The Send Gate Is Part of the Operating System

The first obvious risk in AI workflows is bad output.

The more expensive risk arrives after the output gets better. The draft looks polished. The account brief sounds plausible. The customer response reads cleanly.

The work feels ready to move.

That is the moment a workflow can quietly cross from assistance into action.

A draft becomes an email. A support summary becomes a customer response. A roadmap synthesis becomes a product commitment. A finance recommendation becomes a budget move. A deployment suggestion becomes a production change.

The problem is not that AI produced the artifact. The problem is that nobody defined the line between drafting and acting. Nobody chose where the result would be logged either.

That boundary is the send gate.

The failure pattern

Most teams design the generation step before they design the authority model.

They can answer what prompt creates the output, which model is used, which template it follows, and how fast the workflow runs.

But they cannot answer:

• Who decides whether this output may leave the system?
• Which channels are allowed?
• What evidence must be checked before action?
• What requires escalation?
• What contexts are always human-only?
• Where is the outcome logged?
• Who reviews failures or near misses?

That is an operating-system gap: production capability without execution governance, durable memory, or a clear owner.

The operator lesson

Prospecting makes the boundary obvious.

An agent can research companies, organize public signals, draft a buyer-facing artifact, and run a quality review. That is useful work. It is not authority to send.

The stronger workflow separates roles:

• the agent researches and drafts;
• the human chooses whether the target, contact, and channel are valid;
• the human checks evidence, assumptions, tone, and risk;
• the human decides to send, revise, hold, or discard;
• the outcome is logged so the workflow improves.

The lesson applies anywhere AI output can affect another person, move money, create a commitment, or change a production system.

Quality gate versus send gate

A quality gate asks:

Is this output good enough to use?

A send gate asks:

Who is allowed to use it, through which channel, under what conditions, and where will the result be logged?

Those are related, not identical. The quality gate checks evidence, assumptions, risk, relevance, owner clarity, and learning. The send gate sits closer to consequence. It decides whether the artifact can leave the workflow, update a system, trigger a commitment, notify a customer, or change a priority.

A good draft can still need a human approval boundary. That is not a bottleneck. It is part of the operating system.

The six parts of a send gate

1. Action class: what action could this output trigger?
2. Human owner: who can approve, revise, reject, or escalate?
3. Channel boundary: where may this action happen, and where is it prohibited?
4. Evidence threshold: what must be verified before action is allowed?
5. Escalation rule: what conditions force a pause?
6. Outcome log: where is the result recorded and reviewed?

The gate does not have to be heavy. It has to be explicit.

A lightweight approval gate

Use this before AI output becomes external or material internal action:

• Output: what did AI produce?
• Possible action: what could this trigger?
• Affected party or system: who or what could be affected?
• Human owner: who approves or rejects?
• Channel: where may this action happen?
• Evidence: what facts were verified?
• Risk check: what could be wrong, sensitive, stale, confidential, or overclaimed?
• Escalation: what forces a pause?
• Decision: send/use, revise, hold, reject, or escalate?
• Outcome log: where will the result be reviewed?

This is not a committee process. It keeps action rights visible and makes the outcome reviewable.

One action this week

Pick one AI-assisted workflow where output is already close to action. Before improving the prompt or adding automation, write the send gate:

1. What action can the output trigger?
2. Who has approval authority?
3. Which channels are allowed without extra approval?
4. What evidence must be checked?
5. What conditions force escalation or a hard stop?
6. Where will the outcome be logged?

If those answers are missing, the workflow is draft-capable but not operating-system-ready.

For a broader starting point, use the AI workflow inventory template. It will surface owners, source gaps, decision rights, and the next decisions. If AI output is moving faster than your approval model, map your company brain. The diagnostic turns sprawl into a 90-day operating plan with explicit decision rights and approval boundaries.