Home·Insights·Choosing AI use cases: the filter that kills nine of ten

Choosing AI use cases: the filter that kills nine of ten

Idea lists are cheap — any workshop produces twenty. The strategy is the filter that decides which one earns a pilot.

Digital & AI2026-10-016 min readAtlas Strategy Group

Every AI workshop produces the same artifact: a flipchart of twenty ideas, greeted with the enthusiasm only free things inspire. Six months later, the same company is running four pilots with no budget, no owner and no end condition — and the difference between the companies that got value and the companies that got tired is not the quality of the ideas. It is the filter that came after the flipchart.

The five gates

  • Economic weight. Is the task's cost base large enough that even excellent automation of it matters? A brilliant solution to a small cost is a hobby. The candidate list is ranked against the company's actual cost lines, not against what is fashionable to automate.
  • Readiness of the ground. Data exists, is accessible and means the same thing in every system it lives in; the process is structured enough that its steps can be described without argument. Automating an unstructured process does not produce results — it produces arbitration.
  • Measurability. Can the before and the after be read in numbers the company already trusts — hours, error rates, throughput? A case whose value cannot be measured will be «believed successful» and cannot be scaled on evidence.
  • Cost of error. What happens when the system is wrong — and it will be? Cases where a wrong output is cheap to catch are fundable; cases where a wrong output quietly corrupts something expensive are not, whatever the demo looked like.
  • An owner who wants it. A named person whose week this improves, who will defend the change in the process it disturbs. Automation without an internal beneficiary dies politely, in committee.

How the filter is used

Each candidate goes through the gates in order, and the order matters: weight first, because the gates that follow cost effort; a case that survives all five is not «approved» — it is admitted to a pilot with a defined size, a defined metric and a written kill-condition. Cases that fail one gate are not discarded forever; they are parked with the failed gate named. The parking lot is the strategy's memory: next year, when the data or the owner changes, the case re-enters through the one gate it failed, not from scratch.

Nine of ten ideas should die at the filter. That is the filter working. The tenth arrives at its pilot with a budget, a metric and an owner — which is the only way pilots end in anything but fatigue.

The filter is the second page of the one-page AI strategy, and the question it serves is the one we opened earlier: when automation actually reaches the P&L. Running it is standard opening work in the digital and AI transformation practice.

All insights →