How Do You Decide Which AI Ideas Are Worth Funding?

Most leadership teams aren’t short of AI ideas. Every department has a list. Finance wants invoice matching. Operations wants demand forecasting. Customer service wants a chatbot. Someone in marketing has already bought three tools nobody signed off.

They all sound promising. Some will deliver real value. Some are nice-to-haves. A few would create more risk than they’re worth.

Finding ideas is the easy part. The hard part is choosing which one gets a budget. Here’s the scoring model we use in our client workshops, the mistakes that sink AI business cases, and how to get from a long list to three funded use cases.

Why Good Ideas Still Fail

The numbers aren’t kind to AI projects that skip this step. Gartner reports that by the end of 2025, at least half of generative AI projects had been abandoned after proof of concept, citing poor data quality, weak risk controls, rising costs or unclear business value. (Gartner) gartner

Look at that list again. Unclear value is a value problem. Poor data is a feasibility problem. Weak risk controls is a risk problem. Rising costs touches all three.

Those projects weren’t killed by the technology. They were killed by questions nobody asked before the money went in.

The wider picture tells the same story. In McKinsey’s latest global survey, only 39% of respondents attribute any EBIT impact to AI, and most of those say it accounts for less than 5% of their organisation’s EBIT. (McKinsey) Plenty of activity. Not much on the bottom line yet. mckinsey

Three Questions, Scored Together

We score every AI opportunity against three questions.

Value: what measurable outcome does it move, and by how much? “It’ll save time” isn’t an answer. “It cuts invoice processing from four days to one, across 2,000 invoices a month” is. If nobody can name the number, the idea isn’t ready for a budget.

Feasibility: do we have the data, systems and skills to deliver it? The data needs to exist, be accessible and be good enough. The solution needs to fit the systems where the work actually happens. And someone needs to build it, run it and fix it when it breaks.

Risk: what could go wrong, and who carries it? A wrong answer to a customer. A decision in a regulated process. Sensitive data in the wrong place. The second half of the question matters as much as the first. If the customer, the compliance team or the frontline staff carry the risk, they need a voice in the decision.

How the Scoring Works

Keep it simple. A model nobody understands won’t survive contact with the boardroom.

• Score each question from 1 to 5.
• Score risk in reverse, so 5 means low risk.
• Count value twice. It’s the reason you’re doing this in the first place.

That gives a maximum of 20. Then apply two rules.

Any score of 1 is a stop sign. Don’t let a strong total average away a fatal flaw.

Every score needs a sentence of evidence. A 4 with no reason behind it is really a 2.

Why “Together” Matters

The scoring is useful. The conversation is where the value is.

When finance, operations, IT and legal score in the same room, disagreements surface early. A sponsor gives feasibility a 5. IT gives it a 2. That gap is the most useful number in the whole workshop, because it’s a problem you’ve found before anyone has spent a penny.

Scored in isolation, every department’s idea comes back a 5. Funny how that works.

Five Mistakes That Sink AI Business Cases

1. Value with no baseline. “Improves efficiency” can’t be measured, so it can’t be proven. If you don’t know how long the process takes today, you’ll never show the improvement.

2. Counting hours as cash. Hours saved only become money if the work is redeployed or the cost is removed. Your CFO knows this. Build it into the case before they point it out.

3. Scoring the pilot, not production. Gartner warns that token costs that look trivial in a pilot can become a serious cost of ownership problem once multiplied across thousands of users and many use cases. (Gartner) Score feasibility and cost at the scale you actually intend to run. gartner

4. Treating data as a phase two problem. If the data isn’t ready, the project isn’t feasible. Pretending otherwise just moves the failure further down the line, where it costs more.

5. Risk with no owner. If the answer to “who carries it?” is “the vendor” or an awkward silence, the risk score should be low. Accountability can’t be outsourced.

From Long List to Three Funded Use Cases

Collect, then filter. Ask for one line per idea: the problem, the outcome and the owner. If someone can’t fill in all three, the idea goes back for more thinking.

Score together in one session. Bring the people who’d fund it, build it, run it and carry the risk.

Sort into three piles. Fund now. Fix first. Park.

Pick three. Three is enough to spread your bets and few enough to resource properly. Funding ten ideas a little usually means none of them get far. A good mix is one quick win that can prove value within a quarter, plus one or two bigger bets.

Write a one-page case for each. Include the metric, the baseline, the target, the owner, the cost at scale, and what result would make you stop.

Test before full spend. A score is still a judgement. Before committing the full budget, put something real in front of users and check the assumptions hold. We’ve written about how to test an idea before you fund the build.

A Shorter Conversation With Your CFO

Your CFO is going to ask three questions anyway. What do we get? Can we deliver it? What’s the downside?

Turn up with those answers scored, evidenced and agreed across the business, and the meeting gets a lot shorter.

The parked ideas aren’t wasted either. Data improves, tools mature and priorities shift. Re-score the list every quarter and you’ll often find that last quarter’s “not yet” has become this quarter’s quick win.

If you’re sitting on a long list and need a shortlist you can defend, our AI Strategy team runs these workshops with leadership teams every month. Where feasibility is the blocker, our Data Strategy & Governance team can help fix the foundations. And for the risk side, our AI Governance service helps make sure someone owns it.