Nordwind Support — AI Draft-Assistant: Team Guidelines (one-pager)

Scope: every support agent using the AI draft-assistant for customer replies · Owner: Support Team Lead · Review: quarterly

Rule of thumb: the AI drafts, you sign. A reply leaves this team only when a human has run the Verify ladder and owns every claim in it.

---

The 5 rules

1. No policy claim without a KB hit

If the draft cites a clause, guarantee, or entitlement, search the knowledge base first. If the KB has no result, the claim does not go out — replace it with the KB wording for the real policy.

Why (war story): the draft-assistant invented “Rate Protection Guarantee, clause 4.2” in ticket NW-4821 — a clause that exists nowhere in our terms. The customer’s finance team quoted it back at us four days later during a renewal conversation.

2. Dates and amounts come from the ticket, never from the draft

Compute dates from the ticket timeline + SLA yourself; refund amounts are confirmed by Billing (policy OWN-3). The assistant may describe process; it may not commit figures.

Why (war story): in NW-4821 the draft promised “€412 by Friday 15 August” — the settlement was approved Tue 12 Aug with a 5-business-day SLA (Tue 19 Aug), and Billing had never confirmed €412. The promise broke on day 4 and the ticket reopened as an escalation.

3. Every commitment carries its ticket reference

Any reply that promises money, dates, or exceptions must name the ticket ID (and quote/order number where relevant) so Finance and CSMs can trace it.

Why (war story): the NW-4821 draft contained no ticket ID or quote number; when the customer escalated, Billing could not link the promise to quote Q-2291 and spent a day reconstructing the paper trail.

4. Brand voice: one plain apology, no drama

Trim intensifiers (“terribly sorry”), exclamation marks, and stacked apologies. Plain, direct, warm — per the brand voice list.

Why (war story): early pilot drafts opened with “I am terribly sorry for the confusion…!” — reviewers flagged that over-apologizing reads as guilt and inflates the customer’s sense that something is seriously wrong, when usually it is a normal billing question.

5. When in doubt, escalate the claim, not the ticket

If a draft contains a claim you cannot verify in under two minutes (legal wording, pricing exceptions, SLA deviations), route the claim to its owner (Billing / Legal / CSM) and send a process-only reply meanwhile.

Why (war story): before this rule, agents either deleted unverifiable claims silently (customers got vague non-answers) or sent them hoping they were right. NW-4821 showed the cost of “hoping”: one invented clause became a customer-facing commitment we then had to walk back.

---

The Verify ladder (run on every AI draft)

#	Rung	Check	Pass means
1	KB match	Every policy claim searched in the KB	0 results → claim is out, KB wording is in
2	Ticket math	Every date recomputed from ticket timeline + SLA	Date in reply = date from your own math
3	Brand voice	Tone against the brand list	≤1 apology, no intensifiers, no “!”
4	Owner for money	Any amount/exception has a named human owner	Billing/CSM tagged before send

A draft that fails any rung gets fixed at that rung before moving up. Typical ladder time on a billing reply: 40–90 seconds.

Escalation path

1. Agent fixes rung failures in the draft (this is normal work, not a failure).
2. Team lead decides when a claim sits between policies, or the customer pushes back on a corrected reply.
3. Billing / Legal / CSM owns the claim itself — the agent sends a process-only reply and names the owner + expected date.
4. Log it: every invented fact you catch gets tagged #ai-catch in the ticket. The weekly count is our quality signal and feeds the vendor feedback loop.

---

Fictional disclosure: Nordwind Software, Rate Guard, ticket NW-4821 and all persons are a fictionalized composite built from realistic support-desk patterns, created for this portfolio. Guideline drafted with AI assistance; rules, thresholds and final copy are human-owned.