A support bot that stops leaking prices and dead links
We audited and hardened a public support bot built by the client's team: source cleanup, pricing guardrails, a regression set, and a weekly audit.
- Chatbase
The problem
The public support bot was answering. It was also quoting dealer pricing to the public and sending customers to product pages that no longer existed, and nobody was reading the conversations.
What we built
We did not build the bot. We cut its crawled sources to a fraction of their size, added a structured product reference and fault-code table, wrote pricing and confidentiality guardrails, built a 174-question regression set, and set up a weekly audit that classifies every conversation and checks every link.
The bot answers customers from the hardened source set. In the weekly audit, a model classifies each conversation by topic and outcome so a human reviews a ranked list, not every transcript.
Deterministic pricing and confidentiality checks on the bot's output, a regression set that asks the same questions every week and diffs the answers, a link checker, and scheduled source sync.
Outcome
Dead product links fell from 77 percent to 3 percent of answers in three weeks and pricing leaks went to zero. The link rate regressed later when sources changed, which is why the weekly audit exists.
Who this fits
- Companies with a support bot someone built fast and nobody audits
- Dealer and channel businesses where pricing is confidential by tier
- Teams that want a bot they can leave public