Retail
AI customer assistant
A grounded assistant that answers order and product questions from real systems, and hands over to a person the moment it should.
- Fewer tickets reaching an agent
- 42%Fewer tickets reaching an agent
- Answers with a verifiable source
- 91%Answers with a verifiable source
- Median response time
- < 2sMedian response time
- Client
- A multi-channel retail group
- Sector
- Retail
- Duration
- 5 months
- Stack
- FastAPI
- pgvector
- PostgreSQL
- Next.js
- Redis
The problem
Support volume grew faster than the team, and roughly two-thirds of contacts were routine order-status and returns questions. An earlier off-the-shelf chatbot had been withdrawn after it answered confidently and incorrectly about refund eligibility.
What we did
- Grounded every answer in retrieval over the actual policy documents and order system, with the source cited in the reply.
- Built an evaluation set from real historical tickets and gated deployment on measured accuracy rather than impressions.
- Defined explicit refusal and escalation rules for anything touching payments, personal data or refund decisions.
- Instrumented the handover so an agent receives the full conversation and the retrieved sources, not just a transcript.
The outcome
Routine enquiries resolve without an agent, and the assistant declines rather than guesses on anything with financial consequences. Agents now spend their time on the cases that needed a person all along.
Illustrative engagement profiles. Figures show the shape of outcomes we design for, not published client results.
Services
Practices used here
AI Agents & Workflow Automation
Language models, retrieval and agents built to run in production with evaluation and guardrails.
Learn more about AI Agents & Workflow AutomationWeb & Enterprise Applications
Typed, tested platforms that integrate with the systems already running your business.
Learn more about Web & Enterprise Applications
More work
Other engagements
Healthcare
60%
Reduction in onboarding time
Healthcare learning platform
Replacing a classroom-and-paper induction programme with a digital learning platform that new clinical staff complete at their own pace.
Read the engagement — Healthcare learning platformManufacturing
35%
Improvement in operational efficiency
Manufacturing operations dashboard
Bringing machine data from four decades of equipment into one view, then using it to predict failures instead of reacting to them.
Read the engagement — Manufacturing operations dashboard
Start with a conversation
Tell us what you are trying to build and what is in the way. We will tell you whether we are the right people for it.