Split the claim into checks the agent can finish and decisions it cannot. The agent confirms the product, the purchase date and the warranty term, collects fault evidence, and issues a repair or return authorisation for clear in-cover claims. Refunds, replacements above your limit, suspected misuse and anything the customer disputes go to a named person.
What is a warranty claim, and who owns the process?
A warranty claim is a customer asking you to honour a promise made at sale: the product will work, for a stated period, or you will repair, replace or refund it. Three questions decide every claim. Is the product covered? Is the fault the kind the warranty covers? Which remedy applies?
The head of customer support usually owns the process, but the warranty terms belong to product or legal, and the remedy budget belongs to finance. Write those three owners down before you start. The agent will refer to each of them by name when it holds a case. For the wider position on what support agents should and should not do, see our guide to the AI support agent that knows when to say no.
How does the agent handle a claim, step by step?
- Agent, depth 1: identifies the product and purchase. It matches the serial number or order reference to the order record, reads the purchase date, the warranty term and any extended cover. Evidence: the order record cited by source.
- Agent, depth 1: checks the claim history. Previous claims on the same unit, previous replacements, open returns. A third claim on one unit is not routine. Evidence: history read and attached.
- Agent, depth 2: collects the fault description and evidence. Photos, a short video, error codes, when the fault started and how the product was used. Evidence: the customer's statements marked as declared, files stored against the case.
- Agent, depth 2: classifies the fault against the published fault list. Manufacturing defect, wear outside cover, accidental damage, or unclear. It does not invent categories. Evidence: the classification and the fault-list entry it relied on.
- Agent, depth 3: issues a repair or return authorisation. Only for in-term, in-cover claims with a clear defect classification, under the value limit you set. It books the collection or sends the returns label. Evidence: the authorisation number and the verdict that allowed it.
- Person, warranty lead: decides held claims. Out-of-term claims, misuse disputes, goodwill gestures, and any refund. Evidence: the held verdict, the decision and the reason given.
- Agent, depth 2: drafts the outcome message for held cases, which the warranty lead sends or edits. For allowed cases it sends an approved template at depth 3.
Registering the claim in the system of record is its own capability; set it up as described in letting an AI agent register claims in your claims system.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Identify | Matches serial or order to the purchase | 1 | Nothing |
| History | Reads prior claims and returns | 1 | Repeat-claim units |
| Evidence | Collects photos, codes, account of the fault | 2 | Nothing |
| Classify | Maps the fault to the published list | 2 | Unclear or disputed faults |
| Authorise | Issues repair or return authorisation | 3 | Claims above the value limit |
| Remedy | Prepares the case | 2 | Refunds, goodwill, out-of-term |
| Outcome | Sends approved template or drafts reply | 2 or 3 | Any reply to a dispute |
Why does the agent stop at repair and return?
A repair or return authorisation is reversible. If the unit arrives and the fault is not what the photos suggested, the service centre reports back and the claim is reassessed. Nothing has left your accounts.
A refund is different. Money leaves, the customer may already have bought a replacement elsewhere, and recovering it is awkward. That is depth 4 territory, and most support teams should keep it with a person until the agent has a long record of accurate classification. Compensation for the inconvenience of a failure is a separate matter again; see handling compensation claims for service failures.
Replacements sit in the middle. Shipping a new unit costs money but is recoverable if the faulty one is returned. Allow them at depth 3 only below a value limit and only when the faulty unit is collected first.
Which controls matter, and what evidence do they leave?
Cover is read, not believed. A customer saying "I bought it last spring" is declared. The purchase date in your order record is sealed. Only the sealed date decides whether the claim is in term. Evidence: the source class on the purchase date.
Fault list as the only vocabulary. The agent classifies against the published list or says "unclear". Evidence: the list version cited on each classification.
Value limit enforced at the action. The authorisation is checked against the customer's own limit at the moment it is issued, not when the conversation started. Evidence: the verdict naming the bound that decided.
Customer rights stated, not decided. In many markets a customer has rights against the seller that sit alongside any manufacturer warranty. The agent tells the customer the claim is being reviewed rather than declaring that they have no remedy. Evidence: the held verdict on any refusal.
Every refusal goes to a person. The agent never sends a final "not covered" on its own. Evidence: no refusal message without a named approver. The detail of when the agent should decline is covered in whether an AI agent can handle warranty claims.
What should you measure?
- Classification accuracy, checked against what the service centre finds when the unit arrives. This is the number that earns more depth.
- Time from first contact to authorisation for in-cover claims.
- Held claims by reason, and how long each waits for the warranty lead.
- Units returned with no fault found, split by whether the agent or a person authorised them.
- Repeat claims within a short window after a repair, which signals either a bad repair or a misclassified fault.
Where does it usually go wrong?
The serial number does not match the order. Gifts, second-hand purchases and resellers all break the link. The agent should hold these, not reject them, because the warranty may still apply.
Photos are treated as proof. A photo shows a crack; it does not show how the crack happened. Treat photo evidence as external and let it support a classification, never settle a misuse dispute.
The fault list goes stale. A new model launches with new known faults, and the agent keeps classifying them as "unclear". Product owns the list and must update it at each launch.
Out-of-term claims are refused flatly. A claim a few days outside term, on a known defect, is often worth a goodwill repair. That is a person's call, so hold it rather than refuse it.
Where Pryme fits
In Pryme, issuing a repair or return authorisation and issuing a refund are separate capabilities, each with its own certificate and depth. The purchase date from your order system arrives as sealed; the customer's account of the fault and their photos arrive as declared or external. If the agent tries to refund, the verdict names the bound that held it and the warranty lead who must decide.
Questions people ask
Can the agent decide that a fault was caused by misuse?
It can flag that the evidence suggests misuse, and record why. The decision to refuse cover on that basis stays with a person, because it is the point customers most often dispute.
What if the customer bought the product from a retailer, not from us?
The agent asks for the retailer's proof of purchase and treats it as external evidence. If the warranty terms allow third-party purchases, the claim proceeds; if they are unclear, it is held.
Should the agent tell the customer it is not a person?
Yes. Say so at the start of the conversation and explain how to reach a person. Under the EU AI Act people must be told when they interact with an AI system, and it is good practice everywhere.