The one-page case

Everything up to now has cost nothing new: redesigning work, training people, setting up the tools you already pay for. Build is different. It means spending new money or engineering time, and it only starts once improving can't close the gap and the gap is worth paying for. Before any of that money is spent, the idea gets one page.

Why one page

AI projects rarely fail because the technology doesn't work. They fail because nobody wrote down what they were for, what they would really cost, or what would make anyone stop. In 2025, Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, and gave three reasons: escalating costs, unclear business value and inadequate risk controls.

Those are three of the five things the one-page case makes you write down before you start. A page is long enough to make those decisions, and short enough that a leadership team actually reads it. It covers value, cost, people, risk and the options.

Value: what it's for

Value comes in three kinds, and the case names which one it's claiming.

• What it frees at the bottleneck. More through the step that limits the aim, valued in the aim's unit. Not hours saved: speed anywhere else is worth nothing until it's converted.
• The cost it removes. A contract that ends, overtime that stops, an outsourced service you no longer need. MIT's research found back-office uses often delivered faster payback and clearer cost cuts than the sales and marketing uses most AI budgets go to.
• What an experiment will teach. For new offerings, paid from the experiment budget. Lesson 4 covers how.

Value at the bottleneck is counted as the extra output the whole organisation gets when it moves, in the aim's unit. Two more quotes a week, say, at the historic win rate and the average deal size. If you put a value on your bottleneck in module 4, the case below picks it up.

Cost: the full running cost

The quote is the part everyone sees. The running cost sits below the waterline: licences, usage and hosting, monitoring and support, and the hours the owner spends looking after it, every year it runs. Gartner found in 2024 that organisations which don't understand how generative AI costs scale can get their cost calculations wrong by 500% to 1,000%.

Building is getting cheaper to start. In McKinsey's 2026 survey, nearly a third of respondents said their organisation had turned down a software purchase in favour of building it with AI coding tools. The running cost hasn't fallen with it. Something built in a fortnight still needs an owner, monitoring and a way to undo it for as long as it runs. Put your own numbers in below and watch where the money actually sits.

People: three names

• An owner, who answers for it once it runs: whether it's healthy, whether it's still worth it, and what happens when the vendor changes it.
• A backup, who can stop it when the owner can't.
• A builder of record, who builds, buys or configures it, and answers for how. In a small firm, that's often the outside builder.

Three names on the page before anything is built is how you avoid orphaned apps: tools that keep running long after the person who made them has moved on.

The case is usually written by the builder of record and the bottleneck lead, starting from what came through the front door. Leadership funds builds at its quarterly meeting, inside the spending ceiling on your aim page. If the case won't fit on one page, the idea isn't clear enough yet.

Risk: four lines

• A data check: the most sensitive data it touches, checked against the data map.
• The stakes: from the two questions in lesson 2. High stakes means the proof in lesson 4 needs evidence from someone who didn't build it.
• A stop-loss: the spending or time limit at which it stops, whatever anyone hopes.
• Kill criteria: the evidence that would end it, written now, before anyone is attached to it.

Options: cheapest first

The case lists five options: reuse, configure, buy, build, or don't. Take them in that order. Each one costs more than the one before and is harder to undo, so each is only worth considering once the one before it has been ruled out.

Buying comes after the vendor checklist: eight questions about your data, access, controls and costs, the same ones every approved tool answered in module 2. And the evidence on buying versus building is worth knowing before you choose.

“Don't” is a real option. A case that ends in don't has done its job: the money stays inside the ceiling, the friction stays visible on the list, and the person who raised it hears why. Now write yours, for the one thing most worth paying for.