# Human approval for AI agents: what a small business should sign off on

> Which AI agent actions need sign-off, how to rank the risk, what Law 25 and Canada's AI code say, and how long review really takes in a small service business.

Updated on October 15, 2026 · https://app.pilotiko.com/en/guides/human-approval-for-ai-agents

An AI agent can draft a reply, a quote or a schedule on its own, but a person should approve any action that touches a customer, money or an outgoing message. Rank actions by risk, decide who approves what, and make sure the review stays real instead of becoming a rubber stamp.

## In short
- Rank every agent action by risk: internal, reversible, binding or irreversible.
- In Quebec, section 12.1 of the private-sector privacy act governs decisions based exclusively on automated processing.
- Canada's voluntary code of conduct on generative AI includes human oversight and monitoring after deployment.
- A 20-second review of a drafted message costs little; a mistake sent to hundreds of customers costs a lot.

## What does human approval look like in practice?

It is a checkpoint: the agent prepares the action and parks it, and nothing goes out until a designated person accepts, edits or rejects it.

An AI agent reads your data (calendar, customer records, history) and proposes an action. Without a checkpoint, the action runs. With one, it lands in a queue showing what the agent looked at and why it is proposing this. The person approves in one click, edits the text or turns it down.

You do not need a checkpoint everywhere. An agent that prepares an internal report or reorders a task list carries almost no risk. An agent that sends a quote or grants a credit does. The skill lies in putting approval in the right places.

## Which AI agent actions should require approval?

Anything that leaves the business or commits money. Internal, reversible actions can run on their own as long as they are logged.

| Agent action | Risk level | Suggested approval |
|---|---|---|
| Summarizing a visit in the internal file | Internal, reversible | None; reviewable log |
| Proposing a technician for a call | Internal, reversible | Confirmed by the dispatcher |
| Drafting a reply to a customer | External | Approval before sending |
| Preparing a quote with a price | External, commits to a price | Approval by someone allowed to set prices |
| Scheduling a social media post | External, public | Approval before publishing |
| Messaging a customer list | External, bulk | Approval plus a sample check |
| Granting a refund or credit | Money, irreversible | Human only; the agent can prepare the file |
| Refusing service to a customer | Decision affecting a person | Human only |

Indicative grid suggested by Pilotiko. The level rises with the number of people affected and how hard the mistake is to undo.

## What does Law 25 say about automated decisions?

Section 12.1 requires you to tell people when a decision about them is based exclusively on automated processing, and to let them be heard by someone who can review it.

Section 12.1 of Quebec's Act respecting the protection of personal information in the private sector applies to any business that uses personal information to render a decision based exclusively on automated processing [1]. The business must inform the person no later than when it informs them of the decision. On request, it must also tell them which personal information was used, the reasons and main factors behind the decision, and their right to have the information corrected [1].

Finally, the person must be given the opportunity to submit observations to a staff member who is in a position to review the decision [1]. The Commission d'accès à l'information lists this rule among the main changes brought in by Law 25 [3]. Genuine human approval before the decision is communicated means, in practice, that it is not exclusively automated. For the rest of the law, see [Law 25 for service businesses](https://app.pilotiko.com/en/guides/quebec-law-25-for-service-businesses).

## What does Canada's generative AI code recommend?

The voluntary code published by ISED in September 2023 sets out six principles, including human oversight and monitoring: system use is monitored after deployment and updated as needed.

The Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems is aimed first at the companies that build or operate those systems [2]. Its six principles are still useful to a small business that uses them: accountability, safety, fairness and equity, transparency, human oversight and monitoring, and validity and robustness [2].

On content, Google likewise stresses that it is critical to manually fact-check all AI-generated content before publishing, including titles and descriptions [4]. The same discipline applies to a customer message drafted by an agent.

## How long does human review actually take?

Not long, if the agent does the preparation. Reviewing a ready message takes seconds; writing it takes minutes.

**Example: 40 customer messages a day**

Assumptions: a small business answers 40 customer requests a day (scheduling, simple quotes, follow-ups). Writing a reply from scratch takes 3 minutes. Reviewing and approving an agent-drafted reply takes 20 seconds, and one in five needs a 1-minute fix.

