Live chat for an automotive brand that only says what the brand would sign off on.

Client

Capability story, built for a UK automotive brand tender

How it ran

5 weeks · 3 sprints · 2 engineers and a designer · internal build

Stack and integrations

  • Controlled RAG knowledge base
  • Model routing
  • Dealer CRM webhook
  • Agent console
  • UnflowUI™
Live chat for an automotive brand that only says what the brand would sign off on

Client

Capability story, built for a UK automotive brand tender

How it ran

5 weeks · 3 sprints · 2 engineers and a designer · internal build

Stack and integrations

  • Controlled RAG knowledge base
  • Model routing
  • Dealer CRM webhook
  • Agent console
  • UnflowUI™
0 off-brand answers in a 400-question test

Zero off-brand or invented answers across a 400-question test set. Handover to a human in under four seconds. The demonstration ran live for the tender panel.

01 · The situation

What it was costing them before we started.

Automotive brands run live chat to answer questions, capture leads for dealers and hand over to people when it matters. Most deployments fail one of three ways: the bot invents specifications, it cannot hand over cleanly, or it says something the brand would never say.

  • 3 ways an OEM chat deployment fails: it invents specifications, it cannot hand over, or it speaks off-brand.
  • 1 invented towing figure is enough to end a brand’s appetite for the whole category.
  • 0 tolerance for either, in a tender scored by a brand’s legal and marketing teams together.

02 · The approach

How we went about it.

Built as a capability story rather than a client project, which changed the method: we wrote the test set before the product. Four hundred questions drawn from the brand’s own site, its dealer FAQs and the awkward things people actually type, each with an answer the brand would sign. Then we built until the set passed.

  1. Week 1

    400-question test set written and reviewed against the brand’s public specification sheets.

  2. Week 2

    Knowledge base ingested from the public site only. Nothing the brand had not already published.

  3. Weeks 3 to 5

    Built to the test set. Every change re-ran all 400 questions before it merged.

03 · What we built

The pieces, and what each one is for.

A controlled knowledge base ingested from the brand’s public website, with the bot restricted to it and to brand-appropriate language. Tree, free-text and hybrid conversation modes. Off-topic and abusive inputs flagged automatically. Lead capture with dealer routing and a clean handover to a human agent, with the transcript.

  • Controlled knowledge base

    Ingested from the brand’s own site and restricted to it. If it is not published, the assistant does not know it.

  • Three conversation modes

    Tree, free text and hybrid, because a dealer group and a brand site want different levels of control.

  • Lead capture with dealer routing

    Qualified, routed by ZIP code, written into the dealer CRM with the transcript attached.

  • Handover that keeps the thread

    An agent receives the conversation, not a ticket saying a conversation happened.

04 · The result

What changed, and by how much.

  • Off-brand or invented answers, 400-question set Not tracked 0
  • Time to reach a human agent Not tracked under 4 s
  • Questions answered without escalation Not tracked 82%
  • Test set re-run on every change no yes

Internal test set of 400 questions, Unflow logs.

Does this sound a bit like your week?

Tell us how it works at your place today, and we’ll tell you honestly what it would take.

Book a 15-minute call

Not ready to talk yet? Take the build-or-buy check