When a team first meets a digital worker, the question usually comes out as "Can we trust it?" It sounds like one question. In practice it is a different question for every task, because a worker's job is a bundle of tasks, and each one carries a different kind of risk.
That is why KAgents sets the mode per task, not per worker. In copilot, the worker prepares the work and a person approves it. In autopilot, the worker finishes the task on its own, inside limits you set, and hands anything outside those limits to a person. The same worker can do both on the same day.
Why the mode belongs to the task
Take Caleb, our customer support worker. Answering "Where is my order?" means looking up the order, reading the delivery status and replying. The answer is either right or easy to correct. Issuing a refund above your limit is different: money leaves the business, and the decision should belong to someone on your team. Same worker, same helpdesk, very different stakes.
The pattern repeats on every desk. Vanessa, the sales worker, can book a meeting when an inbound demo request meets your criteria. A first email to a new target account is another matter. It goes out under your company's name to someone who has never heard from you, so a rep reads it first. Hannah, the finance worker, can send reminders on overdue invoices at the intervals your policy sets. A payment run always waits for a named approver.
If you set one mode for the whole worker, you end up choosing between two bad options. Either everything waits for review, including work nobody needs to check, or everything runs alone, including work that should never skip a person.
Four questions to ask about each task
Four questions do most of the work. None of them needs a scoring model. They need someone who knows the job.
- What happens if it goes wrong, and can you undo it? A draft can be edited. A sent email can't be unsent, and a payment is hard to pull back. The harder an action is to reverse, the longer it should stay in copilot.
- Are the rules written down? Tasks with clear inputs and a documented policy are good autopilot candidates. Tasks where the answer depends on judgment, context or relationships are not, at least not yet.
- Who receives the result? An internal summary posted to your own channel is low exposure. A message to a customer, an applicant or a new contact speaks for your company. An emotionally charged conversation needs a person.
- How alike are the cases? When most cases look the same, review quickly shows whether the worker handles them well. When every case is different, review never gets easier, and the task may belong in copilot for good.
A simple way to map it
Write down every task the worker would take on, one per line. Next to each, note the four answers: reversible or not, rules written or not, who receives the result, routine or varied.
Every task starts in copilot, so the map does not decide where a task begins. It shows which tasks are candidates to move later and which never will be. Tasks that are reversible, rule-based and routine are your first candidates. Tasks that fail two or more of the questions should stay in copilot until your review history says otherwise.
Keep the map next to the worker's rules. It is the plain record of why each task runs the way it does, and it makes the next conversation shorter when someone asks.
Some lines don't move
A few actions sit outside this exercise. In KAgents, payment runs and refunds above your limit always need a named approver. They can't be switched to autopilot. The same goes for decisions about people. Spencer, the HR worker, screens applicants against your criteria and writes a note on each, but a person decides who moves forward and who gets an offer.
You can add your own fixed lines on top. Mark any action as sensitive, such as an offer, a credit or an email to a new contact. It then waits for a named person to approve it, whatever mode the rest of the task runs in.
Moving the line, in both directions
When a task is ready, move that task on its own. Don't move a whole worker at once. Autopilot still runs inside guardrails: the amounts, recipients, channels and hours you set. Cases outside those rules go to a person with a summary of what happened so far. Every action is logged, so you can see what the worker read, what it changed and which rule applied.
The line can move back, too. A new refund policy, a product launch or a busy season can turn a routine task into one that needs a second pair of eyes again. Moving it back to copilot doesn't take a project. You can change the mode for a task anytime. Treat it as normal maintenance, not as a failure.
Over time the map changes, and that is the point. The decision is never made once for the whole worker. It is made task by task, by the people who own the work.
If you have a backlog in mind, bring the task list to a demo. We'll go through it with you, task by task, and mark where each one starts and where it could go next.
Names and glyphs are personas for software agents. They help your team know who handles what.