- Writing by hand (40 × 3 min) : 120 min a day
- Reviewing drafted replies (40 × 20 s) : 800 s, or about 13 min
- Fixes (8 replies × 1 min) : 8 min
- Total with agent and approval (13 + 8) : about 21 min a day
- **= About 99 minutes of capacity freed per day under these assumptions, with no message going out unreviewed**

## How do you stop approval from becoming a rubber stamp?

Cap the volume per reviewer, show what the agent based its proposal on, sample bulk sends and track the correction rate.

- **Show the inputs.** The approval screen should display the data the agent read: customer record, booking, price. People approve better what they can check.
- **Track corrections.** If the correction rate sits at zero for weeks, either the agent is excellent or nobody is reading anymore. Run a spot check.
- **Keep queues short.** After a few dozen approvals in a row, attention drops. Split the queue between two people.
- **Sample bulk sends.** For a message going to 300 customers, review the template and at least five personalized versions.
- **Keep a log.** Who approved what, and when: that is what lets you fix a rule after a mistake.

## How do you set up approvals in a small business?

1. **List the actions** — Write down everything your agents or tools do or could do: draft, schedule, send, invoice, refund.
2. **Rank them by risk** — Place each action in the grid: internal, external, commits to a price, money, or a decision affecting a person.
3. **Name the approvers** — For each level, decide who can approve. A quote is not approved by the same person as a visit summary.
4. **Inform customers** — If a decision can be based solely on automated processing, prepare the notice section 12.1 requires and name the person who can review it.
5. **Review monthly** — Look at the correction rate and the errors caught. Loosen or tighten the grid based on what you see.

## How does Pilotiko handle human approval?

It is a core rule of the product: AI agents read your data and prepare the work; anything touching a customer, money or an outgoing message waits for a person's approval.

Pilotiko's AI agents work in the virtual office, where each department is a room. What they prepare shows up in the relevant room, waiting for a decision. Content drafting follows the same rule: posts are drafted and scheduled, then approved by you.

To place approval within a full automation project, read [Automating a service business](https://app.pilotiko.com/en/guides/automate-a-service-business). To cut no-shows with reminders, see [Reduce no-shows](https://app.pilotiko.com/en/guides/reduce-no-shows). To compare plans, see [pricing](https://app.pilotiko.com/en/pricing).

> Coming soon in Pilotiko: the [AI receptionist](https://app.pilotiko.com/en/product/ai-receptionist) and email marketing. Confirmations and reminders are being commissioned: nothing is sent yet.

## Frequently asked questions

### Can an AI agent send emails without approval?

Technically yes, but it is unwise for any message that commits the business. A few seconds of approval keeps a mistake from reaching hundreds of customers. In Pilotiko, every outgoing message waits for a person's approval.

### Does human approval slow down customer service?

Slightly, if the queue is watched during business hours. The writing time saved more than makes up for a few minutes of waiting. For emergencies, have an on-call person approve first.

### Who should approve an agent's actions?

Whoever would have done the task without the agent, or their manager for riskier actions. A quote goes to someone who can set prices. A refund goes to someone who can commit the company's money.

### Does Law 25 ban AI agents?

No. It governs decisions based exclusively on automated processing, the use of personal information and transfers outside Quebec. An agent that prepares and a person who decides fits within those rules.

### Should I tell customers an AI agent drafted the reply?

If a decision about them is based exclusively on automated processing, section 12.1 requires you to tell them. Otherwise, transparency is still good practice, and Canada's voluntary code recommends it too.

### What if an agent gets something wrong despite approval?

Fix it with the customer, then check the log: which data was wrong, which rule was missing, why the review missed it. Adjust the risk grid or the data, not just the wording.

## Sources

1. LégisQuébec, « Act respecting the protection of personal information in the private sector (CQLR, c. P-39.1) », https://www.legisquebec.gouv.qc.ca/en/document/cs/P-39.1 (accessed on 2026-10-15)
2. Innovation, Science and Economic Development Canada, « Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems (septembre 2023) », https://ised-isde.canada.ca/site/ised/en/voluntary-code-conduct-responsible-development-and-management-advanced-generative-ai-systems (accessed on 2026-10-15)
3. Commission d'accès à l'information du Québec, « Principaux changements apportés par la Loi 25 », https://www.cai.gouv.qc.ca/protection-renseignements-personnels/sujets-et-domaines-dinteret/principaux-changements-loi-25 (accessed on 2026-10-15)
4. Google Search Central, « Using generative AI content », https://developers.google.com/search/docs/fundamentals/using-gen-ai-content (accessed on 2026-10-15)