AI resolves questions that are documented, repetitive, and low risk: access and login, order and delivery status, plan and feature explanations, billing policy, and setup how-tos. It resolves badly anything requiring judgement, empathy, a commercial exception, or information your documentation does not contain.
Key takeaways
- Documented, repetitive, and low risk is the profile that works.
- Judgement, empathy, and exceptions are not automation failures, they are the human job.
- Your topic distribution determines your ceiling, not the vendor.
- The gap log from a live deployment is the most honest data you will get.
Vendor resolution rates are not comparable to each other and they are not predictive of your result. What is predictive is your own topic distribution. Here is the map.
The profile that works
Three characteristics, all of which need to be present:
Documented. The answer exists in written form somewhere you can point a system at. If it lives only in a senior agent’s head, no system resolves it.
Repetitive. The same question, or the same underlying question phrased differently, arrives frequently. One-off problems have nothing to learn from.
Low risk. A wrong answer is embarrassing rather than expensive. Nobody loses money, nobody is unsafe, no commitment is created.
Miss any one of those and the question belongs with a person.
Categories that resolve well
Access and login. Password resets, account lockouts, single sign-on routing, device limits. The highest-volume category in most queues, almost entirely documented, and urgent to the person experiencing it. This is usually the single biggest win available.
Order and delivery status. Where is it, what does this status mean, when will it arrive, what happens if it is late. Dominant in e-commerce and repetitive to a degree that surprises people.
Plan, feature, and pricing explanations. What is included, what is different between tiers, how does this feature work. High volume in software, fully documented if your content is good, and frequently a buying signal as well.
Billing policy questions. Proration, renewal timing, invoice access, why a charge looks different. High emotion, high volume, and entirely documentable. Route disputes to people, but the policy explanation itself is a strong candidate.
Setup and how-to. For software products this is where trial users get stuck, and where a slow answer costs a customer rather than generating a complaint.
Returns and policy questions. What is the window, how do I start one, what condition must it be in. Documented, repetitive, and spiking exactly when your team is busiest.
Categories that resolve badly
Complaints. Not because the words are hard, but because a complaint is a request for acknowledgement from a person. Automating it escalates it.
Cancellations. Explaining how cancellation works can be automated. The conversation itself is the highest-value retention moment you get, and it deserves a person with the context already gathered.
Commercial exceptions. Anything requiring someone to decide whether to bend a rule. That is judgement by definition.
Anything undocumented. A system grounded in your content cannot answer what your content does not contain. It should say so, and that refusal is correct behaviour rather than failure.
Bespoke technical debugging. Account-specific problems requiring investigation. The system can gather context, and the diagnosis needs a person.
Vulnerability and financial difficulty. These need a human by rule, not by confidence score, and in regulated sectors that is an obligation rather than a preference.
How to predict your own ceiling
Do not ask a vendor. Run this instead:
- Categorise a thousand real tickets by topic.
- Mark each topic documented, documentable, or judgement.
- Sum the volume in each column.
Documented is your realistic ceiling today. Documentable is what content work would add, and it is usually larger than expected. Judgement is what your team will always handle, and it is the work worth having people do.
A queue that is sixty percent documented plus documentable supports a very different outcome from one that is twenty percent. That is the number that predicts your result, and no vendor can tell it to you.
Why documentation quality is the real variable
Two companies buying identical products get very different results, and the difference is almost never the model.
It is whether the content covers the questions, whether it contradicts itself, and whether procedures are kept whole rather than split across pages. We wrote about structuring documentation for retrieval separately.
This is also why the audit before deployment matters more than the vendor comparison. It improves every option you are considering simultaneously.
The most useful output of a live deployment
Not the resolution rate. The gap log.
A grounded system should record every question it could not answer from your content. That log is a continuously updated list of what your documentation is missing, in real customer wording, ranked by frequency.
Most teams find that log more valuable than the automation itself in the first quarter, because it converts a vague sense that the documentation could be better into a ranked, evidence-backed roadmap.
The honest summary
If your queue is dominated by documented repetitive questions, automated resolution recovers real capacity and improves the experience for those customers, because they get an instant answer instead of a queue.
If your queue is flat, bespoke, and judgement-heavy, it will not do much, and you should know that before a sales process rather than after one. That is a genuine outcome and the queue analysis tells you which you have in an afternoon.
Frequently asked questions
What can AI resolve in customer support?
What should AI never handle?
What resolution rate should I expect?
How do I predict my resolution rate before buying?
Resolve, don't deflect.
See Fidiora resolve a ticket, capture a lead, and keep the bill predictable.