All field notesWorkflow Ownership

An Approval Button Is Not a Control

Putting a human sign-off in front of an AI agent only protects you if the reviewer can see enough, has the authority to say no, and only gets asked about the steps that matter. A practical guide for SME operators.

When an SME owner asks how to keep an AI agent safe, the usual answer is "put a human in the loop." Add an approve button, make someone click it, done.

This week gave us three good reasons to be more careful with that answer. A product engineer argued in public that an approval button on its own records consent without creating oversight. A securities regulator said reviewing every agent step is unrealistic. And two mainstream vendors showed what a better gate looks like. If you run finance, operations or a small team that is starting to hand work to agents, this is the part worth getting right.

What happened

The rubber stamp problem. In a piece published on 2 November, Quadri Ismail, a product engineer at a B2B fintech, describes a loan tool that flags an application as high risk and a staff member who clicks Approve.[2] The system logs a human decision, but nobody can say what the person actually reviewed, whether they could challenge it, or whether disagreeing would have changed anything. His summary: "the approval step may be recording consent without creating oversight."[2] He points to examples the UK Financial Conduct Authority published on 19 October, which warned that human review can turn into box-ticking when staff lack the information, expertise, authority or escalation path to push back.[2]

A regulator admits the obvious. Speaking at Hong Kong FinTech Week on 2 November, SFC chief executive Julia Leung said the regulator will write "practical, principles-based guidance" for more autonomous AI, recognising that human review of every intermediate step is unrealistic.[3] The SFC sketched a gradient: start autonomy with low-impact, reversible back-office work, and apply "progressively stronger safeguards" to client advice, portfolio decisions and market-facing execution.[3] The test it wants firms to pass is whether they can "reconstruct a decision, identify the action taken, explain the outcome, and determine accountability when things go wrong."[3] These were panel remarks, not rules, and no thresholds were published.[3] On the same day the HKMA launched Project SAGE, which includes work on verifying AI agents' identities and delegated authority.[5]

Two vendors show the better pattern. Intuit's new Chart of Accounts Cleanup in QuickBooks Online finds duplicate, misnamed, unused and wrongly typed accounts and suggests fixes. "Nothing changes in your books until you review and approve each suggestion."[1] By default it won't touch transactions before your closed-books date, and some recommendations are shown only to accountants.[1] Intuit is also blunt that you can't undo changes once approved.[1]

Over at AWS, Step Functions workflows can now be handed to an agent as a tool.[4] The agent decides when to run a workflow and what to pass in, but the steps run exactly as defined. AWS's own example is a refund where "the eligibility check, the manager approval, and the credit still happen in that order every time, because the ordering was never the model's to improvise."[4]

What I take from it

Put those together and you get a simple rule for SMEs: don't ask for approval everywhere. Ask for it at the few points where a mistake is expensive or permanent, and make that approval real.

Too many approvals is its own failure. If your finance lead has to click through forty agent suggestions before lunch, by the tenth they are skimming. You've built the rubber stamp the FCA was warning about.[2] Too few, and the agent is quietly posting entries nobody looked at.

The QuickBooks design is worth copying even if you never use QuickBooks. It keeps some recommendations for accountants only, it protects closed periods, and it tells you plainly which actions can't be reversed.[1] That's an approval policy written into the product.

A practical way to set your gates

Before you switch an agent on, sit down with the person who owns the process and sort every action the agent can take into three groups.

1. Let it act, then spot-check. Reversible, low-impact work: drafting a reply, tagging an invoice, flagging a duplicate vendor, renaming an account to fix a typo. Review a sample weekly. This is where the SFC suggests autonomy should start.[3]

2. Let it prepare, a named person approves. Anything that changes a record of value or is hard to undo: reclassifying transactions, merging accounts, issuing a refund, sending to a customer. One approval, from someone with the authority to say no.

3. Keep it out entirely. Closed periods, payroll changes, bank details, anything regulated. If the tool can't be locked away from these, don't connect it yet.

Then make the group 2 approvals mean something. Ismail's checklist translates well to an SME back office.[2] When the agent asks for sign-off, the approver should be able to see:

  • what exact change is being proposed,
  • which records or documents it is based on, and how fresh they are,
  • what is missing or uncertain, and
  • what they can do instead: edit it, reject it, park it, or escalate it.

If your current tool shows a one-line summary and a green button, that's a gap. Raise it with your vendor or implementer.

Finally, fix the order of steps in the workflow, not in the prompt. The AWS refund example works because the approval sits in a fixed sequence the agent can trigger but can't reorder.[4] For a small team the same idea might be a simple form or an ERP approval rule that the agent submits into. The point is that the agent never decides whether the check happens.

The record matters as much as the click

The SFC's test is useful for any business, regulated or not. Six months from now, if a supplier disputes a payment, could you show what the agent proposed, who approved it, what they saw when they approved it, and what happened next?[3] If the answer is "we'd have to dig through chat logs," your control lives in someone's memory, and that's a weak place to keep it.

So when someone tells you their AI rollout is safe because "a human approves everything," ask two follow-up questions. Approves what, exactly, with what in front of them? And how many times a day? The answers tell you whether you have oversight or a habit of clicking.

This post draws on public reporting and vendor documentation published between 27 October and 2 November 2026. The SFC remarks are supervisory intent, not regulation.

Sources

[1] https://quickbooks.intuit.com/learn-support/en-us/help-article/chart-of-accounts/clean-chart-accounts-quickbooks-online/L5aVdBfnq_US_en_US — Clean up your chart of accounts in QuickBooks Online (Intuit, updated 27 Oct 2026)

[2] https://hackernoon.com/a-human-approval-button-is-not-an-ai-safety-strategy-for-fintech — A Human Approval Button Is Not an AI Safety Strategy for Fintech (HackerNoon, 2 Nov 2026)

[3] https://kya.fi/news-2026-11-02-sfc-agentic-ai-guidance — Hong Kong's SFC says reviewing every agent step is unrealistic (Know Your Agent, 2 Nov 2026)

[4] https://aws.amazon.com/blogs/aws/aws-weekly-roundup-gpt-6-1-terra-on-bedrock-step-functions-as-agent-tools-and-more-november-2-2026 — AWS Weekly Roundup, 2 Nov 2026 (AWS News Blog)

[5] https://www.hkma.gov.hk/eng/news-and-media/press-releases/2026/11/20261102-4 — HKMA launches Project SAGE (HKMA press release, 2 Nov 2026)

Continue the work

Turn a capable model into dependable execution.

Nexius Labs helps SMEs design the context, tools, permissions, approval gates, and evidence trails around useful Digital Coworkers.