The first one carries extra weight
A first project is a test of whether the approach works here, in this business, with this data and these people. If it stalls, the second rarely gets funded, whatever the reason for the first one failing.
That argues for choosing on deliverability rather than on which problem is largest. The biggest problem is usually the worst first project.
Five things a good first candidate has
- A decision made repeatedly. Weekly or daily, not a handful of times a year - there has to be enough history and enough opportunity to show benefit.
- Outcomes already recorded. You need to know what actually happened. If nobody recorded it, that is a data collection project first.
- A measurable current method. Something to compare against, so improvement is provable.
- A willing owner. Someone in the business who wants it and will use it.
- Tolerable errors. Where being wrong occasionally costs time rather than safety, money at scale, or a regulator's attention.
All five matter, and the fourth is the one most often ignored. A technically ideal project with no enthusiastic owner will not be adopted.
Candidates that usually work well
| Project | Why it suits a first attempt |
|---|---|
| Demand forecast for the top products | History exists, baseline clear, benefit measurable |
| Support ticket routing | Plenty of labelled history, errors cheap |
| Which overdue invoices to chase | Clear outcomes, immediate cash benefit |
| No-show prediction for appointments | Simple data, direct operational use |
| Spend or product categorisation | Self-contained, useful regardless |
Tempting options to avoid first
- Anything safety-critical. The error tolerance is wrong for a first attempt.
- Your most strategically important process. Too much depends on it, and caution will slow everything.
- Something requiring a new data source. Doubles the project and adds a dependency outside your control.
- Decisions about individuals - hiring, credit, tenancy. Legally sensitive and needing careful handling.
- Anything needing a system change to act on the output. The integration becomes the project.
That last point is worth dwelling on. A prediction nobody can act on without a six-month system change delivers nothing, however accurate.
Plan the second project from the start
The first project should leave you with more than a model: a clearer view of your data, a working pipeline, people who understand what this involves, and evidence for what it is worth.
Choosing a first project adjacent to a likely second one compounds that. Building a demand forecast that later feeds stock allocation is better sequencing than two unrelated efforts, because the data work carries over.
Choose the first project on whether it will finish, not on whether it is the biggest.