The box looks battered. The claim looks straightforward. The photograph leaves rather a lot to the imagination.
For a business handling deliveries, returns or supplier defects, that’s a familiar problem. Someone has to inspect the evidence, find the order and decide what happens next. Meanwhile, the customer wants a replacement and the warehouse wants the item off its bench.
Liquid AI’s d1-ultra offers an interesting approach to this work. It evaluates text and images and returns probabilities rather than conversational answers. Its damage-claim and defect-checking examples suggest a practical starting point: sorting cases for review.
The useful question is where it could help your team make a quicker, better-supported decision.
Give the model a narrow job
A sensible first task might be assessing whether a delivery photograph shows visible packaging damage.
That’s a smaller responsibility than deciding whether a customer deserves a refund. A refund decision can involve delivery records, contract terms and evidence the photograph simply doesn’t contain.
A crushed carton could contain a perfectly intact product. An immaculate box could contain a broken component. A blurry image could mean someone has a poor camera, rather than a flourishing career in claims fraud.
Keep those distinctions in the process. Ask the model to assess a defined visual condition, then use that assessment alongside the rest of the case.
For an initial pilot, let it recommend a review route. Leave financial decisions with an authorised person.
Treat probabilities as scores to test
A probability sounds reassuringly precise. Precision on a screen, however, needs checking against real outcomes.
Suppose a system assigns a high probability to “visible damage”. You still need to establish how often cases receiving that score actually show damage under your agreed definition. Test that using examples your team has already reviewed.
Your evidence will vary. Warehouse lighting changes. Customers photograph parcels on patterned carpets. Reflective wrapping can make an intact product look cracked.
Before choosing a threshold, compare the model’s assessments with human judgements across those conditions. Review disagreements carefully. Sometimes the model will be wrong; sometimes your existing guidance will be unclear.
Don’t assume a probability measures everything you care about. A score for visible damage says nothing, by itself, about who caused it or whether the photograph belongs to that order.
Build a route for uncertain cases
The most valuable output may be a better-organised review queue.
Clear evidence can go to the appropriate reviewer with the relevant records attached. Poor-quality photographs can trigger a request for better evidence. Ambiguous cases can reach someone experienced enough to investigate.
You’ll need to define those routes before connecting the model to live work. Assign an owner to each queue and set a review deadline. Give staff a clear way to override the recommendation and record why.
Keep a fallback for technical failures too. If the service is unavailable or the image cannot be processed, the claim should still reach a person. A customer’s replacement shouldn’t disappear into a digital cupboard because an integration had a difficult afternoon.
During the pilot, keep human review across all routes so you can spot cases being sent to the wrong place.
Separate damage checks from fraud checks
“Damaged delivery” and “dodgy photo” make a neat headline. Operationally, they need different evidence.
A damage assessment looks for a condition in an image. A fraud investigation considers whether a claim is deliberately misleading. One cannot safely stand in for the other.
A model might identify something that deserves closer inspection, but that doesn’t establish intent. Similar photographs might have an innocent explanation. Missing details might reflect confusing instructions.
If you want to investigate image reuse, order mismatches or altered evidence, define those as separate checks. Establish what tools and records each check requires. Don’t assume a general image assessment provides image authentication.
Customer-facing language matters here. “Please send a clearer photograph of the damaged corner” is a useful request. An unsupported accusation is a complaint waiting for a reference number.
Keep customer information under control
Delivery photographs can capture more than a dented box. Labels may expose names and addresses. Backgrounds may show people, homes or confidential paperwork.
Before uploading images to an external service, establish what data leaves your systems and where it goes. Check retention arrangements, access controls and whether the provider can use submitted material for training. Review the relevant contractual terms and data protection obligations.
Collect only the evidence needed for the task. Where practical, crop or obscure unrelated personal information before processing. Keep the original securely when your claims process requires it.
Staff also need an approved route for submissions. Otherwise, the careful pilot can become several people uploading customer photographs through personal accounts. Convenient, until somebody asks where the data went.
Measure the whole job
A fast assessment can still create more work overall.
If reviewers spend extra time correcting recommendations, chasing unnecessary photographs or finding missing order details, the business gains little. Measure the process from the claim arriving to its resolution.
Useful measures include handling time, repeat requests for evidence and the number of decisions reviewers overturn. Track missed damage as well as undamaged items flagged for investigation.
Look separately at different products and image conditions. Performance on large cardboard cartons may tell you little about scratched metal fittings or transparent packaging.
Agree your acceptance criteria before the pilot begins. Include the cost of mistakes. An unnecessary review and an incorrectly rejected claim have different consequences.
Start with a controlled pilot
Choose one product group and one clearly defined assessment. Use a set of previously resolved cases, with appropriate permissions and safeguards, to test the approach before live use.
Then run it alongside your existing process. Compare recommendations with staff decisions and examine disagreements. Confirm that records remain traceable and that reviewers can understand what the score represents.
Expansion should follow evidence that the process helps. Keep checking after deployment, particularly when products, packaging or image sources change.
For UK businesses, the practical opportunity is a more orderly claims process with people handling the judgements that need context. That’s worth exploring carefully.
If you’re considering image-based claims triage, talk to Morse Networks about the IT questions to resolve before a pilot, from data access to integration and fallback arrangements.
Let’s make IT one less thing to think about. morsenetworks.co.uk.
Sources
- d1-ultra: Introducing the World’s Best Decision Model — Liquid AI · 15 Oct 2026