Picking Your First GenAI Project: What to Look For
Starting with generative AI can feel exciting and overwhelming at the same time. Many teams jump straight into a complex use case, only to find that the data is messy, the stakeholders are uncertain, or the expected benefits never materialise. The trick is to choose a first project that is small enough to finish quickly yet meaningful enough to show real value. That is the thinking behind a focused, hands-on effort such as the GenAI Builder Sprint, where the goal is to ship a working prototype in two days. But even in a compressed timeframe, the project you pick makes all the difference.
This article offers a practical checklist for selecting that first generative AI project. The points here are not theoretical. They come from watching what works and what stalls when people try to apply large language models and similar tools to everyday business problems. If you keep these criteria in mind, you will increase your chance of finishing with something that people actually want to use.
Focus on a repetitive task that consumes time
The best candidates for generative AI are tasks that people do over and over, following a pattern that does not change much each time. Think of drafting routine emails, summarising meeting notes, extracting key data from invoices, or writing first drafts of standard reports. These are tasks where the output format is predictable and the main variation comes from the input data.
Why does this matter? Generative AI models excel at producing plausible text based on a prompt. If the task requires creativity, judgment, or deep domain expertise, the model will often produce something that looks good on the surface but is wrong underneath. A repetitive, pattern-based task gives the model a fair chance to succeed. It also means that the time saved per repetition adds up quickly, making the project easier to justify.
When you look at your own workflow, ask yourself: which task do I or my team do every day that feels like a template with blanks to fill? That is your starting point.
Check that the input and output are clear
A generative AI project needs a well-defined input and a well-defined output. The input might be a short text prompt, a document, a set of data fields, or a combination of these. The output should be something you can describe in one sentence: a one-page summary, a draft reply, a list of action items.

If the input is vague or the output is open-ended, the model will produce inconsistent results. For a first project, you want a tight scope. For example, instead of asking the model to "improve customer satisfaction," ask it to "write a polite response to a customer complaint about a delayed shipment, using these three bullet points from the agent's notes." The more specific you can be, the easier it is to test whether the output is good enough.
Ensure you have at least a small set of examples
Generative AI models learn from examples. Even with a well-crafted prompt, you will get better results if you can show the model a few correct input-output pairs. This is often called few-shot prompting. The examples do not need to be a huge dataset. A dozen carefully chosen samples can be enough to guide the model toward the right tone, structure, and level of detail.
If you cannot find even a handful of real examples from your own work, the project may be too abstract. Either the task is not well understood yet, or the output format varies too much to be automated. In that case, spend time creating examples manually before you start the AI work. That exercise alone will clarify what you really want.
Identify a process owner who will use the output
Every successful automation project has a person who cares about the result. That person is the process owner. They are the one who does the task today, or who manages the people who do it. Without a committed process owner, the project risks becoming a technical experiment that nobody adopts.
Talk to the process owner early. Ask them what they find tedious about the task, what they would change, and what a good output looks like. Their answers will shape your prompt design and your success criteria. They will also be your best ally when you need to show that the prototype actually helps. If the process owner is sceptical or indifferent, the project will struggle to gain traction even if the technology works perfectly.
This is where the UCC approach to workforce training becomes relevant. In a short, intensive format like the GenAI Builder Sprint, participants bring their own use cases and work through them with guidance. The process owner is often the participant themselves, which solves the commitment problem from the start.
Look for automation potential that is visible but low-risk
Generative AI can make mistakes. It can produce plausible-sounding nonsense, omit important details, or reflect biases in its training data. For a first project, choose a task where a mistake is inconvenient but not damaging. Avoid tasks that involve legal advice, medical diagnosis, financial calculations, or any decision that could harm someone if it goes wrong.

Good low-risk candidates include internal communications, draft documents that a human will review before sending, or data extraction where the original source is still available for verification. The goal is to save time, not to replace human judgment entirely. If the cost of an error is low, you can afford to experiment, iterate, and learn without fear.
Estimate the return on effort, not just the return on investment
Return on effort is a more practical measure for a small project. Ask: how many hours per week does the task take today? How much time would the AI save if it produced a usable first draft 80% of the time? Multiply that by the number of people doing the task, and you have a rough sense of the benefit.
Compare that to the effort of building the prototype. In a two-day sprint, the effort is capped by design. That makes it easier to pick a project where the return on effort is clearly positive. If the task takes someone an hour a day, saving even half of that across a team of ten people adds up to meaningful time. If the task happens only once a month, the savings may not justify the work.
Be honest about the trade-off. A project that saves ten minutes a day for one person is fine as a learning exercise. A project that saves an hour a day for a whole team is a genuine win. Both have value, but knowing the difference helps you set expectations.
Test with a small sample before you commit
Before you spend time building a full prototype, run a quick test. Pick five to ten real inputs from the task and feed them to a general-purpose model using a simple prompt. See what comes out. Does the output look usable? Are the errors predictable? Does the model understand the context without excessive coaching?
This quick test costs almost nothing and reveals whether the use case has basic feasibility. If the model produces something close to what you need, the project is promising. If the output is consistently off-target, either the task is too hard for the current technology, or your prompt needs more structure. Either way, you learn before you invest.
This is exactly the kind of validation that a focused sprint environment encourages. In the UCC classroom model, participants move from idea to working prototype in two days, so the emphasis is on rapid testing and iteration rather than long planning cycles.

Plan for the handoff, not just the build
A prototype that sits on a laptop and never reaches the people who need it is a missed opportunity. Think about how the output will be delivered. Will it go into an email draft? A shared document? A chat interface? The simplest delivery method is often the best for a first project. A custom interface can come later if the prototype proves itself.
Also consider how users will interact with the AI. Will they paste input and copy output? That is fine for a start. The key is to remove friction, not add it. If the prototype requires technical steps that the process owner cannot do, the project will not survive past the demo.
Choosing a first generative AI project is a balancing act. You want something feasible, valuable, and low-risk, with a clear owner and a realistic delivery path. The checklist above is not exhaustive, but it covers the factors that most often separate projects that ship from projects that stall. If your candidate use case checks most of these boxes, you have a strong starting point. If it misses several, consider a different angle or a simpler task. The goal is not to build the most impressive AI application on the first try. The goal is to build something that works, that someone uses, and that teaches you what to do next.