Plan: pick a mandate and read the plan I would bring
Mandates3
Each mandate is something a sponsor might hand over. Pick one with the arrow keys.
An example plan that shows my method. Not a past client engagement.
VP of Customer Support says
“Use AI to get our support backlog under control.”
Where AI is the wrong tool
Billing disputes with legal exposure stay with people. The agent gathers the facts and stops.
- 01
OutcomeWhose number moves, by how much, by when?
First response time on the highest-volume ticket categories, measured weekly. The sponsor owns the number, and the target is set against the current baseline, not a guess.
- 02
Users and decisionsWho acts on the output, and what does a mistake cost?
Support agents accept, edit or reject a drafted reply. A wrong refund or a wrong policy answer costs money and trust, so those never send without a person.
- 03
Data and systemsWhat exists, where does it live, and who owns it?
Ticket history in the help desk, the policy and help-center articles, and order data behind an internal API. Each has an owner and an access rule.
- 04
LayersWhat ships value on its own?
- VisibilityClassify and route every ticket. Show the queue by category.
- AssistanceDraft replies with citations for an agent to approve.
- AutonomyResolve only the categories that hold the eval gate over time.
- 05
ConstraintsWhat can stop this, said out loud?
PII in tickets, refund authority, the help desk API limits, and a support team that has to trust the drafts.
- 06
First sliceWhat is the smallest thing worth deploying?
Routing plus cited drafts for one high-volume category, with every draft approved by a person.
Tools and gates5 tools
Where AI is the wrong tool
Billing disputes with legal exposure stay with people. The agent gathers the facts and stops.
Agent tools, with scope and approval
search_help_centerautoread: published articlesget_orderautoread: one order, for the ticket's customerdraft_replyautowrite: drafts only, never sendsissue_refundhuman approvalwrite: refunds, cappedescalateautowrite: hands the ticket to a human queue
Stop rule. No cited source, or low confidence: escalate with a summary instead of answering.
Retrieval
Hybrid search over articles chunked by heading, then reranked. Resolved tickets join the corpus only after cleanup.
Internal-only policy never reaches a customer-facing draft.