AI agents can now browse, compare, draft and act across several steps without someone typing each prompt. The pitch is an agent that “runs your marketing”. The useful question is much narrower: which job can it do reliably, with the information and authority you’ve given it?
Start with a job, not a department
“Manage our social media” isn’t a task anyone can hold an agent to. “Every Monday, check these competitors’ prices on these products and flag any change” is. Define the inputs, the expected output, the checks and the actions that need a person before you connect a single tool.
Design for the exceptions
The routine cases are easy. The damage happens in the exceptions: a price that excludes delivery, a listing that’s out of stock, two product names that look alike but aren’t. A good agent workflow makes those uncertainties visible instead of resolving them with a confident guess. When a source can’t be read, that’s a reason to stop, not to fill the gap.
The same lesson applies without AI. At AdvanceQuip, when two excavator models sold out mid-campaign, the right move was to stop their ads straight away. A process that kept running on yesterday’s assumptions would have spent money advertising machines we couldn’t sell. Automation needs the same off switch.
Separate recommending from doing
Summarizing competitor offers is low risk. Changing your own prices, publishing an ad or committing budget is not. Keep those as separate levels of permission, and make sure the person approving can see the sources and understand the limits.
Earn autonomy with evidence
Before giving a workflow more freedom, run it alongside the manual process on real examples, and record the errors and how long the review takes. Autonomy should be earned by category, this kind of task at this level of risk, not granted because a demo looked impressive. It’s the principle behind my free Blueprints: the machine proposes, a person approves.