Most KPI commentary just restates the chart in prose: “revenue grew 5%”. Leadership already saw the 5%. A good prompt forces the model to skip the number and answer the three questions that actually matter — what changed, why, and what to do about it. The 12 templates below are written to do exactly that, and they paste straight into ChatGPT, Claude, or Gemini.
TL;DR: Feed the model the raw KPI (prior vs current, plus 6-12 periods of history), pin the audience seniority, cap the length, and forbid filler phrases explicitly. For commentary on an uploaded spreadsheet, GPT-5.5 (ChatGPT for Excel went generally available across all plans on May 5, 2026) and Gemini 3.1 Pro are the two strongest picks as of June 2026. Never let the model invent the numbers — paste real data or connect a real source.
Who this is for
- Operators preparing weekly or monthly metrics
- Founders writing the KPI section of an investor update
- BizOps and FP&A analysts adding narrative to dashboards
- Anyone who keeps getting “so what does this mean?” replies to their reports
When not to use these prompts
- You have not verified the numbers yet (the model will confidently narrate a typo).
- Nothing actually moved — flat is fine, but don’t manufacture a story (see template 11).
- The data is confidential and your account isn’t on a plan with data controls. ChatGPT Free, for example, has shown ads in the US since February 2026 and trains on Free/Plus chats unless you opt out; check your workspace settings before pasting sensitive figures.
The six-part prompt skeleton
Every template below carries the same six elements. Borrow the skeleton when you write your own:
- Role — who the model plays: FP&A analyst, chief of staff, the CFO’s deputy.
- Context — team, org, product, and the time window of the data.
- Goal — one deliverable: a table, three talking points, a 150-word section.
- Constraints — word count, must-include fields, audience seniority.
- Tone — neutral and factual for an exec; more explanatory for an IC.
- Anchors — 1-2 examples of past commentary so the format matches your house style.
Best for
- Weekly metrics commentary
- Monthly business review (MBR) narrative
- Investor-update KPI sections
- Anomaly callouts and trend-vs-noise calls
- Trimming a bloated KPI doc
12 copy-ready prompt templates
Placeholders use [brackets]. Swap them for your real values before sending. Paste the surrounding history (last 6-12 periods) whenever the prompt references variance, pace, or a cohort split.
1. What changed, why, what to do
KPI: [kpi] moved from [prior] to [current] ([delta]). History (last 12 periods): [paste].
Write 3-sentence commentary: (1) what changed, factual; (2) the single most likely cause; (3) the action or no-action implied. Do not restate the number. No hedging filler.Variables to swap: kpi, prior, current, delta
2. Trend vs noise
Is this [kpi] move a trend or noise? Data, last 8 periods: [paste].
Output: (1) compare this move to the period-over-period variance; (2) signal vs jitter; (3) verdict — "Trend", "Noise", or "Wait one more period" — with one sentence of reasoning. One paragraph.Variables to swap: kpi
3. Anomaly callout
This week, [kpi] is [N] standard deviations from its rolling mean. Rolling stats: [paste].
Write a 2-sentence callout: (1) magnitude and direction; (2) two plausible explanations worth investigating. No "we need to dig in" filler.Variables to swap: kpi, N
4. Cohort-explained delta
Total [metric] moved [delta]. Breakdown by cohort: [paste new vs existing, segment, region].
Identify which cohort drove the move and quantify it. Output format: "X% of the +Y came from cohort Z because ...". One line per material cohort.Variables to swap: metric, delta
5. Goal-pacing commentary
Quarter target for [kpi] is [target]. Current value is [current] with [percent] of the quarter elapsed.
Commentary: ahead of / on / behind pace, the run-rate needed to land the target, and one adjustment to ask for. No "we're crushing it".Variables to swap: kpi, target, current, percent
6. Multi-KPI dashboard summary
Here are this week's KPIs with prior values: [paste 6 rows].
Write a 5-bullet exec summary: the most important movement, one risk, one tailwind, one anomaly, one ask. 200 words max. Lead each bullet with the takeaway, not the number.Variables to swap: the 6 KPI rows
7. Investor-update KPI section
Write the KPI section of a monthly investor update from this data: [paste].
Cover (1) the headline metric movement; (2) three supporting numbers; (3) what changed in the funnel; (4) what we're doing differently next month. 150 words max, confident but not promotional.8. Cohort retention commentary
Retention curves by signup cohort: [paste week-0 through week-N for each cohort].
Commentary: (a) where the cliff is; (b) whether week-N retention shifted vs the prior cohort; (c) one explanation; (d) one action. Do not restate the table.Variables to swap: the retention table
9. Conversion funnel narration
Funnel, this period vs last: [paste each stage with both periods' conversion].
Identify (1) the biggest drop-off stage; (2) whether it changed this period; (3) the likely root cause; (4) the lowest-effort experiment to validate it. Skip "needs further analysis".Variables to swap: the funnel table
10. KPI doc trim
Our weekly KPI doc has [N] metrics: [paste list with 90-day history].
