Home·Insights·Assessing an AI opportunity before spending

Assessing an AI opportunity before spending

A two-hour assessment that kills a bad case is worth more than a six-month pilot that should have died in a meeting.

Digital & AI2026-10-016 min readAtlas Strategy Group

The expensive way to find out whether an AI opportunity is real is to run the pilot. The cheap way is to assess it first — a deliberate two-hour exercise that answers, on paper, most of what the pilot would answer in two quarters. Most companies skip it, because assessing feels like delay and pilots feel like progress. The arithmetic runs the other way: a pilot costs months and credibility; an assessment costs an afternoon and one uncomfortable conversation.

What the assessment actually asks

Four numbers and one admission. The volume: how many times does the task occur — per day, per operator, per year? Automation of a task that happens twice a month is a rounding error wearing a demo. The human cost: what does each occurrence cost in time, wages and error cleanup? The variance: how standardized are the occurrences? Ten variants the model handles reliably are worth more than a thousand variants it handles approximately. The savings threshold: what share of the occurrences would the system need to process without human review before the case pays for itself — and is that share plausible, or is it the optimistic edge of a marketing curve?

The admission is the hard part: what happens to the last ten percent. A system that handles ninety percent of occurrences still needs a human ready for the tenth percent — which means the savings are not ninety percent of the cost; they are the cost, minus the price of keeping the exception handler. Cases die here more often than in any technical review, and the assessment is where they die cheaply.

Two traps the assessment catches

  • Automating a broken process. If the process produces rework, disputes and workarounds today, automation will produce them faster. The assessment asks the question the demo never raises: is this process worth automating at all, or does it need to be fixed first — and is the fix itself the real project?
  • The exception masquerading as the rule. Teams describe their work by its most interesting cases; automation lives in the boring majority. The assessment counts occurrences from logs and timesheets, not from interviews — interviews describe the exceptions, the logs describe the work.
Ask for the numbers before the pilot. If nobody can produce the volume, the cost or the variance, that is the assessment's answer: the company does not yet know the process well enough to automate it.

The assessment is deliberately boring — four numbers, one admission, a page. It is also the highest-leverage page in the AI strategy, because it converts enthusiasm into arithmetic before enthusiasm has a budget. It is how the digital and AI transformation practice opens an engagement, and the survivors of the assessment go to the filter described in use case selection.

All insights →