How to Write a One-Page AI Business Case for a Small Business

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write a One-Page AI Business Case for a Small Business.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write a One-Page AI Business Case for a Small Business.

Fit it on one page under six headings: the problem in numbers, the change you propose, the one-off and monthly costs including staff time, the benefit with its assumptions shown, the risks with a stop rule, and the decision you're asking for. Use your own measurements rather than vendor claims, and state the payback in months.

The one-page limit does real work. If the case can't be made on one page, the project is usually too big or too vague, and the fix is to shrink it to a first phase that fits. Your readers, whether a co-owner, trustees or a bank manager, will give it about five minutes, so every line has to answer a question they would otherwise ask you in the meeting.

Follow me on Instagram@sagnikteaches

The six boxes on the page, and what goes in each

Before the stages, here's the shape you're aiming for. Each box has a job and a rough length, and the lengths add up to a page at a normal font size.

Connect on LinkedInSagnik Bhattacharya
BoxIts jobLength
1. ProblemWhat happens today, measured, and what it costs2 to 3 sentences with numbers
2. Proposed changeWhat the tool does, who checks its work, what stays the same2 sentences
3. CostsOne-off money, monthly money, staff hours, a contingency lineA 4-line table
4. BenefitWhat improves, by how much, and the assumptions behind it3 lines, assumptions in brackets
5. Risks and stop ruleThe three biggest risks, the control for each, and when you'd stop3 bullets and 1 sentence
6. Decision neededWhat you want approved, by when, and the review date1 to 2 sentences

Everything else, such as vendor comparisons, screenshots and the detail of your time log, goes in an appendix that nobody has to read to make the decision. Allow four to five hours in total across the stages below, most of it spent measuring.

Subscribe on YouTube@codingliquids

Stage 1: measure the problem for two weeks

Time: ten minutes a day for two weeks, then half an hour to add it up.

A business case built on "it feels like it takes ages" loses to the first sceptical question. Pick one process and log it: how many times it happens, how long each takes, who does it, and what goes wrong. A shared spreadsheet with four columns (date, item, minutes, problems) is enough. Two weeks catches the ordinary variation without dragging on.

Compare two versions of the same problem statement:

  • Before: "Our fundraising admin is overwhelmed, and AI could really help us."
  • After: "Thank-you letters for gifts take a median of nine days to go out. Over two weeks the fundraising officer spent 11 hours 40 minutes on them, about six hours a week, and three donors asked whether their gift had arrived."

The second version gives the reader a cost (six hours a week), a symptom they care about (slow thanks) and evidence that it matters to the people who fund you. It also makes no mention of AI, which is right: the problem box should stand on its own, so that the reader agrees there's a problem before hearing your solution.

Stage 2: describe the change in two sentences

Time: fifteen minutes.

Name the tool, say what it does, and say who checks its output. Two sentences force you to decide the scope, and they stop the page turning into a tour of everything AI can do.

A weak version: "We will implement AI to transform our supporter communications." A usable one: "The fundraising officer will draft each thank-you letter with an AI assistant set up with our letter guide and approved paragraphs, then check and personalise it before sending. Donor names, amounts and dates will come from the donor database mail merge, never typed by the AI." The second tells a trustee exactly where the human stays involved and where the data comes from, which answers two of their likely questions before they ask them.

Stage 3: cost it, including the hours nobody invoices

Time: about an hour.

Split the costs into four lines: one-off money, monthly money, staff hours (both setup and ongoing checking), and a contingency. Use list prices from the vendor's pricing page, not a first-year promotion, and note the billing basis (monthly or annual, per seat, minimum seats).

Three details trip up small organisations:

  • Seat minimums. Claude Team and ChatGPT Business both require at least two seats, so a one-person team pays for two or uses an individual plan instead.
  • Discounts you may qualify for. Claude for Nonprofits prices Team at $8 a user a month, and OpenAI for Nonprofits prices ChatGPT Business seats at $8 a user a month on annual billing ($10 monthly). Eligibility rules vary, so check before you put the figure on the page; AI tool discounts for charities and non-profits lists what to claim.
  • Checking time. Every AI output someone reviews costs minutes. Put the review time in the ongoing cost, and the benefit line will be honest.