Audit and trim to 6. Flag (1) which haven't moved meaningfully in 90 days; (2) which are derivatives of others; (3) which we'd never actually act on. Justify each cut in one line.Variables to swap: N, the metric list
11. Narrative for a flat KPI
[kpi] is flat ([value], unchanged for [periods]).
Write 2-sentence commentary that doesn't fake significance: (1) flat, plus whether that's good or bad given context; (2) the one thing that would move it next month. No padding.Variables to swap: kpi, value, periods
12. Forecast vs actual
Forecast: [forecast]. Actual: [actual]. Variance: [diff] ([beat/miss]).
Commentary: (1) what drove the variance; (2) whether the forecast model needs adjusting and how; (3) the implication for next period's forecast.Variables to swap: forecast, actual, diff
A worked example
Input you paste:
KPI: paid signups moved from 412 to 358 (-13.1%). History (last 8 weeks): 401, 388, 420, 433, 412, 419, 412, 358.
Write 3-sentence commentary... [template 1]Weak output (what to reject): “Paid signups fell 13.1% week over week to 358, a notable decline that warrants further investigation.”
Strong output (what template 1 produces with the history attached): “Paid signups broke an 8-week band of 388-433 for the first time, so this is a level shift, not normal jitter. The drop lines up with the checkout-page deploy on Tuesday, which is the first place to look. If that’s confirmed, roll back and the rest of the funnel needs no action.”
The difference is the history. Without the prior periods, the model can’t tell a real break from noise and falls back on hedging.
Which AI model to use for the data
The prompt does the thinking; the model handles the spreadsheet. As of June 2026:
| Tool | Best for | Cost (consumer) | Notes |
|---|---|---|---|
| ChatGPT (GPT-5.5) | Spreadsheet upload + Python analysis | Free $0 / Go $8 / Plus $20 / Pro $100-$200 | ChatGPT for Excel and Google Sheets went GA across all plans on May 5, 2026; runs sandboxed Python (pandas) on uploaded files up to ~50 MB. |
| Gemini 3.1 Pro | Finance / spreadsheet reasoning, dashboards | Google AI Pro $19.99 (was “Gemini Advanced”) | Google specifically tuned 3.1 Pro for finance and spreadsheet usability; 1M-token context handles large workbooks. |
| Claude (Opus 4.7 / Sonnet 4.6) | Long-form narrative, dense chart screenshots | Pro $20 / Max $100-$200 | 1M-token context; accepts high-res images (~3.75 MP) for reading dashboard screenshots; strongest at polished prose. |
Practical default: drop the raw export into ChatGPT or Gemini, let it compute the deltas and variance, then run the commentary template on the result. Claude is the pick when the deliverable is the writing itself (an investor update, a board narrative). Whatever you use, paste or connect the actual data — every one of these models will fabricate plausible figures if you ask it to “estimate” the numbers.
Common mistakes
- No history attached. Variance, pace, and anomaly prompts are guesses without the prior periods.
- Skipping the fact-check. The model narrates whatever you give it, typos included.
- Vague audience. “Write commentary” overshoots for a CEO and undershoots for an IC.
- No word cap. Anything past line five doesn’t get read.
- One template for every situation. Readers tune out a fixed format; match the prompt to the move.
- No decision framing. If there’s no ask, the reader doesn’t know what you want.
- Letting the model invent numbers. Connect a real source; never accept “estimated” figures.
How to sharpen the output
- Name the audience level explicitly: IC, manager, VP, or CEO.
- Cap length up front — one page for tactical, three bullets for executive.
- Lead with the ask or decision; context goes after.
- Attach the source-data link so the reader can verify without a follow-up email.
- Ask for two versions and pick the tighter one; AI is cheaper to over-produce than to under-edit.
- Use the model to draft and self-audit, then have a human skim the numbers before it ships.
FAQ
- How long should KPI commentary be? Match the audience. Tactical reviews: one page. Executive or investor updates: three to five bullets plus a link to the underlying data.
- Can AI replace the analyst? For first drafts, formatting, and variance decomposition, yes. For the judgment call on what to do about a number, no — that stays with a human.
- Which AI is best for spreadsheet-based commentary in 2026? GPT-5.5 (now generally available for Excel and Google Sheets) and Gemini 3.1 Pro, which Google tuned for finance and spreadsheet work. Use Claude Opus 4.7 when the writing quality of the narrative matters most.
- Will the model use my data to train its model? Depends on the plan. ChatGPT trains on Free and Plus chats unless you opt out; business and enterprise tiers exclude your data by default. Check the data-controls setting before pasting sensitive figures.
- Should I include risks in the commentary? Always. Pretending no risk exists erodes trust on the next update, when the risk shows up anyway.
- Can AI generate the underlying data itself? No. It will invent plausible-looking numbers. Paste real exports or connect a verified data source.
Related
- Weekly report prompts
- Monthly business review prompts
- Executive summary prompts
- Excel analysis prompts
- Write the Narrative Around Your KPI Movement With AI
- Write a Weekly KPI Summary With AI: Numbers In, Executive-Ready Summary Out
- Productivity & Office Prompts hub
External references: OpenAI — ChatGPT for Excel and OpenAI Help Center — Data analysis with ChatGPT.
Tags: #Prompt #Productivity #KPI #Commentary