An assistant takes a question and returns an answer; a person decides what to do with it. An agent takes a goal and does the steps: reads the file, checks the rules, drafts the letter, updates the record, schedules the follow-up. That is where the hours are, and that is where the risk is. Every vendor now sells “agents.” Very few of them will tell you which rung of the ladder their product sits on.
The autonomy ladder
How much an agent may do on its own is a setting, not a fact. The useful way to talk about it is a ladder with three rungs. Rung one: it drafts, and a person sends. Rung two: it acts, and a person reviews the log. Rung three: it acts, and only escalates exceptions. Most organizations should start on the first rung with one workflow and move up only when the log has earned it. We put the rung in writing before we build anything.
Where agents earn their keep
The good first workflows share a shape: repetitive, rule-bound, high volume, and low stakes per item. Remittance reconciliation. Intake forms that need to be checked against a checklist. Apprentice scheduling and reminders. Drafting routine letters from a template. Answering staff questions from your own manuals. Each of these can be handed to an agent on rung one in a few weeks, with a person still pressing send.
Where they should not start
Anything that decides about a person: eligibility, discipline, selection, dispatch order, benefit amounts. An agent may prepare the file and flag what looks off. It should not decide, and in several states from January 2027 it legally cannot decide without notice and human review. Keep those workflows on rung one indefinitely.
Three things every agent build needs
First, an approval gate before anything that affects a person, a record or a dollar. Second, a log a non-technical manager can read: what it saw, what it did, what it asked. Third, an off switch that does not require a phone call to the vendor. If a proposal does not name all three, ask for them in writing before you sign.
What to ask a vendor
Which rung does your product run on by default, and can I change it? What does the log look like, and who can read it? What happens when the agent is wrong? Who owns the prompts and configuration if we leave? Those four questions separate a governed agent from a demo.
Share this note
Forward the link, or email it to the person who should read it: send by email.
