By KYN AI Advisory Team — AI implementation specialists, Singapore
Place a human approval at every point where an action leaves the building or cannot be undone: outbound messages, money movement, filings and deletions. Let the AI draft, match, sort and log without asking. The rule is simple because the test is simple. Ask what happens if the agent is wrong. If the answer is "we fix a row", let it run. If the answer is "a customer sees it" or "money moves", a named person approves first.
This is a design question, not a trust question. You are not deciding whether the AI is good. You are deciding where an error becomes expensive.
Why not review everything?
Because reviewing everything defeats the purpose, and people stop reading. A reviewer asked to approve 300 items a day clicks through. The approvals become a ritual that records consent without giving any control.
Reviewing nothing fails the other way. One wrong message to a customer, one wrong payment, one deleted record costs more than all the minutes you saved.
The target is a small number of high-consequence approvals, each fast to give and each read.
Which actions need a named person?
Four classes, in the order we apply them.
| Class | Examples | Why it waits | |---|---|---| | Outbound | Email or WhatsApp to a customer, supplier or new contact | It carries your name and cannot be recalled | | Money | Payments, refunds, credit notes, payroll runs | Errors cost cash and are slow to reverse | | Filings | GST returns, tax submissions, regulatory forms | A wrong filing creates a record with an authority | | Deletions | Removing or merging records, closing accounts | You cannot recover what you did not back up |
Anything outside these classes is a candidate to run unattended: reading, classifying, drafting into a queue, matching, calculating, writing to a log, raising a flag.
For deeper guidance on the most sensitive cases, see how to set approval gates for high-risk AI agent actions.
Where does the approval go in the flow?
At the last step before the action, not at the first step. Let the agent do all the preparation, then stop at the gate.
For an outbound message:
- The agent reads the account and finds a checkable fact.
- The agent drafts the message.
- The draft lands in a review queue with the source of the fact beside it.
- A named person approves, edits or rejects.
- Only after approval does the system send.
- The log records the draft, the approver, the time and the result.
The reviewer sees one screen: the draft, the evidence, and three buttons. The work they do is judgement, not typing.
For a payment run:
- The agent matches bills to purchase orders and bank lines.
- The agent proposes a batch and flags any bill that does not match.
- A named approver sees the batch, the flags and the total.
- The approver releases the batch in the bank's own approval step.
The agent never holds the power to release funds.
How do you keep approvals fast?
Four design choices decide whether review takes minutes or hours.
Show the evidence next to the draft. If the reviewer must open three other tabs to check, they will not check. Put the source line, the matched document or the previous message on the same screen.
Batch by type. Fifty similar drafts in one queue take far less time than fifty items scattered through an inbox.
Pre-sort by risk. Flag the unusual: a first-time supplier, an amount above a threshold, a reply that mentions a dispute. The reviewer gives those their attention and clears the rest quickly.
Allow edit-and-approve. If the reviewer can fix a word and approve in one action, they fix instead of rejecting. Rejections that restart the process feel expensive and slow the queue.
What should the agent do with an unusual case?
Stop and hand to a person. Write the rule down.
A reply that mentions a complaint, a legal threat, a regulator or a refund is a stop condition. So is a bill with a bank account that differs from the supplier's usual account. The agent labels the item, explains why it stopped in one line, and routes it to a named person. It does not guess.
Unusual cases are where agents are most confident and most wrong. The rule that protects you is not "the agent should know". It is "the agent must stop".
Who counts as a named person?
A specific individual, with an account, whose approval is logged. Not "the team" and not a shared mailbox login.
Name a backup too. An approval chain with one person stalls when that person is on leave. Decide in advance who covers, and what happens when no one responds in a set time. The safe default is that the action does not happen. We cover this failure in the approval chain timeout trap.
How do you know the gates are in the right place?
Review three measures after the first month.
- Approval rate. If the reviewer approves 99 of 100 drafts unchanged for weeks, the gate may be safe to relax for that class, or the reviewer may not be reading. Sample a few to find out.
- Edit rate. A high edit rate on one type of draft means the prompt or rules need work, not that the gate should go.
- Time in queue. Items waiting more than a day show a staffing problem, not an AI problem.
Relax a gate only for a narrow, low-consequence class, and keep the log. Do not relax a gate for outbound messages to new contacts, money or filings.
What about regulated or audited work?
An approval is also evidence. Auditors, insurers and regulators ask who authorised an action. A log that records the approver, the time and what they saw answers that question. A chat message saying "looks fine" does not.
This is the same reasoning as in bank reconciliation: the agent proposes matches, a person approves exceptions, and the record shows both.
Where KYN fits
In KYN's systems, outbound messages, money, filings and deletions wait for a named person. Everything else, such as research, drafting into a queue, sorting replies and writing to the log, runs on its own. Every action is logged with the actor, and agents work from their own restricted accounts.
On our own books, bills come in from Gmail and receivables are chased at 3, 14 and 30 days, with the sends waiting on a person. The first 14 days of an engagement are a review where we map each workflow and mark where the gates belong. A full build ships in under 90 days. See operations and security.
Frequently asked questions
Does human review make an AI workflow slow?
Not if the gates are few and the evidence is on the screen. A reviewer clearing a batch of drafts with sources beside them spends minutes. The delay comes from reviewing everything or from missing context.
Can I remove approvals once the agent proves reliable?
For low-consequence classes, such as internal labels, yes. For outbound messages to new contacts, money, filings and deletions, we keep a named person in the loop, because the cost of one error outweighs the minutes saved.
Who should be the approver?
The person who would be accountable if the action were done by hand. For a customer message, the salesperson or account owner. For a payment, the finance lead. Name a backup for each.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
