Skip to content

/articlesConcepts

Human approval before an assistant acts

Human-in-the-loop means an AI assistant asks a person before risky actions, like sending a quote or an email. Which steps should wait, and what approvers see.

Updated
Published
min read
06
On this page
09
On this page09

What does human-in-the-loop mean?

A person takes part in a decision before the AI assistant acts on it. The assistant proposes one specific step, stops, and someone on your team approves or rejects it.

Take a Riyadh office-supplies store. A customer asks on WhatsApp for 20 cartons of A4 paper. The assistant checks the price in Odoo and prepares a quote for SAR 2,900. Before the quote is created, a salesperson sees it and taps Approve.

This is different from watching after the fact, where people review actions that already happened. An approval sits between the plan and the result, the one point where stopping a mistake costs nothing.

Good approvals are selective. Looking up a price runs straight away. Sending an email, creating a quote or changing a record can wait for a person.

Why does an AI assistant need approvals?

Some actions cannot be undone. A sent email, a quote with the wrong price or a cancelled order reaches real people at once.

Assistants can also be misled. They read emails, messages and web pages written by outsiders, and some of that text may try to steer them. A person who checks the actual step before it happens stops the worst outcomes.

Someone must own the decision. An approval links the action to a named person at a known time.

Rules are moving the same way. Article 14 of the EU AI Act requires high-risk AI systems to be built so people can oversee them and stop them. Most business assistants are not high-risk under the Act, but approvals are the practical way to show a person, not a model, allowed the action.

Which steps should wait for approval?

Give every action a clear default, then let your rules handle the middle ground. Three settings cover almost every case.

  • When two rules disagree, the stricter one wins: block beats ask, and ask beats allow.
  • One assistant can be made stricter than your general rules, but never looser.
  • An action set to always ask cannot be switched to run on its own by any rule.
SettingWhat it meansExamples
Runs on its ownNo approval; for actions that only readCheck a price, read an order status, search a mailbox
Follows your rulesYour rules decide per action and per assistantCreate a quote in Odoo, save an email draft, add a label
Always asksWaits for a person every timeSend, reply to or forward an email from Gmail or Outlook

How Botify decides when to ask

Botify sorts each action by what it can do: read, change data, delete, or send something outside your company. Reading runs on its own. Sending and deleting wait for approval unless your rules say otherwise, and some sends, such as email from Gmail or Outlook, always wait.

On your website chat, where anyone can type, sending outside and deleting always need approval.

Inside Odoo, when an employee asks the assistant to change a record, the employee confirms the change first by default.

For a quote that customers request on WhatsApp, you choose: turn on approval for that assistant, or add a rule for your whole workspace.

What should the approver see?

Enough to decide in seconds without opening another system. A good approval request answers: what will happen, to whom, with what content, and how risky it is.

  • One line that says what will happen: “Create a quote: 20 cartons of A4 paper, total SAR 2,900.”
  • Who it affects: the customer, the email recipients, or the record that changes.
  • The exact content, with passwords and keys hidden.
  • Two clear buttons: approve or reject.

Where does your team approve?

In Botify, every request that needs a person lands in the Inbox in the console. Team members with the right role approve or reject from there.

On WhatsApp and Telegram, the approval can also come as buttons in the chat, so a manager can approve from their phone. Requests from Slack go to the Inbox.

How does an assistant pause and resume?

Agent frameworks call the pause an interrupt. In LangGraph, for example, the run stops, its state is saved, and it continues from the same point once a person answers. Its documentation warns that the paused step starts again from the top when it resumes.

That detail matters. Anything the step did before the pause runs again, so a careless build can send the same email twice. Check that your platform saves the approval before pausing and runs each approved action only once.

Expiry, audit and approval fatigue

Every approval request needs an end time, and running out of time must mean “not done”, never “done by default”. A reply to a customer may be stale after a few hours; a record change can wait a day.

Keep the whole chain in the audit log: what was asked, which rule applied, who decided, and the result. In Botify every step of an assistant run is in the audit log.

The common failure is approving without reading. If people approve dozens of small requests a day, they stop looking. Keep approvals for steps that cannot be undone or that leave your company, and review the rejected ones: each shows where the assistant or your data needs work.

Common questions

01

What is the difference between human-in-the-loop and human-on-the-loop?

In human-in-the-loop, a person approves an action before it runs. In human-on-the-loop, actions run on their own and people watch and step in afterwards. Assistants that send messages or change records usually need the first for those steps.

02

Do approvals slow the assistant down?

Only at the steps that wait. Lookups run at full speed, and the assistant prepares everything up to the risky step, so the approver decides on a finished proposal instead of doing the work.

03

What happens if the approver rejects the step?

The action does not happen. The assistant can tell the customer that a person will follow up, and the rejection stays in the audit log for review.

04

Who can approve in Botify?

Team members with the member role or above. They decide from the Inbox in the console, or with buttons in WhatsApp and Telegram.

05

Is human approval required by law?

Article 14 of the EU AI Act requires effective human oversight for high-risk AI systems. Many business assistants fall outside that group, but approvals remain the simplest way to show that a person, not a model, allowed an important action.

Sources

  1. [1]LangGraph documentation: Interrupts
  2. [2]EU AI Act, Article 14: Human Oversight

Where you’ll find this in Botify

Start with one assistant on WhatsApp or your website.

Book a demo and we’ll show you an assistant answering questions like your customers’. We set up the first one with you and review the results together before you add more.