Add a contingency of about 15% on the one-off costs, and say what it's for (usually extra setup time). A case with no contingency looks as if the writer hasn't done this before.

Stage 4: state the benefit with its assumptions showing

Time: about an hour.

The benefit box is where most cases lose credibility, because the writer states the best outcome as a fact. Show the arithmetic and put the assumptions in brackets, so a reader can disagree with an assumption without dismissing the whole page.

Count only benefits that turn into money or into time with a named use. If the fundraising officer saves four hours a week, write down where those hours go: grant applications, which bring in money, or fewer paid hours. If you can't name a use, the saving is real but soft, and it belongs in a separate "unquantified" line. The same goes for benefits you hope for but can't measure yet, such as better donor retention from faster thanks. List them, but don't count them.

Then give the payback in months: one-off costs divided by the monthly net benefit. If you want a yardstick for whether the figure is good, how soon AI should pay for itself sets out payback targets by project type.

Stage 5: risks, a stop rule and a review date

Time: thirty minutes.

Pick the three risks a reader would raise and pair each with a control. For most small AI projects, the three are data (where information goes and who can see it), accuracy (what happens when the AI gets something wrong) and dependency (what happens if the tool changes, rises in price or disappears). If you already keep a risk register, take them from there; if not, a simple AI risk register gives you a format to start one.

Then write the stop rule: a measurable condition, checked on a named date, under which you'd end the project. "We will stop if, after eight weeks, the median time per letter hasn't fallen below 10 minutes, or if any letter goes out with a wrong name or amount." A stop rule turns a request for money into a controlled experiment, and it's the single line most likely to get a cautious board to say yes.

A realistic mistake shows why. In this invented but familiar scene, a charity manager's first draft had a solid problem, cost and benefit, and no stop rule. The first trustee question in the meeting was "What if it doesn't work?", and the manager's answer ("we'll review it") sounded like no answer at all. The decision was deferred to the next quarterly meeting. The second draft added one sentence of stop rule and was approved in ten minutes.

A finished one-page case from a small charity

Here is the whole page, filled in. The charity and its figures are invented; the format is the part to copy. The charity has four staff and about 30 volunteers, and its fundraising officer handles thank-you letters, regular-giver updates and grant reporting.

SectionContent
ProblemThank-you letters take a median of nine days to go out. A two-week log showed the fundraising officer spending about six hours a week on them (11h 40m over 10 working days). Three donors asked whether their gift had arrived.
Proposed changeDraft each letter with Claude (Team plan) set up with our letter guide and approved paragraphs; the officer checks and personalises every letter. Names, amounts and dates come from the donor database mail merge, never from the AI.
CostsOne-off: 12 hours of setup ($228 at $19/hour) plus 15% contingency ($34). Monthly: two Team seats at the nonprofit price of $8 each ($16). Ongoing checking time is included in the benefit sum below.
BenefitTime per letter falls from about 12 minutes to about 5 including checking (to be confirmed in the trial). At roughly 30 letters a week, that's about 3.5 hours a week, or 15 hours a month, worth $285 at $19/hour. The freed hours go to grant applications. Payback: $262 one-off divided by $269 monthly net, so about one month. Unquantified: faster thanks may help donor retention; we'll track it but haven't counted it.
Risks and controlsData: Team plan, which doesn't train on our content by default; no bank details or health information in prompts. Accuracy: mail merge supplies every name and figure; the officer reads each letter. Dependency: the letter guide and approved paragraphs are kept in our shared drive, so we could move to another tool within a day.
Stop ruleStop if, after eight weeks, median time per letter is still above 10 minutes, or if any letter is sent with a wrong name or amount.
Decision neededApprove $262 of setup time and $16 a month for an eight-week trial starting next month. Review at the trustee meeting after the trial, with the time log and letter samples.

Look at what the page leaves out. There's no description of how large language models work, no list of other AI tools considered, and no claim about what "charities like us" achieve. Each of those would take space from the six boxes and invite a debate the decision doesn't need. They can go in the appendix if a trustee asks.

