NisusNisus

Checkpoint: Human Approval for AI Agents

AI Agents are running. But who fixes the wrong record?

Payment Agent
Supplier invoice read
Vendor account matched
Bank details change detected

Approval Request

Checkpoint

Amount€48,900RuleIBAN change + above thresholdApproverFinance Manager
Click the buttons. You make the call.

Automation is fast. Mistakes are permanent.

Reversals grind work to a halt

A wrongly issued invoice, a wrong ledger entry, a document sent to the wrong customer. Fixing it takes longer than the work itself.

Checkpoint
Stops before the action is taken

Some mistakes can't be undone

The declaration is filed, the payment is out, the document has reached the customer. A reversal doesn't cover the loss.

Checkpoint
Gates the irreversible step behind approval

The agent doesn't get the fine

A wrong declaration means a tax penalty; a wrong payment means a seven-figure loss. The agent makes the mistake, the company gets the bill.

Checkpoint
Stops before the penalty; a human decides

These steps happen every day

Irreversible

Customs declaration filed with the wrong HS code

Several HS codes match the product. The agent picks the closest one, prepares the declaration and sends it to the customs broker.

Checkpoint

Stops when more than one code matches. The trade specialist picks the right code; only then does the declaration go out.

Irreversible

Payment sent to the wrong IBAN

A supplier sends a request to update its bank details. The agent updates the vendor record and the payment order goes to the new account.

Checkpoint

Bank-detail changes and above-threshold amounts never pass automatically. An authorized person approves, and the approval is logged.

Irreversible

Shipment leaving without a valid A.TR

The movement certificate is missing from the file or expired. Nobody notices, the goods are loaded.

Checkpoint

Document requirements are checked per shipment. If something is missing, a warning goes out before the goods leave.

Hard to fix

Export invoice issued to the wrong customer

Two similar accounts exist within the same group. The agent picks one, issues the invoice and the document goes to the customer.

Checkpoint

Stops when the account match isn't certain. Both candidates are shown side by side; the user picks the right one.

Hard to fix

Journal entry posted to a closed period

The invoice date belongs to a prior period. The agent posts it to the open period or triggers the period to be reopened.

Checkpoint

On a period mismatch, nothing is posted. The accounting lead corrects the date or approves it.

Irreversible

Backdated social security filing

An employee change was reported late. The agent prepares the filing and submits it.

Checkpoint

Backdated filings never go out automatically. HR approves, and the justification is logged.

Hard to fix

Bulk price update

A new price list arrives. The agent writes thousands of items into the system. One item has the wrong unit or exchange rate.

Checkpoint

Stops on above-threshold price deviations. The category owner reviews the outliers and approves; the rest passes automatically.

Irreversible

Campaign message sent to a customer without consent

The campaign list is pulled from the CRM. The agent prepares the message and sends it to thousands of people by SMS and email. The list includes people who never opted in, or who withdrew consent.

Checkpoint

No record without verified consent enters the list. Unconsented records are set aside, the marketing owner reviews and approves them, and only then does the send go out.

Which steps should require approval?

Let's map the irreversible steps in your processes together in 30 minutes.

Not a reporting tool. A permission mechanism.

it prevents it.

it never posts the record.

it gates the risky step.

it documents the decision.

How will you trust the AI Agent?

You won't. You don't need to.

You define the risky step, the agent stops there, you grant permission. The question of trust disappears.

Four steps

  1. 01

    Define

    You define the risky steps. Amount threshold, document type, transaction type.

  2. 02

    Run

    The agent runs the process. It doesn't stop until it reaches a risky step.

  3. 03

    Stop

    It stops there and puts the task on the authorized person's screen.

  4. 04

    Approve

    Permission is granted, the agent resumes where it paused. The decision is on record.

It can't pass without your permission.

Let's decide together which steps need approval, in a 30-minute call.