AI · Saudi Arabia · Six-week vertical slice
A knowledge assistant staff would actually trust
Professional services firm
They asked for an AI assistant. They got retrieval over maintained SOPs, citations, and a CRM write-back — with refusal when the corpus had no answer.
Challenge
A generic chatbot was already live. It invented policy and pricing. Containment rate looked good; accuracy was unmeasured.
SOPs lived in a shared drive with no owner, no kill date, and two conflicting pricing sheets.
Approach
Named a document owner before choosing a vector database. Canonical answers for the ten questions already asked in Slack.
Chunked on headings. Metadata for locale and service line. Evaluations on thirty real questions the operations lead wrote.
High-risk answers required a citation. No citation, no send — open a ticket instead. CRM note on every resolved thread.
Result
Staff used it because they could see the source paragraph. That is a different product from a personality widget.
The first slice stayed one channel, one corpus, one CRM write. Expansion waited until the eval set stopped regressing.
Some “AI work” was actually knowledge-base hygiene. That was treated as a successful discovery, not a failed chatbot.
Composite of an AutoFy Tech AI engagement. We will not publish a hallucinated accuracy percentage.
All selected work