# Returns that resolve themselves. Almost.

> Returns intelligence is an illustrative veridive engagement for a multi-brand home and living retailer in Türkiye. Over ten weeks, a Discovery Sprint and a Pilot to Production engagement produce a returns assistant that checks orders and photos, cites the policy paragraph behind each recommendation and leaves the final decision to a person. Figures are pilot targets.

An illustrative engagement for a multi-brand home and living retailer: a returns assistant that prepares each decision with the evidence attached, so the team confirms, adjusts or escalates.

- **Client (illustrative):** A multi-brand home & living retailer · 120 stores and online · Türkiye
- **Scope:** Discovery Sprint → Pilot to Production · 10 weeks · Team of four

## Key results

Illustrative pilot targets.

- **Return decision time:** 3 days → same day
- **Cases needing full manual review:** 100% → about 35%
- **Policy consistency across channels:** measured weekly

## The situation

Returns arrived through stores, the website, marketplaces and the contact center — each with its own form and its own rules. A small team reviewed every case by hand: reading the customer’s message, checking the order, looking at photos, and applying a policy that had grown to forty pages. Decisions took days, and the same case could get different answers in different channels.

## The question

Could a system prepare each return decision — with the evidence attached — so the team only had to confirm, adjust or escalate?

## Evidence first, then a person decides.

What we built:

1. A triage assistant that reads the request, pulls order and product data, and checks photos against the product record.
2. A policy engine that cites the exact paragraph behind every recommendation.
3. A review screen where the team confirms, edits or escalates in one click.
4. Weekly quality reports comparing recommendations with final decisions.

The returns flow: order & product data (read-only) → photo check → policy retrieval with citations → recommendation + confidence → human review → decision logged. A person decides at step five.

## Ten weeks, as a stream.

How we worked:

- **Weeks 1–2 · Align & validate.** Walked the returns process in two stores and the contact center. Built an evaluation set from 400 past cases. Agreed acceptance criteria.
- **Weeks 3–6 · Build.** Weekly demos with the returns team. Integrated order data and the policy library. Designed the review screen with the people who’d use it.
- **Weeks 7–8 · Pilot.** Ran alongside the existing process on live cases and compared decisions daily.
- **Weeks 9–10 · Embed.** Trained the team, wrote the runbook, set up monitoring and a monthly review.

## Evidence attached. Judgment where it counts.

What changed:

- Most standard cases now arrive with a recommended decision, the evidence and the policy paragraph behind it.
- The team spends its time on unusual cases: damaged goods, repeat returns and high-value items.
- Customers get an answer the same day in most channels.
- Store managers see the same policy applied everywhere.

## Guardrails in this project

- Refunds above a threshold always need a person.
- The system never contacts customers directly.
- Photos are processed in-region and deleted after the decision.
- Every recommendation shows its sources.

## What’s next

Extending the same approach to warranty claims and marketplace disputes.

## Have a workflow like this?

[Start a project](https://veridive.com/contact/) · [All work](https://veridive.com/work/) · Next engagement: [Supplier document desk](https://veridive.com/work/#supplier-desk), food manufacturing. One desk for every supplier promise.