TaskbyTask.Practical AI guides for work
ChatGPT, Claude, Gemini or Copilot · Practical walkthrough

Write a decision memo that doesn't call saved time saved money

You've been asked which of two AI tools the team should try. One frees more hours, the other looks better on paper, and someone will read "saves £190 a month" as a budget line.

Task by Task14 min read + practice
A hand lifts a translucent layer on a decision memo to reveal assumptions, beside two option folders and a neutral current-option card.

What you'll make

a memo of under 450 words with a conditional recommendation, the few calculations that matter, the assumptions that could change the choice and the approval still needed.

What you'll need

ChatGPT, Claude, Gemini or Copilot in your work account, your notes on the options (including doing nothing), the constraints and who decides. About 30 minutes.

Example result.

Conditional recommendation: evaluate Tool B first, subject to the pilot gates. At 180 summaries a month, B releases about 9 staff hours and A about 12. Valued at an assumed £30 an hour, after subscriptions, that’s £190 of capacity for B and £180 for A. This is time released, not cash saved. Approval needed: a pilot of B. Any purchase is a separate decision.

1. Lay out the decision and the real options

Open a new chat. Paste your notes and say who will decide, what the hard constraints are and whether doing nothing is an option. Mark every guess as an estimate, or it will become a fact by the third copy. If your notes include a spreadsheet, ChatGPT can analyse an uploaded file, depending on your plan and workspace settings.

To practise first, use this fictional pack:

Decision: which AI tool to evaluate first for case summaries. Decision-maker: operations lead.
Today: 180 summaries a month, about 12 minutes each.
Tool A: £180 a month. Estimated 7 minutes per summary. Estimated 80% of summaries would use it. Setup about 8 hours.
Tool B: £80 a month. Estimated 9 minutes per summary, used for all of them. Setup about 3 hours.
Subscription cap: £200 a month, excluding VAT.
Assumed value of staff time, for planning only: £30 an hour.
Pilot gate: 20 representative summaries with no critical factual errors after review.
Still needed: accessibility review and security approval. Neither tool tested. No purchase approved.

Why this works: a memo can only be as honest as its inputs. Labelling estimates now stops them turning into evidence later.

What you should see: a short list of options including the status quo, the constraints, and every number marked as known or estimated.

2. Separate facts, estimates and the calculations

Help me examine [decision] for [decision-maker].
Options, including doing nothing: [options].
Must-have constraints: [requirements every option must meet].
My notes, with their sources, are pasted after this message.

Separate source-backed facts, estimates, judgement calls and missing information. Compare the options against the constraints. Show the calculations with units and formulas, and don't invent scores or weights. Keep cash costs apart from any assumed value of staff time, and one-off effort apart from monthly figures. Say which assumptions could reverse the choice and what small observation would settle each. Don't treat an attractive estimate as evidence that a must-have is met.

[decision] and [decision-maker] name the choice and who makes it. [options] lists them, status quo included. [requirements every option must meet] covers caps, quality gates and approvals.

Why this works: the model shows its arithmetic before it writes persuasive sentences, so you can check the numbers while they’re still numbers.

What you should see: four short lists, a few calculations you can redo on a calculator, and at least one assumption flagged as able to change the answer.

3. Draft the memo

Recalculate the key figures yourself, then ask for the memo in the same chat.

Draft a memo of no more than 450 words for [decision-maker] using the checked comparison above. Use these headings: Decision requested; Conditional recommendation; Options and evidence; Assumptions that could change the choice; Pilot gates; Approval needed. Include the status quo fairly. Recommend a reversible next step, not a purchase. Call released staff time "capacity", never savings, profit or ROI. Keep estimates labelled. Don't invent approvals, deadlines or test results. Put the calculations in a short appendix.

Why this works: fixed headings make the approval and the gates impossible to leave out, and naming “capacity” stops the most tempting overstatement.

What you should see: a memo whose first two sections tell the reader what they’re being asked and what’s recommended, with conditions.

Check the result

  • Decision: the next step is clear and reversible, and missing quality or security evidence hasn’t been treated as passed.
  • Maths: adoption and timing assumptions are stated, and monthly costs stay separate from one-off setup.
  • Evidence: facts match the notes, estimates are still labelled, and doing nothing gets a fair comparison.

The usual failure: capacity becomes cash. “Releases 9 staff hours, valued at £190” turns into “saves £190 a month”, and a pilot recommendation turns into “buy B”. Spot it by searching the memo for “save”, “saving”, “ROI” and “buy”.

