Turn Service Data into a Verified Monthly Narrative

Tier: Foundation · Workflow area: Communication & leadership · ~25 minutes

This playbook is UC-53: the monthly (or weekly) service report, drafted by AI from data you computed, never from data the AI computed. The model writes prose and proposes explanations; the export remains the single source of every digit.

When to use

• Any recurring service reportmonthly SLA pack, weekly ops summary, QBR input. Recurring structure plus fresh numbers is the ideal AI drafting shape.
• Your reports are numbers without a storya table of SLAs tells leadership nothing about why P2 response slipped or what you are doing about it. Narrative is the job; AI removes its cost.
• Different team members write wildly different reportsa pinned-numbers prompt standardizes voice and structure across whoever owns the report this month.

When NOT to use

• Letting the model do arithmeticnever paste raw ticket rows and ask for the SLA percentage. LLMs miscount with total confidence; your ITSM reporting or a spreadsheet computes, the model narrates.
• When the anomaly is unexplainedif you don't know why resolution time spiked, the report says "under investigation," not a fluent guess. AI-generated causes read exactly like real ones; that is the danger.
• Sensitive escalation reportsa report naming an employee's performance or an active security matter is written by a human with HR/security in the loop.

The recipe

1. 1.

   Compute the numbers in your reporting tool, then pin them. Export from ITSM reporting (or Freddy AI Insights, or your pb-44 workbook) and assemble one plain block — this is the only data the model ever sees:

   Copy

   PINNED NUMBERS — July 2026 (source: SN report REP0021, pulled 01-Aug)
   Volume: 1,412 incidents (-6% MoM), 486 requests (+11% MoM)
   SLA: P1 response 100% (4/4) · P2 response 92.1% (target 95) ·
        resolution overall 94.8% (target 90)
   Top categories: Access 31%, Software 24%, Network 17%
   Bot strict deflection: 19.4% (pb-44 workbook) · BSAT 4.1/5 @ 22% response
   Anomalies: P2 response miss driven by 12 tickets during 14-16 Jul
   Freshdesk-to-SN migration freeze; aging backlog down from 41 to 28
   

   Every figure carries its source. If a number isn't in the block, it cannot appear in the report.

2. 2.

   Draft with a role-and-rules prompt.

   Copy

   You are drafting a monthly IT service report for non-technical
   leadership. Use ONLY the pinned numbers below — never compute,
   extrapolate, or invent a figure, comparison, or cause.
   
   Structure:
   1. Headline (2 sentences): overall service health this month
   2. What changed and why — each anomaly gets: the number, the
      explanation I provided, and the action we are taking
   3. SLA table (reproduce my figures exactly)
   4. Watch items for next month (only from my notes)
   
   Tone: factual, no praise inflation, no "delighted to share."
   Length: under 400 words. Flag any gap where you would need a
   number I did not provide by writing [MISSING: what].
   
   Pinned numbers: [paste block]
   My anomaly explanations and planned actions: [paste notes]
   

   The [MISSING:] instruction converts hallucination pressure into a visible to-do list.

3. 3.

   Run the figure-verification pass — every digit, every claim. Print the draft, circle each number and each causal claim, and tick it against the pinned block. Three classic failures to hunt: a derived number the model helpfully computed ("that's a 3-point improvement"), a softened miss ("slightly below target" for a 92.1 against 95), and an invented trend ("continuing last quarter's improvement"). Anything circled and unticked is deleted or recomputed by you.

4. 4.

   Make the anomaly section earn its place. The report's value is concentrated here. Formula per anomaly: the number, the verified cause, the action with an owner and date. "P2 response 92.1% vs 95: twelve tickets breached during the 14–16 Jul migration freeze; freeze-window staffing added to the migration runbook (owner: R. Iyer, before next window)." If you cannot fill all three slots, say what is known and when the rest will be.

5. 5.

   Add the deflection line with pb-44's number, never the vendor's. One sentence, strict definition, with trend: leadership will anchor on whatever figure appears here, so make it the one you can defend.

6. 6.

   Send it as text, not attachment. Paste the narrative into the email or channel with the SLA table inline; attach the full dashboard for the two people who click. Same layout every month — familiarity is what turns a report into a heartbeat.

7. 7.

   Keep the prompt and last month's block beside the template. Next month is a 15-minute job: refresh the numbers, rewrite the anomaly notes, regenerate, verify. Rotate report duty freely; the prompt carries the standard.

Make it yours

• ServiceNow shopschedule the source report to email you the export on the 1st; Performance Analytics gives MoM deltas so you never let the model derive them. Reference the report sys_id in the pinned block for auditability.
• Freshservice shoppull Analytics exports and Freddy AI Insights summaries as input to the pinned block, treating Freddy's generated insights as claims to verify against the export, not as pre-verified prose.
• MSP shopone prompt, per-client pinned blocks, per-client tone parameters; add a contractual-SLA table row and a "credits/penalties" line, and have the account manager do the verification pass — the person who signs the QBR checks the digits.

Quality check

• Every number in the draft exists in the pinned block. The verification pass is mechanical and non-skippable — treat it like pb-01's sampling audit: ten minutes that buy the report its authority.
• Misses are stated as misses. "92.1% against 95" appears with the gap named. Softened language found in the draft means the prompt's tone rules get tightened, not just the sentence.
• Causes are verified or absent. Each causal claim traces to your notes; "under investigation" is a complete and respectable explanation.
• The [MISSING:] flags reach zero before send. Each flag is either filled with a real number or the claim it supported is cut.