Many executive teams are working through some version of the same question: what should we actually do with AI? Vendors are ready with answers. The demonstrations are impressive, the case studies are polished and the pressure to act is real.
That makes it easy to begin with a product shortlist before anyone has identified the work that needs to change.
A demonstration can show what a tool is capable of. It cannot show whether a particular workflow has the data, ownership or capacity to use it. You have to look at the work itself.
Watch one process from beginning to end
When people describe a process, they usually give you the version in the procedure or the version they remember. A live walkthrough shows the exports to spreadsheets, manual checks, judgment calls and workarounds that never made it into the documentation.
On a recent engagement, we used screen-shared walkthroughs to assess workflows before evaluating tools. In the first session, an investor-facing team performed the work while we asked where information came from, what had to be checked and which exceptions required judgment. The observations were then grouped into themes so we could see the dependencies behind the apparent opportunities.
The walkthrough also exposed dependencies on another business function. That matters because improving one team's task can simply move the problem elsewhere if the upstream process is left untouched.
Separate assistance from accountability
Most workflows mix repeatable tasks with human judgment. Gathering, formatting, cross-checking, summarizing and moving information may be good candidates for assistance. Deciding what matters, weighing an exception and accepting the consequence of a decision are different.
AI may reduce the effort required to prepare a decision without being the right place to make that decision. Mapping the workflow at this level turns a broad ambition into a more useful question: which steps consume meaningful effort, rely on usable data and feed an outcome the business cares about?
The tool comes later. Once the work and its constraints are understood, selection becomes a fit decision rather than a response to the most persuasive demonstration.
Turn the observations into choices
A long list of ideas is useful during discovery, but it is not a plan. Each opportunity worth carrying forward needs an outcome, an owner, evidence that the problem is real, the required data and process changes, and a clear view of what happens if the recommendation is wrong.
Those opportunities must then compete with other work for the same people and attention. A technically feasible pilot can still be the wrong priority if the organization cannot support the process change or measure whether the result improved.
Four questions before any AI pilot
- Have we watched the actual work this would change, recently and end to end?
- Which repetitive steps does it remove, and what judgment does it leave in place?
- Is the data it depends on accurate and available enough, today, to support it?
- Who owns the outcome, and how will we know in ninety days whether it is working?
Thin answers do not mean the idea should be discarded. They identify the work required before a responsible pilot can begin. When the answers are solid, the tool decision becomes narrower and easier to defend.
Darryl Kwan is the founder and principal of Factum Partners. He leads technology and operating transformation from assessment through delivery.