01 / The problem
What was getting in the way
As a service business grows, the answers sales reps need become more detailed. Scope, policies, customer expectations, and unusual situations are difficult to remember while someone is waiting on the phone.
A folder of documents was too slow to search during a customer conversation. A general AI assistant was faster, but it could fill gaps with a confident answer that was not actually company policy.
A wrong answer from sales can become a promise the operating team is expected to keep. We needed speed without allowing the assistant to invent pricing, policy, or service commitments.
02 / Our approach
The decisions we made first
My roleI defined the business problem, the approved knowledge boundary, and what a useful answer should look like. We built the search, answer, feedback, and review process around those rules.
Use approved guidance only
The assistant searches a controlled knowledge base and answers from the guidance it retrieves, not from general model knowledge.
Say when the answer is missing
If the knowledge base does not contain the answer, the assistant says so instead of creating a policy or guessing what the company would do.
Write for the customer conversation
We ask for a short talk track that a rep can actually say, with a clarifying question when the customer's real concern is still unclear.
Turn weak answers into better guidance
Helpful, missing, incorrect, and overly long answers can be recorded so the underlying company guidance can be reviewed and improved.
03 / What we built
How the solution came together
A structured knowledge base
Each approved entry separates what to say, internal context, what not to say, source information, confidence, and publication status.
Search that handles real questions
The search removes conversational filler and retries with the meaningful parts of a question when an exact search does not find useful guidance.
A bounded answer step
AI turns the retrieved guidance into a concise customer-ready response, but the prompt requires it to stay inside the supplied material.
Feedback and missing-knowledge records
The system records what guidance was used and lets employees flag answers that were helpful, incorrect, incomplete, or too long.
04 / How it works
The process from start to finish
-
01
Ask
A sales rep enters the customer's question in plain language.
-
02
Search
The system finds the closest approved guidance and its limits.
-
03
Answer
AI turns that guidance into a short customer-ready talk track.
-
04
Clarify
The assistant asks a useful follow-up question or admits the answer is missing.
-
05
Improve
Feedback exposes unclear answers and gaps in the approved knowledge base.
AI can turn retrieved guidance into a concise answer. It cannot create policy, pricing, or service commitments. When approved guidance is missing, it must say that it does not know.
05 / The result
What improved
Business value
- We made approved sales guidance faster to find during customer conversations.
- We made missing or unclear company guidance easier to identify.
- We created a more consistent way to answer customer questions without treating AI as the source of company policy.
Safeguards we kept
- The assistant answers only from retrieved, approved guidance.
- It refuses to invent missing policy or pricing.
- Employees must be authenticated to use the system.
- Only authorized owners or administrators can change the knowledge base.
- Questions, source entries, answers, and feedback are recorded for review.
What this demonstrates
Commercial judgment backed by operating depth.
Capabilities
- Sales enablement
- Knowledge management
- Customer communication
- Responsible AI
- Process design
- Feedback systems
Systems used
- Supabase
- PostgreSQL
- AI model
- Knowledge search
- Streaming responses