Decision Desk · Issue 03 · complete public learning edition
Should AI Answer Your Customers?
Decide which support functions and message classes—if any—AI may classify, retrieve, draft, send, or act on, and where human authority must remain.
Learning time75–90 minutes
Teaching10 cumulative lessons
InstrumentAuthority Map
Operating ruleFail closed by class
Authority ladder
Classify → retrieve → draft → send → act.
Every step increases consequence. Permission to assist at one rung does not imply permission for the next.
Authority
Retrieve approved information for review
Draft a proposed response
Send language to the customer
Change money, access, status, or records
Complete issue curriculum
Ten cumulative lessons.
Use the fictional message classes; never paste a real customer interaction.
01Issue 03Decompose support automation
Open
Decompose support automation
Classification, retrieval, drafting, sending, and account or financial action are different authorities. Evaluate each separately.
02Issue 03Classify the consequence
Open
Classify the consequence
An ordinary FAQ, billing dispute, accessibility request, safety complaint, suspected fraud, and account change do not share the same stakes.
03Issue 03Authenticate before revealing or acting
Open
Authenticate before revealing or acting
Identity and authorization requirements change with the requested information or action. Convenience cannot lower the authentication boundary.
04Issue 03Retrieve from an authoritative source
Open
Retrieve from an authoritative source
The answer needs an owned, current, permitted source. Missing, conflicting, or stale authority requires abstention or human escalation.
05Issue 03Control representations
Open
Control representations
Generated language can make promises, admissions, comparisons, refund commitments, or regulated statements. Assign approval and prohibit unsupported claims.
06Issue 03Recognize material escalation
Open
Recognize material escalation
Complaint, harm, fraud, vulnerability, accommodation, safety, and legal-threat signals must reach a capable human without forcing the customer to start over.
07Issue 03Test unequal failure
Open
Test unequal failure
Evaluate language, disability access, ambiguity, nonstandard phrasing, urgency, and customer vulnerability—not only clean average cases.
08Issue 03Limit send and action authority
Open
Limit send and action authority
A draft is not a send; a send is not an account change. Define confirmation, reversibility, monetary thresholds, separation of duties, and stop authority.
09Issue 03Measure what customers experience
Open
Measure what customers experience
Track escalation recall, repeat contact, corrections, complaints, abandonment, first-contact resolution, harmful false closure, and unequal failure.
10Issue 03Issue a message-class disposition
Open
Issue a message-class disposition
Prohibit, redesign, human-lead, or evaluate a bounded assistive pilot. Do not authorize ‘AI customer support’ as one undifferentiated system.
Critical stop conditions
Do not automate through these gaps.
- material complaint or harm is not recognized
- authentication is weak for the information or action
- the response makes an unsupported representation
- an accessible or language-appropriate alternative is missing
- a financial or account action is irreversible
- the human handoff forces the customer to repeat the problem
- correction, logging, incident response, or rollback is absent
Interactive instrument · device-local
Support Automation Authority Map
Decide authority function by function and message class by message class. Do not enter tickets, names, account details, credentials, disputes, or confidential facts.
0/30authority cells decided
No blanket “AI support” permission.
Failure to recognize a complaint, safety issue, fraud, accommodation, vulnerability, or escalation keeps that message class human-led.
Failure to recognize a complaint, safety issue, fraud, accommodation, vulnerability, or escalation keeps that message class human-led.
| Message class | Classify | Retrieve | Draft | Send | Act |
|---|---|---|---|---|---|
| Ordinary FAQ | |||||
| Billing dispute | |||||
| Accessibility request | |||||
| Safety complaint | |||||
| Suspected fraud | |||||
| Account change |
Restoring your local map…
Completion boundary
The educational product and reusable instrument are complete.
A real deployment still requires current sector and jurisdiction research, exact-product evidence, representative testing, accessibility review, incident operations, and organizational approval.