Which workflows are worth automating first

The first AI automation in an organization does two jobs. It has to deliver value, and it has to establish that this works — which means it needs to be the kind of thing that can be measured without argument.

The scoring criteria we use

• Volume. High enough that a percentage improvement is a real number.
• Variation. Low enough that the edge cases are enumerable rather than infinite.
• Ground truth. A correct outcome that two people would agree on without discussion.
• Cost of error. Tolerable, recoverable, and detectable before it compounds.
• Data access. The inputs and the system of record are both reachable today.
• Owner. A named person who wants this and will act on what it reveals.

Workflows that score well on all six are usually unglamorous: invoice intake, shared-inbox triage, CV screening support, recurring operational reporting, CRM hygiene, document classification. That is a feature, not a compromise.

The traps

Two patterns account for most failed first projects. The first is choosing the workflow with the highest theoretical value regardless of feasibility — typically something touching a system nobody can integrate with. The second is choosing a workflow whose correct outcome is genuinely contested inside the business, so the automation gets blamed for surfacing a disagreement that predates it.

If your team cannot agree on what the right answer is, automating the decision will not produce agreement. It will produce a target.

Instrument from day one

Measure the baseline before you deploy: current cycle time, current error rate, current volume, current exception rate. Without a baseline you will end up in an argument about whether the system helped, and that argument is not winnable after the fact.