or outside your policy stops for a person. When a ticket does need your team, it lands in Zendesk or Salesforce with every check Brainfish ran already attached.

The payoff shows up in resolution. At commerce platform Vrio, Brainfish's AI agent resolves 80 to 85% of developer questions automatically, so engineers now see only the complex edge cases.

Want to see how much of your queue AI could resolve? Bring a real ticket to a Brainfish demo, or see our pricing.

Frequently asked questions.

What is a good AI deflection rate?

There's no universal benchmark, because teams define deflection differently and ticket mixes vary. Support teams tell Brainfish that help articles alone deflect about 20 to 30% of volume. Going beyond that usually takes AI that can act, not just answer. Track your own rate over time, next to resolution rate and CSAT.

Is ticket deflection the same as AI deflection?

Mostly, yes. Ticket deflection is the broader term for any support request handled without a ticket, including help centers and FAQ pages. AI deflection refers specifically to requests handled by an AI agent or AI-powered search. Both are measured the same way: requests that never reached your team, as a share of the total.

Is deflection a bad goal for customer support?

Not on its own. Deflection becomes a problem when it's the only goal, because a bot can hit its number by blocking customers from a person. Aim for deflection that comes from real answers and completed tasks, keep the path to a human visible, and watch CSAT to make sure customers are being helped.

How is AI deflection different from containment?

Containment counts conversations that stayed inside the AI channel without transferring to a person. Deflection counts requests that never became a ticket. They often overlap, but neither one proves the customer's issue was solved. Resolution rate and CSAT fill that gap, so it's worth tracking all four together.

How does Brainfish improve deflection rates?

Brainfish answers customers from your real product knowledge, with a cited source on every answer, and uses Ballet by Brainfish to resolve requests that need action, like refunds and account access. CareMaster used Brainfish to raise its self-service resolution rate from 30% to 76% with no new support hires.

See it working on your own tickets.

Bring a real ticket to the demo and we'll show you Brainfish resolving it, grounded in your docs and systems.

Book a demo

No engineering sprint required.