# AI Customer Service: How AI Agents Resolve Tickets Instead of Deflecting Them

> AI customer service that resolves, not deflects: how deflection, containment and resolution differ, how vendors count them, and how to measure the real thing.

- Canonical: https://buffersoltech.com/blog/ai-customer-service
- Author: BufferSol Editorial Team
- Published: 2026-10-27 · Updated: 2026-10-27

## Key takeaways
- Deflection means the customer did not reach a person. Resolution means their problem was actually solved. AI customer service should be judged on resolution.
- Vendors now separate these outcomes themselves: Intercom distinguishes confirmed from assumed resolutions, and Zendesk distinguishes verified from contained ones.
- A customer who goes quiet may be satisfied, or may have given up and called instead. Check reopen and repeat-contact rates before trusting a resolution figure.
- Real resolution usually needs action in your systems, such as an order lookup or a return label, not just an answer.
- Measure resolution in the system of record and with a weekly sampled review, not only with the AI tool's own dashboard.

AI customer service resolves a ticket when the customer's problem is actually solved: the order is found, the return is created, the address is changed and the customer knows what happens next. Deflection only means the customer did not reach a person. The two can look the same on a dashboard, so the way to tell them apart is to check what happened in your systems and whether the customer came back.

This post explains the difference between deflection, containment and resolution, how vendors count them, why a quiet customer is not proof of a fix, and how to measure real resolution. It draws on the [AI customer support agent](/portfolio/ai-customer-support-agent) BufferSol built for an online retailer, and it complements our full [AI customer support agent guide](/blog/ai-customer-support-agent-guide), which covers design, integrations and the pilot plan.

## Deflection, containment and resolution

Three words get used as if they meant the same thing. They do not.

```infographic
{"type":"compare","title":"Three ways to count an AI customer service conversation","columns":["What it counts","What it misses"],"rows":[
{"label":"Deflection","a":"Conversations that did not reach a human agent","b":"Whether the customer's problem was solved, or they gave up"},
{"label":"Containment","a":"Conversations the AI handled with no human involvement or follow-up","b":"Whether the outcome was complete and correct"},
{"label":"Resolution","a":"Conversations where the problem was solved, confirmed by the customer or by other signals","b":"Little, if it is checked against reopens and repeat contacts"}
],"source":"BufferSol; terms as used in Intercom and Zendesk documentation"}
```

Deflection was a reasonable target when self-service meant help-center articles: if a customer read an article and did not open a ticket, that saved time. With AI agents that can act, it is the wrong target. A support bot that is hard to get past will have excellent deflection numbers and unhappy customers.

## How vendors count resolutions

The major helpdesk vendors now separate these outcomes themselves, which is useful, because their definitions show exactly where the uncertainty sits.

Intercom's documentation on [Fin resolutions](https://www.intercom.com/help/en/articles/8205718-fin-resolutions) counts a resolution when, after Fin's last answer, the customer "either confirms the answer was satisfactory (confirmed resolution), or exits the conversation without requesting further assistance (assumed resolution)." A handover to a person does not count, a reply to a greeting is not an answer, and if the customer returns to the same conversation for more help, the resolution is deducted.

Zendesk announced [changes to its AI agent reporting](https://support.zendesk.com/hc/en-us/articles/10677925692698-Announcing-changes-to-AI-agent-reporting) in May 2026. It now separates a contained resolution, a meaningful request handled by the AI agent with no human involvement or follow-up, from a verified resolution, which has additional signals confirming the outcome was complete and satisfactory. It also tracks assisted escalations, where the AI agent contributed before a person finished the job. Only verified resolutions are billed.

Two lessons follow. First, numbers from different tools are not comparable until you read the definitions. Second, both vendors treat a customer going quiet as weaker evidence than a confirmation, and so should you.

## Why a quiet customer is not proof of a fix

When a customer stops replying, several things may have happened:

- The answer solved the problem. This is the case everyone hopes for.
- The answer was wrong, and the customer did not notice yet.
- The customer gave up and phoned, emailed or messaged on social media instead.
- The customer gave up and did nothing, which may show up later as a refund request, a chargeback or a lost customer.

A wrong answer can also cost more than no answer. If the AI states a policy that does not exist, the customer may hold you to it. That is one reason we ground every answer in approved content and let the agent act only through tools that enforce the real rules.

## Resolution usually needs action, not just an answer

Many requests cannot be resolved with words alone. "Where is my order?" needs a live lookup. "This is the wrong size" needs a return created and a label sent. "Please change my delivery address" needs the order updated before it ships. A system that only explains the returns policy will deflect those tickets, not resolve them.

