AI sales assistant
An assistant that answers product and stock questions from the catalogue, so the sales team spends time selling.
Situation and baseline
A specialist distributor manages thousands of components whose compatibility notes are scattered across PDFs, a catalogue and CRM. Staff spend hours finding specifications; buyers wait and quotations need manual checking.
The challenge
Improve retrieval without letting an agent guess compatibility, stock, lead time or commercial terms controlled by source systems.
Approach and timeline
Diagnose
Inventory catalogue, technical documents and CRM; mark safe questions and cases requiring an engineer or salesperson.
Design
Normalize identifiers, synonyms and compatibility links, preserve provenance and assign owners for source freshness.
Integrate
Build retrieval-first answers with citations, clarifying questions for missing identifiers and human hand-off on uncertainty.
Learn and refine
Shadow a bounded set of real queries, compare with expert review and enable CRM actions only after approval.
Agents, automations and stack
A catalogue agent retrieves exact product records and drafts sourced answers or quote checks; workflow automation logs requests and assigns an owner. Pricing and inventory remain read-only. The stack combines product data, a document index, access-controlled models and analytics.
Deliverables
- Source and permission map
- Normalized product index, cited-answer flow and quote review
Outcome indicators
- First useful response
Baseline: 6 h Target: 2 min
- Queries resolved with sources
Baseline: — Target: 64%
- Quotation correction rate
Baseline: 7.5% Target: 1.8%
- Sales time released
Baseline: — Target: 310 h
How we measure
Response time runs from a complete enquiry to the first expert-approved answer. Self-service resolution excludes corrections and staff intervention. Quote errors are audited per issued quote; released hours are sampled against observed handling time, not headcount savings.
Risks and guardrails
Citations must resolve to current records. Abstain on ambiguous compatibility; never invent stock, certifications, delivery or price. Log access, minimize customer data and retain a human quote approver.
What comes next
Expand one product family at a time, review multilingual terminology with experts and check that faster drafts do not increase corrections or returns.