Before choosing an AI agent platform, track the work you want the agent to improve. The best vendor in the wrong workflow still produces a weak pilot. The simplest worker in the right workflow can create real leverage.
Start with a one-page operating readout. It should explain the workflow, frequency, owner time, handoff points, data sources, approval rules, and the business signal that should improve.
Track workflow frequency.
Work that happens once a quarter rarely makes a good first AI Worker. Work that happens every day or every week gives the team enough repetitions to learn, measure, and improve the lane. Frequency creates proof faster.
Track owner time and repair time.
Do not only estimate how long the task takes when it goes well. Track cleanup, rework, waiting, coordination, and follow-up. Many valuable AI lanes reduce repair time more than they reduce first-draft time.
Track approval load.
If the workflow already requires review, design the worker around that reality. Track how many items need approval, how often drafts are accepted, why they are rejected, and which claims or decisions require escalation.
Frequency, owner time, repair time, data readiness, failure cost, approval load, conversion impact.
These are enough to decide whether the first lane should be bought, built, scoped smaller, or deferred.
Track failure cost.
Some errors are annoying. Others create legal, financial, compliance, brand, or customer damage. If the failure cost is high, start with draft-only or review-required authority. If the failure cost is low and the lane is repetitive, the worker can earn more autonomy later.
Track conversion impact.
For marketing and sales workflows, the metric should eventually connect to pipeline, signup, response, or meeting movement. If the worker makes more content but does not improve the path to demand, it is probably activity rather than growth.