What AI ticket auto resolution means
Auto resolution means the AI reads an incoming ticket, decides it understands the request, sends the answer to the customer and marks the ticket solved, with no agent in between.
It is different from AI drafting, where the AI writes a suggested reply and a person reviews and sends it. Drafting helps agents on every ticket. Auto resolution removes the agent from a subset of tickets.
Most teams should run both: drafting everywhere, auto resolution only where it has earned trust.
Three terms come up constantly:
- Intent: what the customer wants, such as "reset password" or "where is my order".
- Confidence: how sure the AI is that it recognized the intent and that its answer fits.
- Handover: passing the ticket to a person, ideally with a summary so the customer does not have to repeat themselves.
How to choose the first intents
Not every frequent ticket is a good first candidate. A good first intent scores well on four questions.
- Is it frequent? Automating something that happens twice a month saves nothing.
- Is the answer stable? The correct answer does not depend on judgment and does not change every week.
- Is it low risk? A wrong answer is annoying, not costly. Nobody loses money, access or data.
- Is the answer documented? There is an up-to-date knowledge base article the AI can ground its reply on.
Score each candidate on those four and start with the ones that pass all of them.
Usually good first candidates
- How-to questions answered by a single help center article ("how do I add a user", "how do I export to CSV").
- Password reset and login help that points to the self-service reset flow.
- Order status and shipping questions, when the AI can look up the order through your own system.
- Business hours, contact details and policy questions (returns window, supported countries).
- Plan and feature questions ("does the Team plan include live chat").
- Routine IT requests with a documented self-service path, such as connecting to the VPN or setting up email on a phone.
Usually poor first candidates
- Refunds, billing disputes and anything that moves money.
- Account deletion, ownership transfer or access changes.
- Legal, security or data protection requests.
- Angry customers and complaints, where tone matters as much as content.
- Bug reports that need investigation.
- Anything where the right answer depends on a person's judgment or an exception.
Some of these can be partly automated later. For example, the AI can collect the details for a refund request and hand over a complete summary. But do not let it decide.
Find your own intents in your ticket history
Pull the last three months of tickets and group them by what the customer wanted. For each group, note the volume and whether a single article answers it.
Suppose your team gets 3,000 tickets a month and one group, "how do I reset my password", makes up 240 of them with a clear article. That one intent is a better first step than ten rare ones combined.
Setting confidence thresholds
Every auto resolution system has a threshold: below it, the AI does not act on its own. Choosing it is a trade-off.
A high threshold means fewer tickets resolved automatically, but fewer wrong answers.
A low threshold means more tickets resolved automatically, but more mistakes that customers see.
A sensible way to set it:
- Start high. Accept that the AI resolves fewer tickets at first.
- Run in shadow mode for a week or two: the AI decides what it would have sent, a person reviews those decisions, and nothing goes to customers automatically.
- Check the reviewed decisions. How many of the tickets above the threshold would have been answered correctly?
- Lower the threshold in small steps only for intents where reviewed answers are consistently right.
Thresholds can differ per intent. A how-to question can run at a lower bar than an order status answer that involves a lookup.
Guardrails to set before going live
Guardrails are rules that apply regardless of confidence.
- Approved intents only. The AI resolves only intents the team has explicitly approved. Everything else, however confident, goes to drafting or to a person.
- Grounded answers only. Replies must be based on your knowledge base or your own systems, with the source visible to the team. No answers from general knowledge.
- Exclusion rules. Never auto resolve for certain customer tiers, certain keywords (such as "cancel", "lawyer", "refund"), negative sentiment above a level, or tickets that already had a human reply.
- Easy way back. Every automated reply tells the customer how to reach a person, and a customer reply reopens the ticket and routes it to the team.
- Full log. Every AI action is recorded: what it read, what it sent, which article it cited, what confidence it had.
- Data boundaries. Check where ticket data is processed and whether it is used to train shared models.
Handover: what happens when the AI is not sure
Handover is where many AI rollouts go wrong. The AI gives up, the customer lands in a queue, and the agent starts from zero.
A good handover:
- Happens early, not after three rounds of the AI failing.
- Includes a short summary: what the customer wants, what the AI already checked, what it found.
- Keeps the full conversation on one ticket, so the agent sees everything.
- Routes to the right team using the classification the AI already made.
- Does not pretend. The customer knows a person is taking over.
Look at handed-over tickets weekly. If one intent hands over often, either the article is missing something or the intent should not be automated yet.
How to measure whether it works
Automation that looks good in a dashboard can still annoy customers. Track several signals together.
| Metric | What to watch for |
|---|---|
| Auto resolution rate per intent | Growing steadily as you approve more intents |
| Reopen rate on auto resolved tickets | Should stay close to the reopen rate on tickets agents solve |
| Satisfaction on auto resolved tickets | Compare with agent-solved tickets on the same intent |
| Handover rate per intent | High rates point to missing articles or a poor candidate |
| Reviewed accuracy | Share of reviewed automated answers a person rates correct |
| Agent time freed | Hours no longer spent on those intents, and where that time goes |
Read a handful of automated answers every week, even after launch. Ten or twenty tickets per intent is enough to catch drift when a product change makes an article outdated.
Do not judge only by the resolution rate. A higher rate with rising reopens means you are closing tickets, not solving them.
A rollout plan for the first six weeks
- Week 1: pick three to five intents with the four-question test. Update their knowledge base articles.
- Week 2: turn on AI drafting for all tickets and shadow mode for the chosen intents. Agents review drafts and AI decisions.
- Week 3: review shadow results. Set thresholds per intent. Add exclusion rules for anything that went wrong.
- Week 4: go live on the two most reliable intents.
- Weeks 5 and 6: review automated answers weekly, watch reopens and satisfaction, add the next intents one by one.
Tell the team what is happening and why. Agents who understand that auto resolution removes repetitive tickets, not their jobs, give much better feedback on the drafts.
A note on cost models
Vendors price AI resolution in different ways. Many help desks bill AI per automated resolution on top of seats. Others sell AI assistance as a per-agent add-on.
Per-resolution billing has a side effect worth thinking about: the more successful your automation becomes, the larger the bill. Budget owners then have a reason to cap automation even when it works.
When comparing tools, work it through with your own numbers. If you expect the AI to resolve a few hundred tickets a month, calculate the monthly cost under each model at today's volume and at double.