An approval is a step. It has an owner, a deadline, and a fallback if nobody answers. It is not a comment the automation hopes someone will notice, and it is not a chatbot message with no record of the yes.
If a workflow can spend money, message a customer, change a record people rely on, or close a request for good, someone should have to say so where the next run can see it.
Which decisions belong to a person
Not every step needs a human. A person should own the decision when a wrong yes is hard to undo.
That usually means a commitment outside the team, a change to a system of record, a payment or a credit, or a case where two sources disagree. It also means anything your policy already says a named role must sign.
Leave the rest on the automated path. Copying a structured field, filing a known attachment, or posting a status that only restates a decision a person already made. Write the decision in one line. "May we send this date to the customer?" is a decision. "Review the thread" is not.
The anatomy of an approval step
A usable approval step has six parts.
The ask is one sentence. It says what will happen if they say yes.
The packet is the facts they need, not the whole history. The draft, the source message, the record it will change, and what is still unknown.
The owner is a role or a named person, not a channel.
The choices are yes, no, and edit. That is enough for most ops work. A free-text "any thoughts?" recreates the chase.
The clock is when the ask expires.
The fallback is what the workflow does if the clock runs out. A backup owner, a safe stop, or a smaller action that does not need the yes. Never a silent continue.
If any of those are missing, you have a notification, not an approval.
Who gets asked, and what they need to see
Ask the person accountable for the outcome, not the person who is easiest to ping. A backup is for absence, not for convenience.
Slack or email can carry the ping. The yes itself should land on the workflow, so the answer is not trapped in a reaction emoji.
The packet should be short enough to read in a minute. Include the draft if a message will go out. Include the before and after if a record will change. Flag the one field that is uncertain.
Timeouts, reminders, and a fallback owner
An approval without a clock becomes a stuck run that looks like success. The message went out. The work did not move.
Pick a window that matches the decision. A routine handoff might be four business hours. A change that can wait until morning should say so. "As soon as possible" is not a window.
Send one reminder before the deadline, in the same place, with the same ask. If the clock runs out, tell the fallback owner. Do not ping the whole channel.
Decide the timeout before you turn the workflow on. A safe default is to stop, mark the run as waiting on a backup, and do not send the external message. Proceeding because a timer fired hides the miss.
What the agent may prepare, and what it may not send
An agent is good at assembling the packet. It can pull the last reply, draft the note, and point at the field that does not match. It can suggest yes or no, with a reason the approver can check.
It should not send the external message, change the committed record, or close the loop until the approval step returns a yes. A polished paragraph is still a draft.
Also block the agent from widening the ask. If the approval was "send this date," the sent text should be that date, not a new promise the model added. The approver signs a specific action.
How to log the decision so the next run can see it
Write the decision next to the run. Who answered, which choice, the time, and the exact text or change they approved. If they edited the draft, store the edited version, not only the original.
That log is what a later step reads. It is also what you want when someone asks why a message went out. A Slack history is not enough. Threads get truncated, and people leave.
Store a one-line reason when the answer is no, so the next owner is not starting from zero.
A small checklist before you turn the workflow on
Read this once with the builder and the person who will actually approve.
The decision is one sentence. The packet fits on one screen. The owner is a role you can name today. Yes, no, and edit are the only choices. The clock is a real time, not "soon." One reminder goes to the owner. Timeout assigns a backup or stops. The agent can draft and cannot send. The log stores who, when, and what was approved. A test with nobody answering ends in the fallback, not in a customer message.
Innflow is one way to put that approval on a visual workflow, with the agent steps beside the human step, on top of email, Slack, and the tools you already use. See https://innflow.ai.












