AI agents stopped demoing and started working

Amazon introduced continuous workflows within Seller Assistant for third-party sellers on its main marketplace. It runs continuously and acts on seller prompts — alerting when ratings drop, monitoring pricing on specific products — rather than waiting to be asked.

UiPath made its Coding Agents generally available and strengthened governance, data access and deployment controls, on the argument that compliance teams need proof of what agents did, not a description of what they were supposed to do.

Avalara launched Aviator, an agent hub that runs tax and compliance workflows with an audit-defensible record for every action. SiliconANGLE reports that Ema raised a $77 million Series B to deploy "AI employees" across HR, IT and finance.

Four announcements, one shape: continuous execution plus a record of what was executed. The interesting part is not that a model can write a paragraph. It's that software now takes actions in systems that book money, change prices and answer customers.

That's the same shift I lived through at Coop when checkout moved into the customer's own hand: the technology was the smaller half. The larger half was deciding what the machine is allowed to do without asking.

Start where the work is boring, repeated and reversible. Agents earn trust on the unglamorous path before they're allowed near the pricing engine.

Define the approval line in writing before the pilot, not after the first incident. Which actions run silently, which need a human, and who is accountable when the agent is wrong.

Measure cycle time, not model quality. The business case for an agent is days turned into minutes — and if you can't see that on a chart, you bought a demo.

At Stena Line a manual process became a six-second machine learning task. The lesson was not the model — it was redesigning the work around it. Agents are the same: the value appears when you change the process, not when you add a tool to the old one.

That is the core of my AI keynote, and the starting point for advisory work with leadership teams deciding where agents should act first.

Choose a workflow with a clear beginning and end, then map what happens when the agent cannot finish. Who receives the incomplete task? What information travels with it? Can that person resume the work, or must they start again? Make those questions part of the pilot design.

Compare the current process with the proposed one using the same boundaries. Include human approvals, corrections, waiting time and work passed to another team. A fast agent step is not the same as a faster completed workflow. Count a task as complete only when the receiving team accepts the result.

Keep the initial permission narrow: preparing a change is different from publishing it. The trade-off is less autonomy in exchange for a clearer view of errors before they become external commitments. Expand the permission only after reviewing the exception cases.

Keep a human decision point when the action is hard to reverse, the authority is unclear or the exception route is unstaffed. That does not rule out useful automation: the agent can gather information and prepare a recommendation without owning the final action.

What would justify expansion? Set the answer before testing: acceptable completion quality, manageable review effort and an effective way to stop execution. The Stena Line case provides a process-redesign reference, not a promised result for your agent. Read the agent incident analysis alongside the business case.