The Difference Between an Agent Making a Decision and Preparing One

An agent that drafts a reply to a customer is preparing a decision. An agent that sends that reply has made one. The words look similar in a project plan and the difference is everything, because only one of them leaves you somewhere to stand when the output is wrong.

Here is the test, and it takes under a minute. Ask: if this output is wrong, who finds out first, and can they stop it? If a person sees it before anyone outside the team does, the agent prepared a decision. If the world sees it first, the agent made one.

Why managers blur the line

The blur is usually accidental. It happens through convenience rather than intent.

A team sets up an agent to draft weekly updates. A person reviews each one. After a month the drafts are good, review becomes a glance, and by week six somebody adds a rule that sends the update automatically on Friday if nobody has objected by Thursday. Nothing was decided. The agent now makes a decision every week.

Silence is the mechanism. Any workflow where doing nothing means the output ships has moved the agent from preparing to deciding, whatever the documentation says.

The one minute test, written out

Take any task you are thinking of handing to an agent and answer four questions.

• Does the output leave your team without a person reading it? If yes, the agent is deciding.
• If nobody responds, does the output ship anyway? If yes, the agent is deciding.
• Can the effect be undone in a few minutes by one person? If no, treat it as a decision even when a human approves it.
• Would you be comfortable telling a customer or a regulator exactly how this output was produced? If not, you already know the answer.

The third question catches the case people miss. A human approval on an irreversible action is still a decision made under time pressure by somebody who trusted the draft. That needs more than a click.

What preparing well looks like

Preparation is not a lesser job. Done properly it is where most of the value sits.

An agent preparing a decision should give the person three things: the draft or recommendation, the facts it used, and what it was unsure about. The third is the one teams forget to ask for. A recommendation with no stated uncertainty invites a rubber stamp, because there is nothing to check.

Good preparation looks like this in practice:

• The summary comes with links to the records it was built from
• Anything the agent could not verify is flagged rather than smoothed over
• Options are presented with what would change if you picked each one
• The reviewer can see what was left out, not just what was included

Products built this way, SicherOne among them, keep a human approving agent output before it ships. That approval is only meaningful when the agent has made itself checkable.

Where the line should sit

Some rules of thumb that survive contact with real work.

Internal and reversible: an agent can go a long way. Drafting, sorting, summarising, flagging, chasing. A mistake costs a few minutes.

External or hard to reverse: preparation only, which is roughly how departments rank when you order them by risk. Messages to customers, anything with a price on it, anything that changes a record other systems depend on, anything touching someone's employment or money.

Permanent: a person decides, and for the worst cases two people decide. Deleting records, closing accounts, and anything that cannot be walked back deserve a second signature regardless of how good the draft was.

The honest version of the trade

Preparing rather than deciding costs you speed. That is real. If a person must read every output, the throughput of the workflow is capped by that person's attention, and you have not removed a bottleneck so much as moved it.

The answer is to be choosy about what you automate rather than to loosen the rule. Pick tasks where preparation is most of the work and the check is quick. A summary you can verify in thirty seconds is a good candidate. A financial reconciliation you would have to redo by hand to trust is not, and neither is a performance review draft, where the reading is the whole job.

One question to keep asking

Every few months, go back through the agent workflows you run and ask which ones have quietly drifted. Hiring deserves a second look, because some steps there are safe to delegate and some are not. Look for anything that ships on a timer, anything where the approval has become a habit, and anything where the reviewer could not tell you what the agent was uncertain about.

The line between preparing and deciding does not move by itself. It moves because somebody was busy and it was easier. Knowing that is most of the defence.