Hand a repetitive task to AI, and keep the approval.
A supplier invoice, checked against its purchase order · Illustration, fictional supplier
InvoicePackaging supplier · against your purchase order
Arrived by email this morning
Line on the invoiceChecked against the order
Cartons, large
Quantity and price match
Packing tape, clear
Quantity and price match
Pallet wrap
Unit price higher than ordered
Delivery
On the order
Tax line (VAT in the UK, GST in Australia)
Present and correct
Read against the purchase order and the prices agreed with this supplier
Held for the person who approves
Note from the AI
Everything on this invoice matches the order except the pallet wrap, which is charged at a higher unit price than the one agreed. I have not entered it.
Accept the new priceAsk the supplier
Once approved, entered in your accounts, such as Xero or QuickBooks
Some work comes back every week in the same shape: supplier invoices to check against what was ordered, booking requests to enter, figures to gather into the same report. It takes hours of someone’s attention, and a mistake in it costs money or a customer’s trust. Handing it to AI makes sense, and the hesitation is understandable: nobody wants an AI paying the wrong amount or writing to a customer without anyone checking. This page explains which tasks to hand over first, where a person’s approval sits, what is written down before anything is built and what the person approving does each week.
Which tasks to hand over first
A good first task is one your team does often, in the same steps each time, where there is a clear way to tell whether the result is right. Checking a supplier invoice against the purchase order is a typical case: the order says what was agreed, the invoice says what is charged, and a line either matches or it does not. When the AI reads both documents it can prepare the entry and point to the exact line that differs, and the person checking it can confirm the result without redoing the comparison.
Other tasks are better kept with a person, even when they repeat: paying a supplier, setting up a new supplier and their bank details, answering a complaint, or changing a price agreed with a customer. In these the cost of a mistake is high, or the decision depends on judgement and a relationship the AI does not hold. The AI can still prepare the information for them, and the action itself stays manual.
Sorting the repetitive work in a business
AI prepares, a person approves
Supplier invoices checked against the orderDifferences held with a note
Booking requests entered in the calendarClashes sent to a person
Weekly figures gathered into one reportRead before it is shared
Customer details copied from a form into your recordsIncomplete forms flagged
Stays with a person
Paying a supplierThe AI can prepare the payment list
Setting up a new supplier and their bank detailsChecked by a person every time
Answering a complaintThe AI can gather the history
Changing a price agreed with a customerA decision, not a check
Where the approval sits
For each task, the level of independence the AI gets is decided before it is built, as part of the scope. There are three levels. As a copilot, the AI prepares the work and a person carries it out. Semi-automated, the AI carries the task out and nothing counts until a person approves it. Automated within written limits, the AI completes the cases that fit the rules on its own and sends every case outside them to a person.
CopilotThe AI prepares
Semi-automatedA person approves each one
Automated within limitsOutside the limits, a person
The AI
Compares the invoice with the order and points out the lines that differ
Prepares the entry in your accounts, with a note on anything that differs
Enters the invoices that match the order exactly, and holds the others with a note
The person
Enters the invoice, using the comparison
Approves each entry before it counts, or corrects it
Decides on every held invoice and reviews a sample of the ones entered
At every level: paying the invoice stays with a person.
The right level depends on the task and on how much is at stake when it goes wrong. We do not push a task to full automation when the variables, the access it would need or the judgement involved make that irresponsible, and a task can move up a level later, once the approvals show that the AI handles it reliably. What the AI is allowed to touch at each level is set like any other access, which we explain in adopting AI securely and on our security page.
Written down before it is built
Before anything is built, the task is written down step by step with the people who do it today. We record them doing it and draw the steps with them, so the written version includes the checks they make without thinking about them, such as which supplier is allowed a different delivery charge or which customers always pay late. If the process is not settled yet, we settle it first, because an AI built on a process that changes every week repeats whichever version it was given.
The written version also says where the AI stops: which cases it handles, which ones it holds, who approves, and what is kept as a record of each decision. The workflow is then built, tested against real past cases from your business, and delivered against that written description, so you can check that it does what was agreed. This written knowledge of how your business works is the same structure we describe in what an AI adoption consultant does.
What the approver does each week
The person approving does not redo the work. They read what the AI has held, with its note on why, and decide: accept the new price, ask the supplier, or correct the entry. On a semi-automated task they also approve the prepared entries, which takes a glance when the comparison is shown next to each one.
Each correction is then written into the rules the AI reads, so the same case is handled the same way the next time. This is what stops the AI from repeating a mistake you already corrected. When you use AI in a chat window, a correction made yesterday is gone today, because nothing was written down where the AI reads it at the start of the next task.
A week for the person who approves the invoices
MondayReads what was heldEach held invoice comes with the line that differs and the AI’s note.
WednesdayConfirms the new price with the supplierThe pallet wrap price was raised on purpose, so it is accepted.Written into the agreed pricesPallet wrap: new unit price from this month
From the next invoiceThe AI checks against the new priceThe same line is no longer held, and a different price would be.
The time it frees
Handing over a repetitive task frees hours for the person who did it, and those hours need a destination decided with you before the work starts: the follow-ups that never get done, the supplier conversations that wait, the customers who deserve a call. We never present the freed time as a reason to remove a role, and the results are measured with your own figures, not with an estimate from elsewhere.
Your team also needs to know how to read what the AI prepares and correct it well, which we cover in AI adoption training for your team. If you are choosing the tool that will run the workflow, we compare Zapier, Make and n8n, including where each one keeps your data for a business in the UK or in Australia.
How we start
The first step with us is a free AI pre-audit: 45 minutes on a call where we look at how you and your team use AI today and which repetitive work is worth handing over first. Bring one task your team does every week and the way it is done today. You leave with a clear next step, with or without us.
If you decide to go further, each workflow is built at a fixed price against its written description, inside the Sprint or on its own. See how we help and who we are. In the news this week, the way a customer’s own AI agent may soon deal with your business: Sierra and Meta’s Personal Agent Protocol.
The AI AdoptionFree AI pre-audit 
Xero or 
QuickBooks