TL;DR: Let AI agents do the work, and let people make the decisions that commit the company. Approvals belong on anything that touches money, customers or commitments: a price, a refund, a contract, a message that promises something. Everything else can run on its own. Designed well, an approval takes one tap, and it is the reason a team ends up trusting agents with more.
"People decide, agents execute" is the first principle we build on at Instanix. This is what it means in practice and how I would apply it in any business.
Why do AI agents need human approval at all?
Because an agent that acts can act wrongly. Language models are good at understanding and drafting, and they are confident even when they are wrong. In a demo that is a curiosity. In a business it is a customer receiving the wrong price.
Approval is how you get the speed of an agent with the judgement of a person. The agent prepares everything. The person decides in seconds. The customer gets a fast and correct answer.
Which actions need approval?
A simple rule works for most companies: approve anything you would not let a new employee do alone in their first week.
| Usually needs approval | Usually runs on its own |
|---|---|
| Sending a price or a discount | Logging a new lead in the CRM |
| Refunds and payments | Sending an internal reminder |
| Contracts and renewals | Summarising a document |
| Promises about dates or delivery | Checking stock or availability |
| First reply to an important client | Updating a task's status |
| Deleting or changing records in bulk | Preparing a draft report |
The list will differ by company. What matters is that someone decides it before the agent goes live, not after the first mistake.
How do you design approvals that don't slow the team down?
Bad approval flows are why many teams give up on human-in-the-loop. A good one has four properties:
- It comes to the person, where they already are. A notification, a WhatsApp message, a button in the tool they use. Not a separate dashboard they forget to open.
- Everything is prepared. The draft, the context, the customer history. The person reads, decides, done.
- One tap to approve, one to edit, one to reject. No forms.
- It is recorded. Who approved, when, and what exactly went out.
When approvals are this light, managers stop seeing them as work and start seeing them as control.
Does approval mean the AI isn't really autonomous?
It means autonomy is earned. When a team sees that an agent's drafts are consistently right for a certain type of action, they can choose to let that type run on its own and keep approving the rest. Trust grows from evidence, one action type at a time.
This is exactly how people are trusted in a company. A new hire gets more responsibility as they prove themselves. Agents should be treated the same way.
What does this look like in a real workflow?
A customer asks on WhatsApp whether a car is available and what the final price would be.
- An agent reads the message, finds the car and the customer, and drafts a reply with the price and two viewing times.
- Because the reply contains a price, it waits for the sales manager.
- The manager sees the draft on their phone, checks the price, taps approve.
- The reply goes out, the viewing is booked and the CRM is updated, all logged.
The customer gets an answer in minutes instead of the next morning. The manager stays in control of the price. You can watch a flow like this run on the home page of this site.
Frequently asked questions
What is human-in-the-loop AI? A design where AI prepares and executes work, but a person approves specific decisions before they take effect.
Which tasks should never be fully automated? Anything that commits the company: prices, payments, contracts and promises to customers, at least until there is a strong track record.
Does approval add a lot of delay? Usually seconds to minutes, because the person only reviews a prepared draft. Compare that with the hours a request waits when everything is manual.
How does this relate to reliable AI? Approval is one part of it, together with limited permissions, monitoring and logs. I cover the full picture in Why AI in business must be as reliable as infrastructure.
Who is Amir Diab? I'm the founder of Instanix, a Dubai company that builds AI agents with human approval built in. More about me.
Want this working in your business?
Tell me which task eats your team's week. I'll tell you honestly whether an agent can take it on.