Reliable output
Which AI jobs still need a person to sign off
The AI can now send the quote itself. The question is whether you want it to, and what happens on the day it sends one wrong.
Which AI jobs need a person to sign off?
Four kinds of job need a person to approve them: anything that moves money, anything you can't take back, anything sent under your name, and any decision about a person. If a job does none of those, the AI can usually do it alone with a spot check. If it does one, keep the approval.
The test is simple to run on any job. Ask: can it be undone, who sees it, and does it touch money or a person? The question gets sharper once AI stops only drafting and starts acting, which is the difference we explain in AI agents vs chatbots.
Anything that moves money
Payments, refunds, credit notes, discounts. A wrong payment is slow and awkward to get back, and a wrong refund is money gone. The AI can prepare all of it. A person presses the button. The National Cyber Security Centre is blunt on this: in Exercise caution when building off LLMs it advises against letting these systems make transactions on a customer's behalf, and says to design so you could accept the worst outcome.
Anything you can't take back
Deleting records, cancelling a booking, closing an account. A draft can be thrown away; a deleted customer file usually can't be brought back. If the action is final, the decision should be yours. Even a legitimate deletion request should run through a fixed process with a grace period, which is part of what the GDPR Compliance Kit sets up.
Anything sent under your name
Quotes, replies to complaints, messages to a new customer. Once it's read, it's said. A venue manager can live with a clumsy internal note; a clumsy reply to an angry customer is a different matter. And a confident wrong fact is the most common way it goes wrong, as our post on why AI makes things up explains.
Decisions about a person
Turning down a job applicant, refusing a refund, flagging a customer as a risk. These affect someone's life or money, and UK data protection law has specific rules for significant decisions made only by a machine. The rules were rewritten by the Data (Use and Access) Act 2025: Article 22A of the UK GDPR now treats a decision as solely automated where there is no meaningful human involvement in taking it. Keep a person making the call, with the AI preparing the facts.
Where the sign-off goes, job by job
| Job | Can it be undone? | Sign-off |
|---|---|---|
| Sorting and labelling the inbox | Yes | None needed; spot-check weekly |
| Drafting internal job notes | Yes | None needed; spot-check weekly |
| Drafting replies to enquiries | Yes, until sent | A person approves before it sends |
| Booking confirmation to a known customer | Mostly | Approve at first, then a written rule |
| Posting on your social accounts | Partly, once seen | A person approves every one |
| Sending a quote | No, once read | A person approves every one |
| Replying to a complaint | No, once read | A person approves every one |
| Chasing an overdue invoice | Partly | Approve the first rounds, then a written rule |
| Issuing a refund | No | A person approves every one |
| Deleting customer records | No | A person approves every one |
| Turning down a job applicant | No | A person decides |
What makes a sign-off real?
A sign-off is real when the person approving has the facts, the time and the authority to say no, and sometimes does. A button pressed without reading is a delay, not a check. The law draws the same line: involvement has to be meaningful, not a name on the end of a machine's decision.
Rubber-stamping is not approval
If you approve forty drafts in two minutes, you are not reviewing them. People who are busy, under time pressure and used to the system being right are exactly the ones who stop looking. A systematic review of automation bias in the Journal of the American Medical Informatics Association found workload, time pressure and trust all push people towards over-relying on automated advice. Fewer, better approvals beat many quick ones.
The approver needs the facts, not just the draft
Approving a refund email means nothing if you can't see the order, the payment and the customer's history. Microsoft's guidance on overreliance on AI says the risk is higher wherever AI can act without giving the user a chance to review, and that verification should be made easy. Put what the decision depends on next to the button.
High-risk decisions get a person, every time
This is not only small-business caution. The UK government's own AI Playbook sets meaningful human control at the right stage as one of its ten principles, and says humans should validate any high-risk decisions influenced by AI. If government departments with whole teams on this keep a person in the loop, a five-person firm should too.
People can ask for a person
Where a significant decision about someone is automated, the UK GDPR now requires safeguards. The ICO's draft guidance on the ADM safeguards lists them: tell the person about the decision, let them make representations, let them get human intervention, and let them contest it. Keeping a person deciding in the first place is simpler than building all four.
How to set up a sign-off step people actually use
- 01
Put the approval where you already are
If you live in your email, the drafts should land in your email. An approval step in a tool you never open becomes a queue nobody clears, and then someone switches it off.
- 02
Show what matters, not just the draft
A good approval shows the customer, the amount and anything unusual at the top. You should be able to say yes or no in seconds without opening three other screens. Automatic checks can clear the obvious mistakes before you look, as set out in how to check AI's work before a customer sees it.
- 03
Make no as easy as yes
Rejecting or editing a draft should take one step, with a short note on why. Those notes are how the AI's instructions improve, and each one should become a rule, as in every AI mistake should become a rule.
- 04
Set limits in numbers, not feelings
"Small refunds are fine" means different things on different days. Write the line down: refunds up to an amount you choose, from customers you already know. Above the line, it waits for you. The same goes for what the AI can reach at all, which we cover in give AI only the tools it needs.
