Case Study
AI Chatbot Case Study: Supply SA Vendor Certification Assistant
A production AI chatbot that compresses a 30–40 step certification process into 7–11 screens — with zero hallucinated compliance answers.
“This is AMAZING!”
At a Glance
01
Executive Summary
Supply SA partnered with Soulwire Studio to eliminate a staffing bottleneck in vendor certification. Each applicant previously required a 30–45 minute live walkthrough with a specialist just to learn which documents applied to their situation. At 10–15 vendors weekly, that consumed roughly 34 hours of specialist time every month.
The solution: a production AI chatbot that compresses the entire intake process into 7–11 adaptive screens with zero improvised answers. Routine walkthroughs now run unattended and correctly.
02
The Problem
Vendor certification requirements at Supply SA vary by application type, certifications sought, entity structure, industry, jurisdiction, and ownership structure. No two vendors are identical. The result: every first-time vendor needed dedicated staff time just to learn what applied to them.
- Every new vendor required 30–45 minutes of live specialist time before any real work began
- An earlier prototype built in Voiceflow hit credit-cost constraints and couldn't scale
- LLM hallucination posed a real risk — incorrect guidance on government certification requirements could expose vendors to compliance failures
03
The Vision
Replicate a trained Certification Navigator's judgment on-demand — at zero marginal cost per conversation — with the reliability and accuracy that compliance-critical guidance demands.
- Exact document computation with zero improvisation
- Explanations grounded in real internal guidance — not model inference
- Clear distinction between certification and contracting processes
- Clean escalation to staff for genuine judgment calls
- Continuous production reliability with no drift over time
04
The Solution
The core architectural insight: separate what AI is good at from what it isn't. A deterministic JavaScript rules engine computes which documents apply. Claude narrates results conversationally — without inventing anything it wasn't given.
The knowledge layer was sourced from a 93-rule validation workbook, the Certification Navigation SOP, and Supply SA's FAQ. Every answer the chatbot gives traces back to a known document.
Quality Safeguards
- Hallucination caught during testing — architecture rebuilt before any vendor exposure
- Substring-matching bug identified and resolved before launch
- Proper scoping between certification vs. contracting enforced at the logic layer
- Low-confidence and eligibility judgment calls route to email support automatically
05
How It Works
Vendors answer a series of profile questions — application type, certification sought, entity structure, industry, jurisdiction, ownership structure. The rules engine branches based on answers, surfacing only the 7–11 document screens that apply to that specific vendor.
Free-text “why” questions match against grounded knowledge before the model responds. When a question falls outside the chatbot's scope — eligibility decisions, complex edge cases, or explicit staff requests — it escalates cleanly to email support and the team page.
06
Measurable Results
07
Business Impact
Routine walkthroughs now run unattended — and correctly. Specialists are free for eligibility judgments, complex cases, and relationship-building: the work that actually requires a human.
The certification process no longer scales linearly with headcount. Whether 5 vendors or 50 submit in a week, every one receives identical, accurate answers at any hour. The chatbot doesn't get tired, doesn't improvise, and doesn't miss edge cases the rules engine was built to catch.
08
Why This Matters
AI earns trust through constraint, not cleverness. The compliance-critical logic in this system runs as auditable code with no model inference involved. Natural language explanations leverage AI for what it does well. The result is a system that's both trustworthy and genuinely useful.
For organizations with compliance-critical processes — government agencies, nonprofits, healthcare systems — this architecture is the pattern: rules engines for correctness, language models for communication.
Stack

