How to Run an AI Opportunity Assessment in Two Weeks
Most opportunity assessments produce a list of things AI could theoretically do. That list is worthless. Here is what to produce instead.
Erin Moore
Fractional Chief AI Officer
An AI opportunity assessment should end with three to five specific, owned, sized candidates and an explicit list of what you rejected and why. If it ends with a long menu of possibilities, you paid for a brochure.
Two weeks is enough. Longer assessments do not get more accurate; they get more comprehensive, which is a different thing and usually a worse one.
Week one: find where the work actually is
Go where the labour is, not where the excitement is. Interview the people doing repetitive work, not the executives describing it — the description and the reality diverge sharply, and the divergence is where the opportunity hides.
Four questions, asked of people rather than of systems. What takes you longest that feels mechanical? What do you redo because of someone else's mistake? What do you copy from one system into another? What do you wait on?
That last question matters more than it looks. Queue time is usually a bigger drag than task time, and queue problems are often fixed by routing and triage — which is cheap, boring, unglamorous AI work that pays back reliably.
Alongside the interviews, count. How many times a month, how long each, what a mistake costs. You will be wrong by a factor of two. Being wrong by two is fine; not knowing the order of magnitude is not.
Week two: apply the four tests
Every candidate that survives has to pass all four.
Volume. Is this frequent enough that a percentage improvement is worth anything? Something taking two hours once a quarter is not a candidate no matter how annoying it is.
Tolerance. What happens when the system is wrong? If the answer is "a person notices and corrects it", you can move now. If it is "a customer is harmed and we find out in ninety days", you need controls first and a longer runway.
Data. Does the input actually exist in a usable form? This is where most candidates die, and they die quietly — see data readiness for AI.
Owner. Is there a named person whose job gets better if this works? Not a sponsor — an owner. Unowned projects are the ones that reach demo and stop.
Rank the survivors by value against reversibility rather than by value against effort, for the reasons set out in AI use case prioritization. NIST's AI Risk Management Framework treats this framing — context and consequence established before capability — as the map function, which is exactly what a good assessment is doing.
What the output looks like
One page per surviving candidate: the current process in three sentences, the volume, the proposed change, the tolerance for error, the named owner, a cost range, and what would have to be true for it to fail. Plus one page listing everything rejected, with the reason. That rejection page is the part executives actually reread, and it is what stops the same idea returning every quarter.
Then a sequence. Not everything at once — the first candidate should be the one that teaches you the most about your own readiness, which is rarely the one with the largest number attached.
What to be sceptical of
Benchmark-driven assessments that tell you what companies in your sector are doing. Sector averages tell you about fashion, not about your constraints. Stanford HAI's AI Index is worth reading for the macro picture precisely because it does not pretend to be a plan for your business.
Also be sceptical of any assessment that finds no obstacles. Every real business has data it cannot use, a process too political to touch, and a system nobody will let you integrate with. An assessment that does not surface at least a few of those did not look hard enough.
Frequently asked questions
What should an AI opportunity assessment produce? Three to five specific, owned, sized candidates in sequence — plus an explicit list of what was rejected and why. A long menu of theoretical possibilities is not a usable output.
How long should it take? About two weeks. Longer assessments become more comprehensive rather than more accurate, and the extra breadth usually delays the first real decision.
Who should we interview? The people doing the repetitive work, not the executives describing it. The described process and the actual process differ, and the gap is where most of the opportunity sits.
What is the most common reason a candidate fails? Data. The input does not exist in usable form, or it lives somewhere nobody will grant access to. This kills more candidates than technical difficulty does.
Further reading
Tagged with:
Ready to Automate Your Business?
Let's discuss how AI automation can deliver measurable ROI for your organization in 90 days or sooner.