- 05
Review the line every month
Look at what you approved without changes and what you had to fix. That tells you which jobs are ready for less sign-off and which still need more.
A worked example: invoice chasing at a small accountancy practice
Here is one job moving through the stages, from full sign-off to a written rule. The practice and its numbers are an example, but the steps are the ones we would use.
A three-person bookkeeping practice spends part of every Friday chasing overdue invoices. The AI is set up to draft each reminder from the accounts software: client name, invoice number, amount and days overdue.
Weeks one to four: every reminder approved
Every draft lands in the partner's inbox. She reads each one against the ledger, sends most unchanged and edits a few. Two of her edits share a cause: the AI chased an invoice the client had already disputed. That becomes a rule. Disputed invoices are never chased automatically, and they are tagged in the accounts software so the AI can see it.
Weeks five to eight: first reminders on a rule
The first, polite reminder to a client who has paid on time before now goes out on its own, under a set amount. Second reminders, new clients and anything above the amount still wait for her. She reads a sample of the automatic ones each Friday.
What never moves
Final notices, anything mentioning fees or legal action, and any message to a client who has complained stay with a person for good. They are the messages where a wrong word costs a client, and the time saved by automating them is small compared with that.
Common mistakes with sign-off
Most sign-off steps fail quietly. Nobody decides to remove them; they wear away. These are the patterns to watch for.
The same sign-off for every job
If sorting the inbox needs the same approval as a refund, you will soon approve both without reading. Match the sign-off to the risk, using the table above, so the approvals that remain get real attention.
Letting the AI decide what needs approval
"Ask me if you're unsure" sounds sensible, but the AI is often sure when it is wrong. The rule for what needs approval should be written by you, in numbers and categories, not left to the AI's judgement.
Removing sign-off after a good week
A good week is a small sample. A job earns less sign-off with a written record over weeks, across its normal mix of cases, including the awkward ones. Our guide on how to trial AI on one job sets a pass bar you can hold it to.
Forgetting the sign-off when something changes
A new price list, a new tool or a provider updating its model can all change what the AI does. When any of those happen, go back a stage for a few weeks and let the record rebuild.
Hiring with AI and no one checking
Sorting CVs is tempting to hand over, and it is where unfairness hides best. The government's Responsible AI in Recruitment guide warns that hiring tools can perpetuate existing biases. Let AI summarise; let a person shortlist and decide.
When can a job lose its sign-off?
A job can lose its sign-off when it has a written record of approvals you didn't need to change, and when a mistake would be small and reversible. Money, deletions and decisions about people usually keep a person for good. Everything else can earn its way off, one step at a time.
A clean record, written down
Count the drafts you approved untouched and the ones you fixed. When the fixes stop, and the last few were caught by a check rather than by you, the job has earned more freedom.
Move one step at a time
Don't jump from drafts to fully automatic. Let it send to known customers first, or only during working hours, or only below a set amount. If something goes wrong, you have only widened one thing. For paperwork, Document Automation builds invoices and reminders from the job record itself, so the figures a person approves are the figures that go out.
Keep a morning report, even without sign-off
A job that runs alone should still tell you what it did. A short daily list of what was sent, what was held back and what it refused is enough to spot drift early, and it is the habit our safety rules are built around.
Questions
08Asked and answered
Yes, for low-risk, reversible messages once the job has a clean record, such as booking confirmations to customers you already know. Anything new, about money, about a complaint or to a first-time customer should wait for you. Widen what it sends alone one step at a time. Our safety page lists what AI never does without a person.
AI can sort and summarise applications, but a person should make the decision. UK data protection law has specific rules and safeguards for significant decisions made without meaningful human involvement, and they were rewritten in 2025. The ICO's guidance on automated decision-making explains them.
Not if you approve without the facts, the time or the authority to say no. The UK GDPR now looks for meaningful human involvement in a decision, set out in Article 22A, and a quick click on forty drafts is hard to call that. For your own case, take legal advice.
AI should never move money beyond a limit you set, delete records, give anyone access to your accounts or change its own rules. Each of those is hard or impossible to undo, so the AI prepares and a person decides. We cover what access to give in give AI only the tools it needs.
No. Approving a good draft takes seconds and writing it takes minutes, so you still save most of the time while keeping control. Automatic checks cut the reading further by clearing the obvious mistakes before you look. We show how in how to check AI's work before a customer sees it.
Run it with sign-off first and keep a written record of every draft you changed and why. When the changes stop for several weeks, across the job's normal mix of cases, widen one thing at a time. Our guide to trialling AI on one job sets a pass bar you can hold it to.
Yes, if it works behind limits and leaves anything that needs a decision for you in the morning. The limits matter more than the hour it runs. Autonomous Agent Operations runs repeat jobs overnight behind a safety lock and leaves you a morning report of what ran, what changed and what it refused.
This is general guidance, not legal or financial advice. For your own situation, speak to a qualified adviser.