ProcessArtificial Intelligence
How to prioritize AI use cases with separate value, feasibility and risk scores
Most AI idea lists stall because every proposal gets one blended score that hides why it might fail. This method keeps value, feasibility and risk as three separate scores, treats some risks as gates rather than weights, and ends with a short portfolio: one or two quick wins paired with foundation work that later use cases reuse. It is the scoring core of the Assess and Prioritize step in ColdAI's AI approach1.
On this page
- Why long AI idea lists stall
- Six steps from idea list to funded portfolio
- How a candidate moves through the scoring gates
- Scoring anchors for value, feasibility and risk
- Why multiplying scores hides deal-breakers
- A hypothetical insurer narrows twelve ideas to three
- Traps that distort an AI use case ranking
- Questions and answers
- Sources
Why long AI idea lists stall
A typical idea list mixes vendor pitches, executive requests and tools a team has already tried. Each entry is described in its own way, so none can be compared with another. Three gaps recur: no shared definition of done, no named owner for the workflow that would change, and no baseline showing how the work performs today. Without a baseline, every value estimate is a guess, and the most persuasive presenter wins.
The fix lies less in a clever formula than in disciplined inputs. A ranking holds up only if each candidate is described the same way, scored by people who know the work, and checked against deal-breakers before any averaging happens. This page sets out that method. For candidates that survive it, the AI readiness assessment goes further into the data and platform evidence.
Six steps from idea list to funded portfolio
Write a one-page card for each idea
Capture the workflow, the accountable owner, the decision or action the system would affect, monthly volume and how the work performs now: time per case, error or rework rate, backlog. Ideas that nobody will own do not go forward to scoring.
Score value with the units visible
Estimate value as volume multiplied by time saved, or by the cost of errors avoided, then note any revenue effect and the fit with stated strategy. Keep the arithmetic in hours or currency on the card, not as points, so reviewers can challenge each assumption.
Score feasibility on evidence
Check three things: whether the data exists and may be used for this purpose, whether the system can reach the applications it needs, and whether outputs can be checked against real past cases with known right answers. The last test is the one most often skipped.
Score risk on its own scale
Record the likely risk class under the EU AI Act (Regulation (EU) 2024/1689)2, what a wrong output would cost, whether it can be caught and reversed before harm is done, and who bears it: staff, customers or the public.
Plot, then apply the gates
Place candidates on a value-against-feasibility grid, then remove or reshape any that fail a risk gate. Never multiply or average the three scores into one number.
Sequence the shortlist as a portfolio
Pair one or two quick wins with a foundation item, such as governed access to a key data source or a shared evaluation harness, that several later candidates depend on. Set a date to re-rank once the first results are in.
How a candidate moves through the scoring gates
- Use-case card
The same fields for every idea; no owner means no score.
- Value score
Volume, time or error cost and revenue effect, in real units.
- Feasibility score
Data rights, system access and whether outputs can be checked.
- Risk gate
Unmanageable risk stops or reshapes the idea at this point.
- Value-feasibility grid
Survivors compared side by side rather than by a blended number.
- Sequenced portfolio
Quick wins paired with foundation work later ideas will reuse.
Scoring anchors for value, feasibility and risk
Anchors keep different scorers consistent. Write your own examples next to each anchor from cases your organization recognizes.
| Criterion | Low | Medium | High |
|---|---|---|---|
| Value: frequency and effort | Occasional task taking minutes per case | Weekly or daily task with noticeable effort | High-volume task where time or errors drive cost |
| Value: strategic fit | Useful, but tied to no stated goal | Supports a current departmental objective | Serves a board-level priority or a revenue line directly |
| Feasibility: data | Scattered, undocumented or not permitted for this use | Available with known gaps; permitted for a pilot | Accessible, documented and permitted for production |
| Feasibility: evaluability | No agreed right answer; quality judged by opinion | Experts can grade a sample, with some disagreement | Past cases with known outcomes form a ready test set |
| Risk: cost of a wrong output | Cheap, and caught before it matters | Causes rework or annoys customers | Causes financial, legal or personal harm |
| Risk: reversibility | Every output is reviewed before use | Errors can be corrected after the fact | Actions are irreversible or affect people directly |
On the two risk rows, a High rating counts against the candidate and can trigger a gate.
Why multiplying scores hides deal-breakers
A hypothetical insurer narrows twelve ideas to three
Traps that distort an AI use case ranking
Choosing by demo appeal
Early signalThe shortlist mirrors the presentations that impressed the steering group.
MitigationScore from cards and evidence only, and watch demonstrations after scoring rather than before.
Ignoring change effort
Early signalValue assumes staff adopt the tool immediately with no change to how work is done.
MitigationAdd training, process redesign and role changes to the card, and discount value until adoption is planned.
Building before measuring the baseline
Early signalA pilot starts with a plan to measure the old process later.
MitigationMeasure time, volume and errors on the current process first; without that, no pilot can show improvement.
Proposers scoring their own ideas
Early signalEvery card rates itself high on feasibility.
MitigationHave data and platform owners score feasibility, and risk owners score risk, independently of whoever proposed the idea.
Questions and answers
How many AI use cases should we fund in the first wave?
Usually two or three: one or two quick wins that can show results soon, plus one foundation item, such as a governed data extract or an evaluation harness, that later candidates depend on. Funding more at once spreads scarce data, platform and risk expertise thinly and makes it harder to learn which assumptions held before the next round.
Who should sit on the panel that scores AI use cases?
A small panel with clear lanes: the process owner and a finance partner score value, data and platform owners score feasibility, and risk, legal or compliance score risk. The executive sponsor chairs and makes the final call. Proposers present their cards and answer questions, but they should not score the feasibility or risk of their own ideas.
How often should an AI use case ranking be revisited?
Re-rank when the first results arrive and at each budget cycle, and sooner if something material changes: a foundation item lands and unblocks several candidates, a regulation or internal policy shifts a risk gate, or a provider changes its terms or prices. A ranking that never changes usually means nobody is feeding production evidence back into it.
What happens to a use case once it makes the shortlist?
Each funded candidate gets a fuller business case that costs it in detail, including running and oversight costs; the consulting practice's AI business case page covers that step. Design then starts with an agreed evaluation set and acceptance thresholds, so the pilot is judged on the same terms that were used to choose it.
Sources
- Artificial Intelligence capability: Assess and Prioritize step — ColdAI
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) — EUR-Lex · checked 10 October 2026