Human Approval by Role: The Gates That Make an Agent Trustworthy
Human approval by role: how to design gates signed by directors so nothing reaches a client without verification, and your AI agents become truly trustworthy.
The problemYour agents produce fast, but you do not know who approved what reached the client or on what basis.
Tuesday, 4:50 p.m. A creative agency sends its biggest client the visual proposal for a campaign. At 5:30, the reply comes in: the headline mentions a product benefit the client is prohibited from advertising under regulations. The account director asks who approved sending it. Nobody knows for sure. Someone on the team gave an "OK" from their phone between two meetings.
The agency's agents did their job quickly and the copy was well written. What failed was the last gate. This article explains how to design human approval by role, the mechanism that makes an agent system trustworthy without making it slow.
Why "someone should review it" isn't enough
Most companies using AI have some form of human review. The problem is that it tends to be informal: anyone can approve, at any time, with no clear criteria and no record. That creates three risks.
- The wrong person approves. An assistant signs off on something only a director can commit to.
- Things get approved without the important parts being seen. The person receives a ten-page document and approves it on trust.
- There is no accountable owner. When something goes wrong, there is no way to know who signed off or with what information.
A review with no owner, no criteria and no record is not a control. It is a habit.
What a role-signed gate is
A gate is a point in the workflow where work stops until a condition is met. In a well-designed agentic system there are two kinds of gates, in this order:
- Automated verification: an evaluator agent, separate from the one that produced the work, checks it against explicit criteria. We explain this in generator and evaluator.
- Human sign-off by role: a person in a defined role reviews a summary prepared by the system and approves, rejects or requests changes.
The rule we propose is simple: nothing reaches the client without verification plus human sign-off. Verification filters what a machine can check. The sign-off brings the judgment and accountability a machine does not have.
In March 2026, Anthropic published a guide on harness design for long-running agents that includes verification gates between roles. Sign-off by role is the enterprise version of the same idea: every important gate has an owner with a name and a title.
What it looks like at a creative agency
Agency OS, the system we built for a creative agency, runs on 27 agents and 9 workflows, with sign-off from the relevant director at every key step. Here is how some of those gates are assigned:
| Gate | Role that signs | What it verifies | What the system prepares |
|---|---|---|---|
| Brief accepted | Account Director | That the brief reflects what the client asked for, scope and budget | Brief summary, open questions, comparison with the sales proposal |
| Creative concept | Creative Director | That the idea is strong, original and on brand | Finalist concepts, discarded baseline version, evaluator feedback |
| Final copy | Creative Director | Tone, permitted claims, brand spelling | Copy, list of claims and the result of the check against the client's guidelines |
| Production budget | Account Director | That the cost fits what the client approved | Breakdown, variances and alerts |
| Client delivery | Account Director | That everything before it is signed and the package is complete | Final package and a record of every prior sign-off |
The full story of the system is in Agency OS.
How to design your gates, step by step
This framework works in any department, not just agencies:
- Map what leaves the company. Proposals, contracts, campaigns, responses to complaints, payments. Every output is a candidate for a gate.
- Assign a role, not a person. "Creative Director" rather than "Anna." If Anna is on vacation, the system knows who covers that role.
- Define what each role reviews. Three to five criteria per gate. The Account Director does not review typography; the Creative Director does not review margins.
- Prepare the decision. The person should not have to read ten pages. The system gives them a summary, what changed, what the evaluator flagged and what needs their judgment.
- Log every sign-off. Who signed, when, which version and with what comments. This feeds into the system's traces, as we explain in observability for AI agents.
- Define what happens if nobody signs. A deadline, a reminder and a backup. A gate that can stay open forever is a bottleneck.
Illustrative example: the same proposal, with gates
Back to the agency from the opening.
Before: the agent drafts the proposal, someone skims it and sends it. The prohibited claim slips through.
After: the copywriting agent drafts the copy. The evaluator checks it against the client's list of permitted claims, stored in the client's profile in the brain, and flags the prohibited benefit. The agent fixes it. The Creative Director sees in their dashboard the copy, the correction made and confirmation that every claim is permitted. They sign off. The Account Director receives the complete package with the creative sign-off on record, checks that everything matches the approved brief and signs off on delivery.
If the client asks tomorrow who approved the campaign, the answer takes seconds: it is in the log, with name, role, time and version.
Common mistakes
- Requiring sign-off on everything. If every step needs approval, people start signing without looking. Reserve gates for what leaves the company, makes a commitment or cannot be undone.
- Approving from the wrong summary. If the system does not show the signer what they need to see, the sign-off loses its value.
- Removing automated verification. Asking a director to catch typos wastes their judgment.
- Not reviewing rejections. Every rejection is information: it tells you which rule or which context the agent is missing.
What to do on Monday
List everything your company sent to clients last week with help from AI. For each one, answer: who approved it, in what role, and where was it recorded? If the answer is "I'm not sure" for more than one, you know where to put your first gate.
At Addabra, we design every system with human approval by role from the start: the agents prepare, the evaluator verifies and the right director signs off. That way you get the speed of agents without giving up control over anything that carries your company's name.
Frequently asked questions
What is human approval by role?
It is a design in which certain decisions in an agent system only move forward with the sign-off of a person in a defined role, such as Account Director or Creative Director. It is not enough for someone to approve: the approval has to come from whoever is accountable for that type of decision.
Don't human approvals slow the process down?
Designed well, they make it faster. The agents prepare everything, the evaluator filters out errors and the person receives a clear summary to decide in minutes. What is slow is fixing something after it reached the client without review.
Which decisions should require human sign-off?
As a general rule, anything that leaves the company for a client, anything that commits money or deadlines, and anything with legal or brand implications. Internal, reversible tasks can proceed on their own, as long as they are logged.
