Turning AI review into a marketplace advantage

Start where the business does repetitive work

A seller submits a listing. Someone has to assess whether it makes sense, whether the photos match, whether it belongs in the selected category and whether it raises concerns. Once approved, the marketplace has another decision to make: where should that listing appear?

At adeka.lk, I connected those two decisions. I built an AI-assisted review workflow whose structured assessments also contribute to listing quality and discovery. One assessment can support an operational decision and become a useful input to the customer experience.

This is where AI adoption becomes concrete: inside work the business already needs to do. The opportunity is to handle routine cases through explicit rules, give people better information for the exceptions, and make the resulting signals useful elsewhere in the product.

Review the listing in its actual context

The review combines a safety check with a marketplace-specific assessment of listing text and selected images. It considers category fit, location consistency, image alignment, description quality and potential risk. The prompt includes the category catalogue, relevant structured fields and Sri Lankan location context, with instructions to handle Sinhala, Tamil, English and mixed-language listings.

The output is structured: a proposed decision, risk level, confidence, flags and reasons for the reviewer. A possible category mismatch can therefore be represented as a specific finding instead of being buried in a paragraph of generated commentary.

That structure is what makes the model useful to an operational system. The application can evaluate the findings, retain the reasons and decide what happens next.

Automate the clear cases. Keep judgement available.

I separated the model’s recommendation from the application’s authority to approve. Automatic approval requires a set of conditions to pass together: low risk, sufficient confidence, a clear description, appropriate category and location checks, acceptable image alignment where applicable, and no review flags.

Uncertain cases remain available for manual review. Even a model recommendation to reject goes to a human reviewer rather than automatically rejecting the listing. If the AI call fails, the workflow records the error and also routes the case for manual review.

This makes automation a controlled path through the existing operation. Reviewers retain the difficult decisions, with recorded findings to help them inspect the case. Model confidence is one policy input, not a measured probability that the listing is safe.

Make the assessment useful twice

Approval and ranking answer different questions. Approval determines whether a listing can go live. Ranking helps decide the order in which eligible listings appear. I kept those responsibilities separate while allowing the assessment to inform both.

The quality scorer records five named signals: information availability, AI-assessed marketplace quality, category fit, media alignment and safety-related findings. Each carries a score, weight and reason. The system combines them through a bounded weighted reciprocal calculation, so the result remains inspectable and the signal set can evolve.

These are internal listing-quality signals. They are not customer reviews, seller reputation ratings or guarantees that an item is genuine. Their purpose is to help the product distinguish more complete, coherent listings from weaker ones.

Rank for quality, interest and freshness

The smart browse feed brings together listing quality, buyer interest and time. Within the quality blend, deterministic completeness contributes 65% and the review-derived quality signal contributes 35%. Observable details remain the larger part of that blend: an AI assessment does not replace the basics of a well-formed listing.

Engagement adds another perspective. Saves and phone reveals carry more weight than passive views, while logarithmic scaling reduces the incremental influence of increasingly large counts. Time decay gives freshness a role, so accumulated activity is not the only factor shaping the feed.

The database maintains the final score, and ordinary client writes cannot directly set the review-derived ranking inputs. The smart browse path orders by score and then creation time; search and nearby discovery have their own paths. Ranking changes order without granting a pending listing public visibility.

The product intent is straightforward: help useful listings surface through more than recency alone. Whether that improves enquiries or completed transactions needs measurement; a scoring formula is a design choice, not evidence of commercial uplift.

A practical starting point for an existing business

The transferable pattern starts with a recurring decision. A marketplace reviews listings; a catalogue team checks product information; an operations team triages incoming requests. Define the evidence needed for that decision and ask the model for structured findings that the workflow can use.

Then define the boundaries: which cases can proceed automatically, which require a person, and what happens when the model is unavailable. Preserve enough detail to inspect the decision later. Where appropriate, reuse those findings for prioritisation, routing or discovery.

Before expanding automation, measure manual-review agreement, incorrect approvals, time to decision and the work that remains for the team. For discovery, track whether people find relevant items and make useful enquiries. These are the outcomes to validate, rather than treating the number of AI calls as success.

Build the operating system around the intelligence

My focus at Adeka was the connection between product judgement and engineering: the review contract, approval policy, exception path, scoring model and the feed that consumes those signals. The value comes from those pieces working together in a usable marketplace.

For a business considering AI, this is an approachable place to start. Choose an operational bottleneck, make the decision explicit, and integrate intelligence into the workflow around it. That creates something people can use, inspect and improve.