Stop Making Your Customers Click Approve

Emil Chetty, Wise15:33 · Nov 2026 · 2,134 views
Thumbnail for Stop Making Your Customers Click Approve Watch on YouTube
TL;DR
  1. 1

    Customers should approve an entire job once, rather than authorizing every action an agent takes.

  2. 2

    A locked, machine-readable agreement lets the agent plan freely while Wise's payment services enforce the approved limits.

  3. 3

    Prompt injection cannot expand the agent's authority because downstream services reject actions outside the confirmed agreement.

Summary

Emil Chetty describes a payment workflow in which the customer approves a complete job instead of supervising every agent step. Wise first uses read-only access and trusted software to gather suppliers, bank details, amounts, and other context. It then presents the customer with a human-readable proposal and a machine-readable version of the same terms. Once confirmed and authenticated, the proposal becomes immutable. Provisioning software grants the agent only the intersection of the approved job, Wise's permitted actions, and the customer's current rights. The agent can choose how to complete the work, retry an existing payment, and re-plan when the plan changes without changing the agreement. A changed quote or recipient requires a new customer decision. Chetty argues that this boundary should be enforced by payment services rather than by the model detecting every attack. That gives Wise an evidence chain showing who authorized the job, what the agent could do, and what happened.

Key ideas
00:01

AI assistants can turn customers into an approval queue

Chetty opens with the familiar experience of repeatedly approving coding-agent actions, website access, and MCP access. The same pattern can affect ordinary customers who were promised less complexity. At Wise, he leads data and AI product and wants to give agents useful autonomy without making the customer the API. Wise moved more than $243 billion in cross-border volume for nearly 19 million active customers and operates under more than 80 licenses. Those regulatory obligations remain when a chatbot or agent is placed in front of the payment system.

03:37

A business job contains many steps the customer should not have to supervise

Chetty uses the example of a small business owner asking an agent to handle the week's approved supplier payments, reconcile the results, and return exceptions. The owner wants the work done, not every lookup and status check coordinated manually. A conventional agent workflow exposes a list of approved actions, such as checking payment account details and scanning an invoice. That creates the same problem seen with early coding agents, where the user watched the terminal and approved each operation instead of getting the work handled.

05:35

Payment authorization asks a different question from software safety

Coding-agent controls ask whether an action looks safe. A payment needs to answer whether this customer authorized this payment to this recipient on these terms. Evals and a harness may help control a model, but they cannot prove that the customer agreed to the job. Chetty's proposal is an immutable agreement that defines the agent's permitted bounds. The customer agrees to the job once. The agent handles the steps, while Wise's services check each payment against the agreement.

06:42

The customer approves a locked agreement in both human and machine-readable forms

The process starts with a vague customer prompt. Trusted software gathers and verifies the relevant information, then shows the customer a proposed outcome. In the example, Sarah sees three suppliers and a total of $775 plus fees, along with the choices she is willing to delegate and the decisions that should return to her. The proposal has two views: a promise in human language and a machine-readable promise for Wise's services. After Sarah confirms and authenticates it, that exact version becomes the authorized job. What she approves is what the system enforces.

10:42

Authority is the intersection of the job, Wise's rules, and the customer's rights

A trusted provisioning platform examines three inputs: the confirmed job, what Wise permits the agent to do, and Sarah's current rights. It grants only the intersection of those permissions. The resulting access is split between a payment worker and a reconciliation worker. The agent can still decide how to achieve the job and can change its plan. It cannot change the agreement. This separates planning flexibility from payment authority, so delegation does not require the customer to supervise implementation.

11:32

Retries follow the existing payment state, while changed terms require approval

An immutable job can be retried after a service restart, a crashed app, or another interruption. Because payment services share the state of what has already run, a worker follows the existing job instead of creating a duplicate payment. A changed or expired quote is different. Sarah approved a specific quote, so the agent must prepare an updated proposal and cannot approve the replacement itself. The customer is shown what changed and decides whether to authorize a new version.

13:25

Downstream payment checks can contain a prompt-injected agent

Chetty considers an agent that is prompt-injected and believes its payment details have changed. The agent's belief cannot alter the confirmed job. When it submits a new recipient or account, the payment service rejects the request because it falls outside the approved bounds. The protection does not depend on the model recognizing the scam. It depends on downstream services checking and enforcing the agreement. Wise can then retain evidence of who authorized the job, what the agent could do, and what happened during execution.

14:18

One evidence chain lets Wise answer regulators without making the customer supervise

Chetty closes with the questions Wise needs to answer: who authorized the job, what the agent could do, what happened when the payment went through, what Wise could stop, and who was responsible. The proposed design connects the customer's decision, the agent's authority, and the payment outcome in one record. His operating rule is simple: agree to the job, let the agent choose the execution steps, and let the services enforce the agreement.

"We need the customer to agree to the job, the agent to choose the steps. And our software to enforce that agreement."07:49
Who should watch
  • You are building agents that can move money or change business records, and repeated human approvals are blocking the workflow.
  • Your system needs to prove what a customer authorized after an agent has planned and executed the work.
  • You are designing controls for prompt injection and want enforcement to happen in downstream services rather than relying only on model behavior.