Revise the memo. [Sentence] presents released staff time as money saved, or a pilot as a purchase. Replace it with the capacity figure, its assumed hourly value and the condition that no cost falls unless the time is redeployed. Keep the recommendation conditional on the pilot gates. Change nothing else.

[Sentence] is the exact line that overstated the result.

Worked example: choosing which tool to pilot

Fictional example. The operations lead uses the pack from step 1. The calculations from step 2:

  • A: 180 × 80% × (12 − 7) ÷ 60 = 12 hours a month. At £30, that’s £360, less £180 subscription = £180.
  • B: 180 × (12 − 9) ÷ 60 = 9 hours. £270 less £80 = £190.
  • Doing nothing: no subscription, no setup, no released time.

The first memo said:

Recommendation: buy Tool B, which saves £190 a month and gives the best ROI of the two.

After the repair prompt, the memo read:

Decision requested: which tool to evaluate first for case summaries. No purchase is requested. Conditional recommendation: evaluate Tool B first, subject to the pilot gates. Options and evidence: the team writes 180 summaries a month at about 12 minutes each. Estimates, not test results: A takes 7 minutes and would be used for 80% of summaries. B takes 9 minutes for all of them. That releases about 12 staff hours a month with A and 9 with B. At an assumed £30 an hour, after subscriptions (A £180, B £80, both under the £200 cap before VAT), that’s £180 of capacity for A and £190 for B. No budget line falls unless those hours are redeployed. Doing nothing costs nothing new and releases no time. Assumptions that could change the choice: above about 200 summaries a month, A’s modelled value overtakes B’s. Setup is 8 hours for A and 3 for B. Higher adoption of A would also narrow the gap. Pilot gates: 20 representative summaries with no critical factual errors after review, an accessibility review and security approval. Neither tool has been tested. Approval needed: a pilot of B. Any purchase is a separate decision.

  • “Saves” became “capacity”. The repaired memo gives the hours, the assumed value and the condition under which money actually moves.
  • The gate survived. The recommendation is to evaluate, and the memo says plainly that neither tool has been tested.
  • The status quo got a line. It’s the option most memos forget, and the one the reader falls back on if the pilot fails.

Try it

Task by Task.Explore the example

Explore the assumption behind the recommendation

Move adoption for option A. Capacity value is not a cash saving.

Explore the assumption behind the recommendationA at 80% adoption: 720 min capacity; £180 net assumed capacity value. B at 100%: 540 min; £190. These are modelled capacity values, not realised cash savings.0106.4212.8GBPA capacity value180B capacity value190
A capacity value£180B capacity value£190

A at 80% adoption: 720 min capacity; £180 net assumed capacity value. B at 100%: 540 min; £190. These are modelled capacity values, not realised cash savings.

Fictional: 180 summaries a month, staff time valued at an assumed £30 an hour. At 80% adoption, A gives £180 of capacity after its subscription and B gives £190. Above about 83% adoption, A comes out ahead. Either way, this is time released, not money saved.

Task by Task · Illustrative example

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Optional training: find the volume where the choice flips

When two options have different fixed and variable costs, work out where they cross. Here, A’s monthly value is 2 × volume − 180 and B’s is 1.5 × volume − 80. They’re equal at 200 summaries. At 240 summaries a month, A gives £300 and B £280, so A comes out ahead, before setup.

Using the same assumptions, calculate the monthly volume at which A and B have equal net capacity value after subscriptions. Then recalculate both at [new volume]. Show formulas. Keep setup effort separate, and say plainly that the crossover is not evidence of quality.

[new volume] is the workload you expect next quarter. A crossover only means something while the timing and adoption estimates hold.

Optional training: ask a sceptical colleague's question

Before the meeting, ask the model to attack its own memo.

Review the memo as a sceptical colleague. Identify the three claims most likely to mislead a reader. For each, quote the sentence, say why it's vulnerable and give a more accurate replacement. Then list the smallest measurements that would replace our most important estimates. Don't add new facts or change the recommendation silently.

Check especially for invented scores, a missing status quo, one-off costs shown as monthly, and “savings” attached to time nobody can actually redeploy.

Optional training: replace the estimates after the pilot

Collect real preparation, review and correction times, adoption, accessibility findings and factual errors. Rerun the comparison with observed figures and date them. Don’t add “validated” to the first memo. Write a new one. Try a two-week AI pilot and count the checking covers counting the time honestly, including checking time.

Sources

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.