The Case Study Problem

Every AI support vendor has a hero metric. "Up to 86% resolution rate" (Intercom, from top-performing deployments; their average across 36M+ resolved conversations is closer to 51%). "80% automation rate" (Zendesk, selected deployments). These peak numbers are real. They're also the ceiling, not the average.

Case studies feature the best implementations at the largest companies with the most mature deployments. Publishing your median customer's deflection rate doesn't make good marketing.

What Average Looks Like

Based on industry data, independent surveys, and conversations with support teams:

Month 1: 15-25% for most teams. The AI handles the obvious stuff (FAQ, hours, basic account questions). Most messages still need humans because the AI isn't configured for them yet.

Month 3: 25-40%. You've added more automation rules, trained the AI on more scenarios, and fixed the biggest gaps. This is where most teams start feeling the time savings.

Month 6: 35-55%. The system is mature. Most easy wins are captured. Improvement slows down because the remaining tickets genuinely need human judgment.

Month 12+: 45-65% for teams that actively maintain and improve. This is where most teams plateau.

By Business Type

E-commerce with simple products: 45-65%. Order status, returns, and shipping questions are highly repetitive and automatable. A Shopify store selling t-shirts will deflect more than one selling custom furniture.

SaaS with simple product: 30-50%. Mix of technical and billing. How-to questions automate well. Bugs and integration issues don't.

SaaS with complex product: 15-30%. API troubleshooting, multi-step workflows, integration debugging. These require investigation that AI can't do.

Service businesses (salons, restaurants, repair): 45-65%. Booking, pricing, and hours are the bulk of inquiries. Highly repetitive.

Healthcare: 10-20%. Regulation limits what AI can say. Most interactions need human verification.

Financial services: 15-25%. Compliance requirements, risk of wrong information, identity verification needs.

Why Your Number Is Lower

You have fewer than 100 knowledge base articles. Vendor case studies often feature companies with 500+ articles. More content means more questions the AI can answer.

Your product is complex. A banking app generates more complex questions than a meal delivery service. Complexity resists automation.

You just started. The vendor's 80% number comes from an 18-month-old deployment that's been continuously tuned. Your 2-week-old deployment hasn't had time to mature.

You're measuring honestly. "Deflection" has different definitions. Some vendors count any conversation where the AI responded and the customer didn't re-contact within 72 hours. Others only count conversations where the customer explicitly confirmed their issue was resolved. The second definition gives lower (and more honest) numbers.

Setting Goals

Quarter 1: aim for 20-30% deflection. Focus on getting the top 5 ticket categories automated correctly.

Quarter 2: aim for 30-45%. Expand to the next tier of automatable categories. Fix accuracy issues from Q1.

Quarter 3+: aim for 40-55%. Fine-tune, add edge case handling, and optimize confidence thresholds.

Don't target 80%. For most businesses, pushing past 60% means either automating things that shouldn't be automated (risking quality) or having an unusually simple support profile.

A 45% deflection rate with 95% accuracy on deflected tickets is a great outcome. It means nearly half your support volume is handled instantly and correctly, and the other half gets the human attention it needs.