The arithmetic also holds up under a quick check. 30 letters a week at seven minutes saved is 210 minutes, or 3.5 hours. At 4.33 weeks a month that's about 15 hours, and 15 times $19 is $285. The monthly net is $285 minus $16, or $269. A reader who redoes the sum gets the same answer, which is worth more than any adjective.

Drafting the page with AI without letting it invent figures

An AI assistant is good at turning rough notes into tidy boxes. It's also inclined to fill gaps with plausible numbers, which is fatal in a business case. Constrain it:

Turn my notes below into a one-page business case with exactly these
headings: Problem, Proposed change, Costs, Benefit, Risks and controls,
Stop rule, Decision needed.
Rules:
- Use only numbers that appear in my notes. Do not add statistics,
  research findings, percentages or benchmarks from anywhere else.
- If a heading needs a number I haven't given, write [NEEDED: what].
- Keep the whole thing under 350 words, plain language, no adjectives
  like "transformative" or "significant".
My notes:
[paste your time log totals, prices, and the change you propose]

A first result might include this (illustrative):

Benefit: The fundraising officer will save around 3.5 hours a week, which can be redirected to grant applications. Sector research suggests faster donor acknowledgement can improve retention by up to 20%, representing significant additional income. [NEEDED: current regular-giver retention rate]

Two parts are good: the 3.5 hours comes from the notes, and the [NEEDED] flag is exactly what the rule asked for. The middle sentence is the problem. The "sector research" and the "20%" came from nowhere; the assistant ignored its own rule once, which is common. Delete it, and check every number on the page against your notes before anyone else reads it. How to catch made-up figures in AI-drafted proposals has a fuller routine for that check.

A second pass worth running: paste the finished page back and ask the assistant, "What are the three questions a sceptical trustee would ask about this page, and does the page answer them?" It won't always be right, but it often spots a missing cost line or an assumption you've stated as fact.

Adjusting the same page for trustees, a co-owner or a lender

The six boxes stay the same; the emphasis moves with the reader.

  • Trustees care most about risk, reputation and whether the charity stays in control. Give the risks box more room, name who's accountable, and make the stop rule specific. If the tool touches supporters' personal information, say how you've checked the vendor's data terms, and bring in your data-protection adviser if you aren't sure.
  • A co-owner or business partner usually wants cash and time. Lead the benefit box with the monthly net figure and the payback, and show what happens to the saved hours. A partner who does the work being automated will also want to know how it changes their day, so describe that in the proposed-change box.
  • A lender or funder reads the page as evidence that you manage money carefully. Show the costs in the same format as your accounts, keep the contingency line, and connect the benefit to something they already track, such as capacity to deliver more of the funded work.

Whoever reads it, the people whose work will change should see the page before the decision is made. A case approved over the head of the person doing the job tends to meet quiet resistance in week two; how to get staff buy-in when you introduce AI covers that conversation. In the charity example, the fundraising officer wrote half the page, which is the easiest way to get it right.

When one page starts turning into three

If the page keeps overflowing, look at why before shrinking the font. Three causes cover most cases:

  • Several processes in one case. "Use AI for letters, grant reports and the newsletter" is three projects. Make the case for the one with the best measurements and mention the others as a possible second phase.
  • A supplier's proposal pasted in. The supplier's pages belong in the appendix. Summarise their price and scope in your costs box, and judge the proposal separately first; how to evaluate an AI implementation proposal covers what to look for.
  • Uncertainty you haven't resolved. If you can't state the costs because you don't know which tool you'd use, the business case is premature. Ask for a small, fixed budget to run a trial instead, and write the full case with real numbers afterwards.

That last option is often the best one. A one-page request for an eight-week trial, with a clear stop rule, is easier to approve and far easier to write honestly than a case for a year-long rollout nobody has tested.

Further reads

Sources: Claude and ChatGPT nonprofit pricing pages (Claude for Nonprofits Team; OpenAI for Nonprofits); Anthropic and OpenAI notes on business-plan data use; Microsoft 365 Copilot Business pricing.

Want help turning your AI idea into a case people approve?

On a 1:1 call we'll pick the process worth making the case for, check the costs and assumptions on your page, and agree a stop rule your co-owner or trustees can accept.

Book a 1:1 call with me