Draft, Review, Send: Splitting One Job Into Three So a Human Stays in Charge

You want an agent to handle a repetitive job, but you do not want it emailing a client something wrong. The fix is to stop thinking of the job as one action. Split it into draft, review and send, and put the human at the cut.

The cut goes exactly where consequences begin. Everything before that point is cheap to redo. Everything after it costs an apology.

Find the point of no return

Take any task and ask a single question: at which step does an outsider first see the result, or does something become hard to undo? That step is the send. Everything leading up to it is the draft. The review is the small window you deliberately open between them.

A few examples of where the line falls:

• Invoicing: pulling hours, applying rates and formatting the invoice is draft. Emailing it to the client is send.
• Client replies: reading the thread, checking the contract and writing a response is draft. Hitting reply is send.
• Offer letters: assembling salary, start date, title and terms is draft. Sending it to the candidate is send.

In all three, the expensive part is not the writing. It is the moment the work leaves your building.

What belongs in the draft stage

Give the drafting stage as much of the boring work as it can hold. This is where an agent earns its keep, because mistakes here cost minutes.

Drafting work usually looks like this:

• Gathering facts that already exist somewhere in your records
• Putting them into the shape your team always uses
• Checking the draft against itself for contradictions
• Flagging anything it could not find rather than filling the gap

That last item matters more than it sounds. An agent that leaves a blank and says why is easier to review than one that produces a confident guess, and where the tool allows it, setting a confidence threshold does the same job. Ask for that behaviour from any tool you buy.

Draft quality depends on what the agent can see. An agent working from a thin slice of your data will invent the rest. SicherOne puts project management, HR and AI agents on one set of records for this reason, so an agent drafting something has the surrounding context rather than a fragment.

What review actually means

Review is not reading. Review is deciding.

A useful review step gives the person three things: what the agent produced, what it used to produce it, and what it was unsure about. With those, a reviewer can approve in seconds when the work is routine and slow down when it is not. Deciding which outputs need a second human is the same judgement in another form.

Review works when a person can see what changed and why, not just the finished text. A wall of polished prose gets skimmed. A short diff with sources gets read.

Keep the queue small. If an agent produces twenty items a day for approval, the review decays into clicking inside a week. Two or three well chosen jobs beat twenty half useful ones.

Send stays with a person

The send step should be dull. Somebody with authority for that kind of decision presses the button, and the record shows who it was.

Some rules that hold up across teams:

• One named approver per queue. A queue owned by everyone is checked by nobody, and two approvers are sometimes worse than one.
• The approver must be able to edit before sending, not only approve or reject.
• Rejections should say why in a sentence, because that is your training data for fixing the draft stage.
• The log keeps the draft, the edits and the approver's name.

In SicherOne a human approves agent output before it ships. That is the shape of the product rather than a setting somebody can quietly turn off.

Applying the split to a real job

Take invoicing. The old version is one person doing everything on the last Friday of the month. The split version looks like this:

• Draft: the agent pulls logged hours, applies the rate card, writes line descriptions from task titles, and flags any entry with no task attached.
• Review: the account owner sees the invoice next to the raw hours, fixes the two odd lines, and notes that one entry was already billed.
• Send: the same person sends it, and the record shows the draft, the two edits and the approval.

Nothing here removes the human. It removes forty minutes of assembly and leaves the judgement in place.

Where the pattern breaks

The split fails in two situations. The first is when the review is impossible to do honestly, because checking the output takes as long as producing it. If that is true, the job is not a good candidate yet.

The second is volume. Approval quality falls as the queue grows, and it falls quietly. Watch how long reviews take over a few weeks. When the average approval time collapses, your review stage has turned into a formality, and the split has stopped protecting you.

Start with one job. Cut it at the point of no return. Give the draft to the agent, keep the send, and make the review small enough that people actually do it.