Helios Retail

AI agents supporting a software company's technical desk

Three agents in production supporting the help desk of a retail software vendor: triage, guided diagnosis and suggestions grounded in the product's real code.

Client
Helios Retail
Secteur
Retail software vendor
Période
2024 - 2025
Technologies
LangGraphGeminiQdrantCloud RunMCP

95%

search accuracy, against roughly 60% before

Every suggestion comes with a confidence score and its sources

A feedback loop that keeps adding to the knowledge base

Helios Retail publishes a suite of point of sale management software. Its technical desk faced the problem every software vendor knows: engineers re-diagnosing incidents that have already been solved dozens of times, with the knowledge locked inside old tickets, phone calls and the memory of a few experts.

Three agents, three uses

Three specialised agents were put into production on a common technical base.

The resolution agent first judges whether the ticket contains enough information to work with, and if not, generates the clarifying questions to ask. It then searches in parallel through similar resolved tickets, fixes already scheduled and the documentation, and drafts a proposal that can be acted on, with a confidence score and its sources. The engineer stays in the loop: the agent proposes, the person decides.

The guided diagnosis agent offers an interactive mode. It prescribes a check to run, interprets the engineer’s answer, including screenshots, and refines its hypothesis.

The self-service agent, on the end user side, answers questions about how to use the product, staying strictly grounded in the product documentation, filtered by retail brand.

Anti-hallucination as an architectural principle

An agent that invents a table name is worse than useless. The answer was to index the product’s real code, over five hundred thousand lines of SQL and around five thousand classes, combining exact, full-text and semantic search. Search accuracy, measured internally, went from roughly sixty per cent to ninety-five, and the agent reasons about the real names of tables, procedures and screens.

Two further mechanisms complete the system. Detection of already-fixed incidents compares the version installed at the customer site against scheduled fixes and flags the cases where the answer is simply to update. And a feedback loop adds to the knowledge base every time an engineer validates a suggestion.

Next step

Where would you start?

Ten days to map where your teams lose time, put a figure on each lever and name the first system to build. If AI is not the right answer, I will tell you so.

Diagnostic

Étape 1

10 days

€9,500 excl. VAT · fixed fee, one area of the business

  • A map of where the time actually goes in the area audited
  • Each lever costed in hours freed, ranked by effort and by gain
  • The data note: what leaves your systems, and where it is processed
  • A three-month plan, naming the first system to build