The first AI project a company ships matters more than any that follow. If it works, it buys trust and budget for the next one. If it stalls, "we tried AI" becomes the reason not to try again for two years. So the choice of what to build first deserves more care than it usually gets.
This is the process we run in the first two weeks of an engagement. You can run most of it yourself.
Start with the work, not the technology
The wrong way to start is "where could we use AI?" That question produces a list of things that sound like AI: a chatbot, a copilot, a dashboard that talks. The right way to start is "where does repeated work pile up?" That question produces a list of actual bottlenecks, and some of them are AI-shaped.
Spend the first few days with the people who do the work, not their managers. Ask what they did yesterday. Ask what they do every week that they wish they didn't. Watch them do it if you can. The candidates come from here.
Score each candidate on four things
1. Value: hours and dollars, honestly estimated
How many hours a week go into this, across how many people? What does it cost when it is done late or done wrong? Put a number on it, with a range. Resist the urge to round up. A use case worth 10 real hours a week beats one worth 40 imagined hours.
2. Data readiness: does the system have what it needs to see?
Where does the input live? Is it in a system with an API, or in a shared drive of PDFs, or in someone's head? Is it reasonably consistent? The best use case in the world is not buildable if the data has to be hand-collected first. Score this harshly. It is where most estimates go wrong.
3. Tolerance for error: what happens when it is wrong?
Some work is forgiving: a draft that a person edits, a classification that gets reviewed. Some is not: a payment, a diagnosis, a filing. For a first project, pick something forgiving, or something where a person reviews the output before it has any effect. You want the model's mistakes to be cheap while you learn its failure modes.
4. Visibility: will anyone notice it worked?
A first project needs a sponsor who cares and a result that is visible to leadership. A back-office win that saves the finance team eight hours a week and that the CFO can see in a monthly report is a better first project than a slightly larger win nobody will hear about.
Rank, then pick the boring one
Put the candidates in a table, score each on the four factors, and rank. Then look at the top three and pick the one that scores highest on data readiness and tolerance for error, even if it is not the highest on value.
This feels wrong. Leadership wants the big one. But the big one usually has the messiest data and the least tolerance for error, which is exactly why nobody has automated it. Take the boring, well-bounded one first. Ship it in a quarter. Then take the big one with a team that has done it once.
Traps we see repeatedly
- The chatbot by default. A customer-facing chatbot is the most visible, least forgiving, hardest-to-measure first project there is. It is rarely the right one.
- The "platform" first. Building a general internal AI platform before a single use case has shipped. You will build the wrong platform.
- Replacing a whole role. Scoping the project as "automate what Dana does." Scope it as "automate the two hours of Dana's day that are the same every day." Dana will help you with the second one.
- Skipping the baseline. Not measuring how long the work takes today. If you do not know the before, you cannot prove the after, and proving the after is how you get the second project funded.
- Confusing a vendor demo with discovery. A vendor will show you what their product does. Discovery is about what your business needs. They are different meetings.
Write down the decision
At the end, you should have one page: the use case, the people it affects, the estimated hours and dollars with assumptions stated, the data it depends on and where that lives, what "success" is as a number, and the risks. That page is what you build from, and it is what you measure against in three months.
If you cannot fill in that page for any candidate, that is the answer too. Sometimes the honest outcome of discovery is that the data is not ready, or the process should be fixed before it is automated. Hearing that early is cheap. Hearing it after a build is not.
Want help picking and shipping your first one?
Discovery is free: a 45-minute call, then a short report of potential AI solutions with estimates and an honest recommendation, whether or not you build with us.
More from KloudDevs: Why most AI projects stall between the demo and the deploymentThe approval checkpoint: letting an agent do real work without losing control