Start From the Decision and Set Safe Ground Rules for AI

Leadership rarely asks for a forecast. They ask a question, such as "Should we open the second site?" A decision brief answers that question. It does not walk through the model.

If you start from the model, you tend to review every input line by line. Fifty inputs get equal attention, and the reader still cannot see which ones matter. If you start from the decision, you have a test for every input: does this change the answer?

So before you open the spreadsheet, write the question down as one decision. Phrase it as a choice that can come out go or no-go:

• Vague: "Review the expansion plan."
• Decision-first: "Should we sign the lease and open the second site in the next financial year?"

If leadership has bundled several choices together, such as which site, when and how big, split them. Pick the one they need answered now. A brief that tries to settle three decisions usually settles none of them.

The rest of this course builds on this sentence. Each later step, from shortlisting inputs to testing them and drafting the brief, refers back to it.

A decision needs a number that answers it. This is the outcome measure: the single result from the model that tells you whether the decision is a good one. Common choices:

• NPV (net present value): future cash flows converted to today's money at a chosen discount rate, minus the investment. A positive figure means the project earns more than that rate.
• Payback period: how long until the cash the project generates equals the money put in.
• Year-two operating margin: operating profit as a percentage of revenue in the second year.

Pick the one that matches how leadership thinks about the decision. Suppose the board cares most about getting its cash back. For the second site, that points to payback.

Example: the site costs £1,800k to open and is forecast to generate £50k of net cash a month. The chart shows cumulative cash: it starts at minus £1,800k and reaches zero at month 36. That crossing point is the payback period.

You can show other measures in an appendix. The brief, though, is built around one. With two headline measures, leadership can get two different answers.

An outcome measure on its own does not say go or no-go. You also need a threshold: the value at which the answer changes. For example: "Go if payback is under 36 months."

The threshold matters for every later step. In lesson 3 you will change one input at a time and see how far the outcome moves. Without a threshold, you can report that payback moved from 36 to 42 months. You cannot say whether that changes the decision. With a threshold, you can: 42 months is over the line, so the answer flips to no-go.

Agree the threshold with whoever owns the decision before you run any tests. If you set it after you have seen the results, people will reasonably suspect you picked it to fit.

Now put the parts together into one sentence:

Should we open the second site? Go if payback is under 36 months.

Use the explorer to see how close the base case sits to the line. At £1,800k and £50k a month, payback is exactly 36 months, with no headroom. A small slip in monthly cash flow would push it over.

A number without its context can be misread. "Payback of 36" could mean months or quarters. "£1,800" could be pounds, thousands or millions. Write down four facts about your outcome measure now, at the start, and carry them into every later step:

• Reporting period: the time frame and the source data it rests on. Example: monthly, from site opening; base figures from the FY2024 budget.
• Currency: GBP.
• Units: thousands, millions or a percentage. Example: £k.
• Calculation method: how the number is worked out. Example: months until cumulative net operating cash flow equals the opening investment; undiscounted; excludes head-office overhead.

The method is the one people most often leave out, and it causes the most arguments. Discounted and undiscounted payback can differ by many months. A margin with or without allocated overhead can give the opposite answer.

Keep these four lines at the top of your working notes. When you later ask an AI tool to draft the brief, give it these lines too, so the draft does not quietly change units or mix periods.

AI tools can speed up this work. They also create a new way for confidential information to leave the business. Two rules apply before anything else.

Rule 1: Use only an AI tool your organisation has approved for internal financial material. A tool approved for drafting marketing copy is not automatically approved for forecast data. If you are unsure whether a tool, or a particular use of it, is approved, check your organisation's policy or ask the person responsible before you paste anything. Text cannot be taken back once it is pasted.

Rule 2: Never put account credentials into any AI tool, approved or not. Credentials include passwords, access tokens and connection strings. A connection string is the line of text a spreadsheet or script uses to log into a database, and it often contains a username and password. These sometimes hide inside model files, in data connections or macro code, so look for them before you share anything taken from the file.

Customer identifiers and nonpublic reporting information, such as unreleased results, must also stay out of unapproved tools.

Even in an approved tool, share only what the task needs. For a decision brief, the AI needs to know which inputs exist, their base values and the ranges you plan to test. It does not need your general ledger, customer lists or login details.

Work through these common items:

• Input names, base values and test ranges: this is what the task needs. Fine in an approved tool.
• Customer names: usually not needed. Replace them with codes, such as "Customer A" or "C-104".
• Unreleased quarterly results: nonpublic reporting information. Keep them out of unapproved tools, and in an approved tool share them only if the task truly needs them.
• Full ledger export: more than the task needs. Give summarised input values instead.
• API access token: a credential. Never into any AI tool.

A useful habit: before you paste, ask yourself whether you would be comfortable if this exact text were forwarded to your data protection lead. If you hesitate, cut it down further.

Now put it into practice. Rather than uploading the model, build a short, safe extract to paste into your approved tool.

Example for the second site:

Decision: Should we open the second site? Go if payback is under 36 months. Period/currency/units/method: Monthly from opening; GBP; £k; undiscounted payback on net operating cash flow, excluding head-office overhead. Inputs: Fit-out cost: base 1,200, range 1,000–1,500. Monthly revenue: base 210, range 170–240. Staff cost per month: base 95, range 85–110. Anchor tenant: Customer C-07.

The extract leaves out the customer's real name, linked data sources, hidden sheets and anything with a password in it. It keeps what the next lessons need: the decision, the context lines and the inputs with their base values and ranges.

Save this extract alongside your working notes. In the next lesson you will use it to shortlist which inputs are worth testing.