Choose your first AI (artificial intelligence) project by starting from one recurring business decision, not a technology. Data Department, an applied machine learning consultancy, recommends a use case that is made often, has an owner, has two or more years of usable data and can be measured against one business metric agreed on day one. This guide is for operations, finance and IT (information technology) leaders in mid-sized and large companies.
Why does the first project matter so much?
The first project sets the company's opinion of AI for years. If it ends as a demo that nobody uses, the next budget request is harder. If it changes a real decision and the result shows up in a number the finance team already tracks, the second and third projects are easier to fund.
Adoption is still early for most companies. In the European Union, about one in five enterprises with 10 or more employees used AI technologies in 2025, up from 13.5% a year earlier (Eurostat, 2025, reported by BTA). Most companies are therefore choosing a first or second project now, often without in-house experience to judge the options.
What makes a good first AI project?
A good first project has six properties. It is built around a decision someone makes repeatedly, such as how much stock to hold, what price to set or which customers to call. The decision has a named owner who will change how they work. The data already exists in an ERP (enterprise resource planning) system, a MES (manufacturing execution system) or a CRM (customer relationship management) system. The value can be measured in money, time or service level. The result can be written back into an existing system. And the risk of a wrong output is manageable, because a person still reviews it or the cost of an error is small.
Projects that fail usually miss one of these. A common pattern is a model that predicts well but feeds no decision. Another is a model that needs data the company does not collect.
How do you score candidate use cases?
List five to ten candidate use cases, then score each one from 1 (weak) to 3 (strong) on the criteria below. The table is the same checklist we use in a 2-week Discovery.
| Criterion | Question to ask | Strong (3) | Weak (1) |
|---|---|---|---|
| Decision | Which decision changes if the model works? | Named, recurring decision | "Better insight" |
| Frequency | How often is it made? | Daily or weekly, many items | Once a year |
| Owner | Who will act on the output? | One accountable manager | Unclear or a committee |
| Data | Is there history at the right level? | 2+ years, item or customer level | Spreadsheets, gaps |
| Metric | How will we know it worked? | Existing finance or operations KPI (key performance indicator) | New metric invented for the project |
| Integration | Where does the output go? | Into ERP, planning or CRM | A separate dashboard |
| Risk | What happens when it is wrong? | Reviewed or low-cost error | Safety or legal exposure |
Add the scores. Anything above 17 out of 21 is a strong candidate for a pilot. Between 13 and 17, the gap is usually fixable during Discovery. Below 13, park it.
Also estimate the value range roughly: a cost base or revenue line multiplied by a realistic improvement. You do not need a precise business case at this stage, only enough to rank options.
Which use cases usually score well?
In our experience the use cases that score highest for a first project are the ones close to money and repeated at scale:
- Demand forecasting and inventory: thousands of items, decided weekly, measured in holding cost and service level. In one project for an automotive distributor, probabilistic forecasting gave about 30% better forecast accuracy and about 10% lower inventory holding cost (case study).
- Dynamic pricing: many products, prices changed often, measured in margin. For an e-commerce retailer, this approach raised profitability by about 4%.
- Churn and retention: which customers to contact with which offer, measured in retained revenue. At an automotive distribution group, renewal and brand-churn models produced about an 8% revenue gain.
- Document processing with generative AI: repetitive reading of supplier documents, measured in hours saved and errors caught.
Use cases that often score poorly as a first project include company-wide "AI platforms", chatbots without a clear task, and computer vision projects that first need new cameras and installation.
What are the most common mistakes?
Starting with the tool. "We should use a large language model" is not a use case. Start with the decision and choose the method after.
Choosing the most visible problem instead of the most tractable one. The board's favourite topic may have no data. A smaller, well-measured win builds trust faster.
Skipping the baseline. If you do not measure how the decision is made today, you cannot show improvement. Record the current error, cost or time before the pilot begins.
Measuring model accuracy instead of business results. Accuracy is an intermediate number. See why forecast accuracy is not enough.
No plan after the pilot. Decide before you start who will run the model in production, where it will be hosted and how it will be monitored.
How long should a first project take?
A realistic path looks like this:
- Discovery, 2 weeks. Rank use cases, check the data, agree on the business metric and make a go or no-go decision.
- Pilot, 6 to 10 weeks. One use case, measured against the agreed metric, ideally running alongside the current process.
- Production and handover. Integration with ERP or planning systems, monitoring and training for the team that will own it.
If a candidate cannot show a measurable result within a 6 to 10 week pilot, it is usually too large for a first project. Split it.
When is AI the wrong tool?
Sometimes the honest answer is that a first project should not use machine learning at all. If the decision follows clear rules, a rules engine is cheaper and easier to audit. If the data is not recorded, fix data capture first. If the volume is small, a well-designed spreadsheet may be enough. We say this in the free 30-minute AI opportunity review when it applies, because a first project that should not have been an AI project damages the next one.
Frequently asked questions
How much data do we need for a first AI project?
For most forecasting and customer models, two or more years of history at the level of the decision (item, store, customer) is a good starting point. Less can work if the signal is strong, and a Discovery phase checks this before any build.
Should the first project be generative AI or classical machine learning?
Choose the method after the decision. Classical machine learning fits numeric decisions such as stock or price; generative AI fits reading and writing tasks such as documents and internal search.
Who should own the first AI project?
The manager who owns the decision, supported by IT. A project owned only by IT or only by a data team tends to stall at adoption.
What if our data is spread across several systems?
That is normal. The Discovery phase maps which systems hold the needed fields and how often they are updated; most first projects need data from two or three sources.