OpenAI's [practical guide to building agents](https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf) describes agents as systems that use tools to gather context and take actions in external systems, within guardrails, and recognize when to hand control back to a person. That is what resolution requires. Our e-commerce support agent is connected to the store, the helpdesk and the shipping provider. It checks orders and live shipment status, creates returns and labels within the store's rules, answers product questions from the catalog and replies in the brand's tone. Refunds above a threshold and unusual cases go to a person with a summary and a suggested reply.

For the difference between answering and acting, see our post on the [AI agent vs a chatbot](/blog/ai-agent-vs-chatbot). For questions that only need an accurate answer, such as staff policy questions, an answer-focused assistant like our [enterprise knowledge assistant](/portfolio/enterprise-knowledge-assistant) can be the right tool.

## How to measure real resolution

Measure resolution in your own systems as well as in the AI tool's dashboard. These checks work together:

| Check | How to measure it | What it catches |
| --- | --- | --- |
| Action completed | The return, update or booking exists in the system of record | "Resolved" conversations where nothing actually happened |
| Reopen rate | Same ticket reopened within a set number of days | Answers that did not hold |
| Repeat contact | Same customer contacts you on any channel within a set number of days | Customers who gave up and tried elsewhere |
| Confirmed vs assumed | Share of resolutions the customer confirmed | How much of the figure rests on silence |
| Sampled review | Support leads grade a weekly sample of AI conversations | Wrong answers, wrong tone, missed escalations |
| Customer rating | Rating on AI-handled conversations, compared with human ones | Whether customers accept the experience |
| Handover quality | Agents rate whether the AI summary was accurate and useful | Escalations that cost the team time |

Pick the windows for reopen and repeat contact with your support team, based on how long your customers typically take to come back, and keep them fixed so trends are comparable. Matching customers across channels needs an identifier you can rely on, such as an email address or order number, so plan that join early.

Then report one number to leadership: resolutions that passed these checks. It will be lower than the dashboard figure, and it will be the one that holds up.

## Designing for resolution from the start

If resolution is the goal, it changes how the system is built.

1. **Start with requests the agent can finish.** Order status, returns within policy, address changes before shipping and booking changes are good first choices because the outcome can be checked in a system.
2. **Give the agent the tools those requests need**, with narrow permissions, validated inputs and approval for high-value actions such as large refunds.
3. **Write policies as rules the tools enforce**, not only as text the model reads, so the agent cannot promise something your systems will not do.
4. **Make handover easy and fast.** A customer who asks for a person should get one, with the context passed along. A quick, well-briefed escalation is a good outcome; a customer trapped in a loop is not.
5. **Tell customers they are talking to AI.** In the EU, the [AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) sets transparency duties, such as telling people when they are interacting with an AI system, and the European Commission lists these as taking effect in August 2026. Being clear also makes the route to a person easier to find.

## Getting started

Export a month of tickets and tag each one by request type and by what resolving it required: an answer, a lookup or an action in another system. Then check your current AI or self-service figures against reopen and repeat-contact rates for the same period. The gap between the two shows how much of today's "resolution" is really deflection, and the action-type tickets show where an AI agent can close it. Our [AI agent development](/services/ai-agents) team can test the strongest request type on your own helpdesk data.

## Frequently asked questions
### What is the difference between deflection and resolution in AI customer service?
Deflection counts conversations that did not reach a human agent, whatever happened to the customer. Resolution counts conversations where the customer's problem was solved. A customer who gives up, or contacts you again by phone, is deflected but not resolved. That is why resolution, checked against reopen and repeat-contact rates, is the better measure.

### How do AI customer service vendors count resolutions?
Definitions differ by vendor. Intercom counts a Fin resolution when the customer confirms the answer helped, or leaves without asking for more help, and does not charge if they return to the same conversation. Zendesk separates verified resolutions, confirmed by additional signals, from contained ones with no human involvement, and bills only verified resolutions. Read each definition before comparing numbers.

### What is a good way to measure real resolution?
Combine three checks. First, confirm the action happened in the system of record, such as the return created or the address updated. Second, track whether the same customer reopens the ticket or contacts you again on another channel within a set number of days. Third, review a weekly sample of AI-handled conversations with your support team.

## Sources
1. [Fin resolutions](https://www.intercom.com/help/en/articles/8205718-fin-resolutions), Intercom
2. [Announcing changes to AI agent reporting](https://support.zendesk.com/hc/en-us/articles/10677925692698-Announcing-changes-to-AI-agent-reporting), Zendesk
3. [A practical guide to building agents](https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf), OpenAI
4. [AI Act](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai), European Commission