Framework library · Technology management
AI opportunity portfolio
The AI opportunity portfolio scores every candidate AI project the same way: the value if it works, discounted by how confident you are and how ready the data is, divided by what it costs to prove. A modest prize that can be settled in six weeks then outranks a large one that takes nine months to test, and a candidate that changes nobody's decision, or has no result that would stop it, scores zero.
Use it when
- An AI programme has more candidate pilots than budget, each with a sponsor, and no common way to rank them.
- You are setting up an AI programme and want the rules for starting and stopping pilots agreed before the first one begins.
- A board asks what the AI budget is buying, and the answer has to be in measures the business already tracks.
- Pilots have run for months without ending, and you need a gate that stops them by arithmetic rather than argument.
Avoid it when
- You are choosing between AI suppliers or platforms for one agreed use. Score the suppliers with a weighted decision matrix.
- The work is mandatory, such as a regulatory requirement. Do it; ranking it by return adds nothing.
- Nobody can say even roughly what a candidate is worth if it works. Map the decisions it would change and the metrics they move first, then score.
How to run it
Write each candidate as the decision it changes
"Whether to send an engineer before the customer reports a fault" is a decision. A candidate that changes no decision is a demonstration, not an opportunity.
Name the decision owner and the falsifier
The owner is the person whose decision changes on the answer. The falsifier is the result that would make you stop, written before the work starts. Without either, the row scores zero.
Estimate the value if it works
In $k a year, in a metric the business already tracks: engineer visits, churn, hours, capital.
Rate confidence and data readiness
Confidence: high when measured somewhere comparable (0.8), medium when reasoned (0.5), low when asserted (0.2). Data: ready and governed (1), usable with work (0.7), partial (0.4), not available (0). Not assessed scores zero on either.
Price the cheapest honest test
Cost to prove is what it costs to learn whether the thing works at all, not what it costs to build. Note the weeks. A proof longer than 26 weeks should be split or dropped.
Rank by return per dollar of proof
Expected value divided by cost to prove. Below 3.0, a candidate does not return three times its own cost of proof in a year, so it waits. Change the bar before you score, not after.
Run the kill gate every month
Stop when the falsifier fires, the owner leaves or the data turns out not to exist. Re-rank when the cost to prove doubles, and move what works into the business with an owner and a date.
Work through it
Answer the questions below, or load the worked example to see a finished one. The drawing updates as you type. Export the result as a PowerPoint deck, a Word document, an Excel workbook, a PDF or plain text.
What you type stays in this browser, so you can close the page and come back to it. It is not sent to Blue Prysm or anyone else, and the exports are made here, on your device. Privacy policy.
Mistakes to avoid
- Entering the cost to build as the cost to prove. Large uncertain prizes then never get a cheap first test, and small builds crowd the top of the list.
- Rating confidence high because a supplier demonstrated it. High means the effect was measured somewhere comparable, in your business or in a published result you can check.
- Agreeing the falsifier after the results are in. It then becomes negotiable, and nothing ever fails.
- Keeping a pilot because someone senior sponsors it. The kill gate is arithmetic so that stopping does not depend on a conversation.
Where it comes from
Blue Prysm's own model, published as the workbook "Ranking an AI portfolio by what it costs to find out" (version 1.0, last updated 29 September 2026), with scales for confidence and data readiness, blocking rules, a 3.0 bar and a monthly kill gate. The workbench below follows its columns, scales and rules. Source.
Use it with
Further reading
Work through it with us
The frameworks here are free to use as they stand. If you would rather work through the question behind this one with us, these are the ways an engagement starts.