05A business case built, not proven
Why the case cannot be proven up front, and how funding in stages builds it: crawl, walk, run, fly.
Boards are used to business cases like a bank's payment engine replacement: requirements set by a standard, a deadline set from outside, a value that is largely binary (compliant or not), costs estimated from similar projects, then a predictable run budget. Such a case can be wrong, but it can be written with confidence before the first euro is spent.
Our rail maintainer's project fails almost every one of those conditions. Its value depends on how often the system is right, how it behaves when it doesn't know, and how much technicians use it, none of which can be known in advance. It depends on a baseline nobody has measured. Part of it is easy to price, such as time saved or fewer calls to experts, but the largest part is a loss avoided: the day the last person who knows an older engine family retires, can the company still maintain it safely? And the costs don't stop at delivery.
Building the case step by step
A business case that cannot be proven up front has to be built as the project progresses, by funding in stages. In GenAI Beyond the Hype, I described this as think big, start small, and progress through four stages: crawl, walk, run, fly. Each stage must return something of value on its own, so that stopping is never a pure loss, and end with an explicit decision based on criteria agreed in advance.
For our rail maintainer, governance shapes the sequence. During Crawl, no answers reach technicians: AI audits the documentation for contradictions, outdated versions and gaps, reviewed by content owners, and helps turn interviews with retiring experts into documents they correct. On that initial scope, the team also builds a first working skeleton of the system and tests it internally against the test set. During Walk, the same skeleton grows into a production foundation and reaches real users for the first time: the system finds the source rather than giving the answer, this document, this page, this revision. The document remains the authority. Direct answers come with Run, on lower-consequence content; safety-critical answers come last, under the strictest criteria, or never.
| Stage | Question it answers | What it returns, even if we stop | Move on if… | Stop or rethink if… |
|---|---|---|---|---|
| Crawl | Can we measure, and may we proceed? | Audited documentation; expert knowledge written down; the baseline; a first expert-validated test set; a working skeleton, tested internally | Use case and data are placed on the governance map, documents are usable, and the baseline shows a real gain | Documents are too poor, the data cannot be used under acceptable conditions, or today's way of working is good enough |
| Walk | Does it reliably find the right source, on one engine family? | A tool that points technicians to the right document, page and revision; expert corrections in writing | The right source is found at the agreed level, and no answer draws on a document the user may not see | Quality stalls despite several cycles, or access rights cannot be enforced reliably |
| Run | Is the value real in daily use? | Direct answers on lower-consequence content in one depot; measured time savings | Technicians use it, the field gain matches the test results, and the system could be operated for years | Usage stays low, or the field gain falls clearly short |
| Fly | Can we extend it, including safety-critical answers, and sustain it? | Coverage across the fleet; knowledge preserved before experts leave | Running costs stay below the value, and safety-critical answers meet the strictest criteria or remain pointers to the source | Running costs grow faster than value, or quality degrades as coverage expands |
Even if the project stops after Crawl, the company has audited its documentation, mapped its data, and started writing down what only retiring experts know. That alone may justify the spending. Stopping because a criterion is not met is not a failure; it is the business case doing its job, at a fraction of the cost of finding out later.
The criteria are written in words here because the right thresholds depend on the company. They must be turned into numbers before each stage starts, by the people who will live with the consequences. Writing them down protects against the most common drift: a pilot that continues because nobody can say whether it has succeeded. And one figure belongs on the table from day one: the cost of doing nothing. The experts are leaving on a schedule nobody can change.