Choosing your first AI project means scoring every candidate idea against four filters. Those filters are business value, implementation effort, data readiness, and risk and reversibility. Score each idea honestly on all four, then pick the option that scores highest on value and readiness while scoring lowest on effort and risk.

Most Irish SME owners approaching their first AI project run into a surprising problem. There are too many ideas, and no reliable way to rank them. Some of those ideas also carry extra weight under Article 6 and Annex III of the EU AI Act. Those articles flag certain uses, recruitment screening and credit decisions among them, as needing a closer look before a pilot begins. [1] If a long list of AI ideas and no way to rank them sounds familiar, that is precisely the gap this scoring method is built to close.

What does it actually mean to choose an AI project?

Choosing an AI project is the act of picking one specific, bounded piece of work from a longer list of possibilities and committing resources to it first, ahead of everything else on that list. It comes after two earlier decisions. The first is whether the business is ready to adopt AI at all, covered in a free readiness self-check. The second is what the full adoption journey looks like end to end, covered in the six-step method for introducing AI into your business. This article picks up where those leave off. A shortlist exists, and one item on it has to go first.

Most shortlists come from a rough task-by-task audit of what is worth automating, built informally over a few conversations. The shortlist itself is rarely the issue. What happens next decides the outcome. Every item looks equally plausible on paper, and the business tends to choose whichever one sounds most impressive over whichever one would actually work.

In summary

Build the shortlist first. Score it second. Never skip straight to a single favourite.

Why do most first AI projects fail?

Most first AI projects fail on selection, not execution. The business picks the most exciting idea on the shortlist, and only later discovers it never scored well against a clear set of filters. MIT's Project NANDA studied more than 300 disclosed enterprise AI initiatives in 2025 and found that 95% delivered no measurable financial return. [2] Most AI investment went into visible areas like sales and marketing. The strongest returns actually came from less visible, everyday work such as customer service and HR administration. In other words, where the money went mattered just as much as which technology was used.

That gap matters for a business picking its very first project. Most people assume the model is where things go wrong. The real cause sits earlier than that, in the shortlist itself, and specifically in which item on it got chosen first. A three-person accounts team spending four hours a week matching invoices to purchase orders is a duller pitch than an AI tool that writes marketing copy. It is still far more likely to be the project that works, because the task is narrow and repeatable, and the people doing it already understand it end to end.

In summary

If you cannot describe the win in one sentence, it is not ready to be your first project.

What are the four filters for choosing an AI project?

Four filters decide whether an AI idea on your shortlist is ready to become your first project. Those filters are business value, implementation effort, data readiness, and risk and reversibility.

Business value asks whether the outcome can be stated in one sentence and measured with one number. That number could be hours saved, error rate, response time, or revenue touched. If it cannot clear that bar, it is not ready to be scored yet. Implementation effort asks how much setup, data cleanup, or integration the idea needs, and whether a first version could genuinely run within a few weeks. Data readiness asks whether the business already holds the information this project needs. That information has to be in a form a tool can use, and the business has to be legally entitled to use it for this purpose. The full version of that check is covered in the five data-readiness checks. Risk and reversibility asks how easily the project could be undone if it went wrong, and whether it touches anything the EU AI Act treats as higher risk.

On that last point, Annex III of the EU AI Act lists domains such as recruitment screening and credit decisions as high-risk. A narrower classification is available under Article 6(3), but only if the system meets specific conditions set out in that provision and does not profile individuals, meaning it does not use someone's personal data to evaluate or predict something about them. A provider relying on that exception has to document the assessment before the system goes live. Under Regulation (EU) 2026/1744, part of the 2026 Digital Omnibus package, the Annex III application date was pushed from 2 August 2026 to 2 December 2027, so this is a planning input for a first project now rather than an August 2026 deadline. [1]

Separately, the business itself is required under Article 4 to put AI literacy measures in place for whoever runs the pilot, matched to their role. It does not set a personal AI literacy requirement for the individual, and it has applied since February 2025 and was untouched by that deferral. [1] A full AI risk assessment covers this classification question, along with the Article 5 and Article 50 obligations that sit alongside it, in depth for any project that needs it.

In summary

Score every idea the same way, or the scoring does not mean anything.

How do you score and compare candidate projects?

Scoring candidate projects means rating each one low, medium, or high on all four filters, then comparing the results side by side instead of trusting instinct.

Picture two genuine candidates from a typical shortlist. The first is an AI tool that drafts replies to common customer queries using existing FAQ content. High value, measured in response time. Low effort, a first version running in days. High data readiness, the FAQ content already exists. Low risk, a human reviews every reply before it sends. The second is an AI tool to screen job applicant CVs. High value on paper. High effort once bias testing and process documentation are added. Patchy data readiness, years of inconsistent CVs and notes. And a risk score that lands it squarely inside Annex III recruitment screening. Scored honestly, the first project clears all four filters. The second stalls on two of them, regardless of how important the department pushing for it, recruitment in this case, looks on the org chart.

That pattern matches what the wider research on enterprise AI outcomes keeps finding. The projects that deliver a return tend to sit in the ordinary, well-understood corners of a business, not the ones that make the best pitch. [2] A simple low, medium, high scoring pass, done honestly on paper, surfaces that pattern in your own business before you have spent a euro.

In summary

Which of your shortlisted ideas can you actually measure in four weeks?

What should you do once you have picked a project?

Once a project is chosen, run it as a single bounded pilot. Give it one owner, one metric, and a fixed few weeks to run in. Not an open-ended experiment.

Resist the pull to start a second or third pilot at the same time. Running several projects at once feels efficient. It usually produces three half-finished pilots and no clear read on what actually worked. Pick the tool for the chosen project only after the project itself is settled. Use a four-question filter for choosing the right tool rather than starting from the tool and working backwards. Whoever owns the pilot day to day should understand it well enough to run it safely under Article 4. That is a training question worth answering before the pilot starts. Ideally, answer it the week before, not partway through.

In summary

One pilot. One owner. One number. That is the whole test.

When should you get outside help choosing?

Outside help is worth bringing in when the shortlist is long, nobody inside the business has the time to score it properly, or the risk filter throws up a use case that nobody feels confident classifying.

Irish businesses that have already used AI cluster around a narrow set of purposes. Business administrative processes and marketing or sales lead the list, with accounting, production, and logistics some way behind. [3] That leaves most of a typical shortlist unclaimed territory, with no established pattern in the wider market to copy from. A second, outside pair of eyes on the same shortlist tends to catch the project that everyone in the room overlooked, usually because it was too dull-sounding to make anyone's shortlist in the first place.

If a long list of AI ideas and no way to rank them sounds familiar, that is exactly what the AI Use-Case Discovery Workshop is built to sort out, working through the specific ideas already on your list.

In summary

A second pair of eyes usually catches the project everyone else missed.