AI for Customer Support. Deflection Without Damage.
Deflection is easy to measure and easy to get wrong. The goal is not fewer tickets. It is fewer tickets that a human needed to touch.
7 min read
- First response
- Under 2 min
- from 4-8 hrs
- Correctly deflected
- Tracked
- Unknown
- Escalation quality
- Full context
- Cold handoff
Support automation has a bad reputation and it earned it.
Most deployments optimised for deflection rate, which is trivially easy to game by making it hard to reach a human.
The metric went up. Satisfaction went down.
The version that works measures something harder.
How many tickets were resolved correctly without a human, and how good was the handoff on the rest.
§01 · The problem
Where the time actually goes.
- 01Deflection measured without checking resolution
- 02No clean path to a human when needed
- 03Escalations arrive with no context attached
- 04Knowledge base out of date with the product
- 05Same question answered differently each time
- 06Agents re-reading full threads on every handoff
An automated answer that is wrong costs more than no automated answer at all.
§02 · The mechanism
The AI workflow.
- 01
Ticket arrives on any channel
Trigger - 02
Agent classifies intent and urgency
AI Agent - 03
Answer drafted from verified sources
Generative AI - 04
Low confidence routes to a human
AI Agent - 05
Escalation carries full context
Automation - 06
Resolution logged and fed back
Validation
A person closes the loop: the workflow ends on a human or a check against a defined outcome.
- Trigger
- The event that starts the workflow
- AI Agent
- Model extracts, classifies, matches, or routes
- Generative AI
- Model drafts something a human will edit
- Automation
- Deterministic. No model involved
- Validation
- Checking the result against a defined outcome
§03 · The difference
Before vs after.
| Before | After |
|---|---|
| Bot answers everything, badly | Bot answers only what it can verify |
| Reaching a human takes four attempts | Escalation is one step, always available |
| Agent starts every escalation from zero | Agent receives the full history and attempt |
| Knowledge base updated once a quarter | Knowledge base updated from resolutions |
| Answer quality varies by agent | One verified answer per question |
| Repeat questions never fixed at the source | Repeat questions flagged for a product fix |
§04 · The measure
Time saved.
| Task | Before | With AI | Comparison |
|---|---|---|---|
| First response time | 4-8 hrs | Under 2 min | |
| Tickets needing a human | 100% | 35-45% | |
| Handoff context | None | Complete | |
| Time per escalated ticket | 18 min | 7 min |
Continue reading
Tell us what you’re building.
Share a few details about what you need, and we’ll get back to you with a quotation.