A Support Assistant That Held Up in Production

Summary

A regional home-services company was drowning in repeat calls — scheduling, pricing, service-area questions — while its answers lived in a help center nobody updated and ticket notes nobody read. We built a support assistant grounded strictly in their own documents, scoped to three weeks, with the client's team owning it from day one.

The Challenge

The Problem

The support team answered the same forty questions all day while complex jobs waited:

• Repeat volume: Routine questions consumed most of the phone queue during peak season.
• Scattered answers: The help center, the price sheet, and the dispatchers disagreed with each other.
• No room to experiment: Leadership wanted AI help but refused a six-month platform project — rightly.

Root Cause Analysis

The problem was never staffing alone. Answers existed but were spread across three systems with no owner, so every channel gave a slightly different one. An assistant could only help once the underlying answers were fixed and owned.

Our Approach

Phase 1: Fix the Answers First (Week 1)

• Audited the forty most-asked questions against the help center, price sheet, and dispatcher practice.
• Rewrote and versioned the correct answers with a named owner for each.
• Deleted or archived everything contradicted or outdated.

Phase 2: Build Small and Grounded (Weeks 2–3)

• Built the assistant to answer only from the versioned answer set, saying "I don't know — let me connect you" otherwise.
• Ran it with two agents on live calls before widening to the team.
• Tracked every escalation weekly and fixed the underlying answer, not the model.

Phase 3: Hand Over Fully (Week 4)

• Moved hosting, prompts, and analytics into the client's accounts.
• Trained two staff members to update answers and review escalations.
• Left a one-page runbook instead of a dependency.

Key Achievements

Quantitative Results

• Roughly a third of routine questions resolved without reaching an agent.
• Satisfaction steady through the transition quarter — no dip during rollout.
• Four weeks from kickoff to client-owned operation.

Qualitative Benefits

• Agents spend their days on complex jobs instead of repeating opening hours.
• Every answer now has an owner and a version — useful far beyond the assistant.
• The client team runs the system without us.

Lessons Learned

1. Answers before assistant: Grounding content must be fixed first; the model is the easy part.
2. "I don't know" is a feature: A graceful handoff beats a confident wrong answer every time.
3. Escalations are the roadmap: Each one points at an answer to fix, not a parameter to tune.
4. Handover is the deliverable: If they cannot run it without us, we did not finish.

Conclusion

This engagement worked because it stayed small until it proved itself: one team, one answer set, four weeks. The assistant earns its place daily now — and the client could fire us tomorrow without losing it. That is how we think every engagement should end.

---

Client: A regional home-services company Duration: 4 weeks Team Size: 2 consultants Outcome: Client-owned assistant resolving a third of routine questions