The highest-impact AI use case is found by mapping real workflows, listing where time or money leaks, scoring each candidate on value, feasibility, and data readiness, and picking the one with the best score that a single team owns. Done with discipline, this takes about two weeks and ends with one use case ready for a pilot.
Companies rarely lack AI ideas. They lack a way to rank them. A brainstorm produces thirty candidates, a vendor pitches five more, and six months later nothing has shipped because everything looked equally promising. Discovery is the process of turning that list into one decision.
This is the method we use at the start of every engagement. It works for a factory, a foundry, a mine, a B2B distributor, or a finance team, because it starts from workflows rather than from technology.
Week one: map the workflows and the leaks
Start with the people who do the work, not with the systems. Sit with operators, planners, quality staff, sales engineers, or analysts and walk through a normal day. Write down each step, who does it, what information they use, and where it goes wrong.
- Where does someone wait for information that already exists somewhere else?
- Where is a judgement made repeatedly on the same kind of input: a part, a document, a quote, a transaction?
- Where does a mistake cost real money: scrap, rework, downtime, a lost order, a compliance issue?
- Where is expert knowledge held by one or two people who cannot be everywhere?
- Where does data get typed from one system into another by hand?
By the end of week one you should have a list of ten to twenty candidate use cases, each attached to a specific workflow and a specific cost.
Week two: score, check the data, and choose
Score each candidate on three dimensions, one to five each. Keep the scoring in a shared table so the reasoning is visible.
| Dimension | Question | High score looks like |
|---|---|---|
| Value | What does the problem cost per year, and who owns that budget? | A named cost line and a named owner |
| Feasibility | Has this class of problem been solved with AI elsewhere? | Proven techniques, clear inputs and outputs |
| Data readiness | Does the data exist, is it accessible, and is it labelled or labellable? | Data already captured in a system you control |
| Adoption | Will the people in the workflow accept and act on the output? | Users asked for it or helped shape it |
Then check the data for the top three. Pull a real sample, look at its quality, and confirm you can access it continuously rather than as a one-off export. Data readiness kills more use cases than any other factor, and it is cheapest to discover here.
What a good first use case looks like
Pick a use case that
- Sits in one team and one workflow, so one person can own the pilot.
- Has a cost you can measure before and after.
- Uses data you already collect, or can start collecting within days.
- Has a safe failure mode, so an uncertain output can be routed to a human.
Avoid, for the first one, a use case that
- Spans several departments and needs everyone to change at once.
- Depends on data that does not exist yet.
- Has no owner with a budget line attached.
- Is interesting to the technical team but invisible to the business.
Common first use cases by sector
- Manufacturing: visual inspection for a defect type that manual inspection misses, or predicting a recurring machine fault from existing sensor data.
- Metals and foundries: optimising the charge mix to cut raw-material cost per melt while staying within specification.
- Mining: forecasting output and constraints from production and equipment data that is already logged.
- B2B commerce: answering product and pricing questions on the website and routing qualified leads to sales.
- Finance: flagging anomalous transactions or automating a recurring report that analysts assemble by hand.
Frequently asked questions
- Do we need an AI strategy before choosing a use case?
- No. A strategy written before any system has shipped is usually guesswork. One use case delivered to production teaches more about your data, systems, and people than any strategy document, and the strategy that follows is grounded in it.
- What if our data is messy?
- Most data is. The question is whether it is messy in a way that can be cleaned in the pilot, or missing in a way that cannot. Sampling real data in discovery answers that before you commit budget.
- How many people should be involved in discovery?
- A small core of three to five: the business owner of the workflow, one or two people who do the work daily, someone who knows the systems, and the AI partner. Wider input is gathered through short interviews rather than large meetings.
- Can the two weeks be shorter?
- For a single well-understood workflow, yes. For a plant or a company-wide review, two weeks is already tight. The point is a fixed timebox that ends in a decision, not a specific number of days.
Ready to find your first use case?
Discovery is how every Adente engagement begins: two weeks, your workflows, and one use case chosen on value, feasibility, and data.
Start discovery