Look at purchases in groups, not one at a time. The agent links requisitions, card transactions and invoices that share a requester, supplier and purpose within a short window, adds them up, and flags any group whose total crosses a limit that each part avoided. It flags before approval where it can and afterwards where it cannot. Approvers and procurement decide what a flag means.
What is a split purchase, and who owns detection?
A split purchase is one need broken into several smaller purchases, each below an approval or quotation threshold, so that it avoids a level of scrutiny the whole would have triggered. Sometimes it is deliberate: someone wants to avoid a slow approval or a competitive quote. Sometimes it is innocent: a team orders as needs arise. Either way, the control the limit was meant to provide did not operate.
The head of procurement owns detection for requisitions and orders. The expenses lead owns it for company cards. Internal audit uses the same analysis after the fact. Approvers own the decision on any flagged purchase. The wider finance position is in AI agents for finance: books you can audit.
How does the agent detect splits, step by step?
- Agent, depth 1: reads purchasing activity. Requisitions, purchase orders, card transactions and invoices without orders, with requester, supplier, date, amount and description. Evidence: the sources and period read.
- Agent, depth 1: groups related items. Same requester or cost centre, same supplier or merchant, similar description, within a window the procurement lead sets. Evidence: each group and the links between items.
- Agent, depth 1: compares group totals with limits. Approval limits by role and quotation thresholds by category. Evidence: the limit and its version.
- Agent, depth 2: scores each group. Items just under the limit score higher; consecutive days, sequential order numbers and identical descriptions add weight. Evidence: the score and its components.
- Agent, depth 2: flags at requisition time. When a new requisition would join a group whose total crosses a limit, the approver sees the group before deciding. Evidence: the flag shown and the approver's decision.
- Agent, depth 2: reports after the fact. Groups found in card spend and invoices, where no preventive check was possible, go to the expenses lead or procurement weekly. Evidence: the report.
- Person, approver or procurement lead: decides. Approves the group as truly separate needs, sends it back for a proper approval or quotation, or refers a pattern to internal audit. Evidence: the decision and reason.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Read | Reads requisitions, orders, cards, invoices | 1 | Nothing |
| Group | Links related items | 1 | Setting the window |
| Compare | Sums against limits | 1 | Setting limits |
| Score | Weights signals | 2 | Nothing |
| Pre-approval flag | Shows the group to the approver | 2 | Approval decision |
| Post-event report | Lists groups found afterwards | 2 | Follow-up |
| Decide | Nothing | None | Approver, procurement, audit |
Why flag rather than block?
Because many groups are legitimate. A team running an event may well buy catering, room hire and printing from the same venue on separate orders. A site manager may order materials as a job progresses. A hard block would stop real work and teach people to vary descriptions or suppliers to get round it. A flag puts the group in front of the person whose approval the limit was protecting, and lets them decide with the full picture.
Flags are worded as facts: "This requisition brings purchases from this supplier by this requester to a total above your approval limit in the last ten days." Not as accusations.
Do the same limits apply to the agent itself?
Yes, and this is easy to miss. An agent that prepares purchase orders, payments or payroll changes within a value limit can, by accident or by being instructed, break one large item into several small ones. The platform should therefore count an agent's actions cumulatively against its limits, not only one at a time. The design of agent limits is covered in how do you set approval limits for an AI agent in finance, and the same discipline applies when an agent drafts items for approval, as in letting an AI agent prepare payroll changes for approval.
Which controls matter, and what evidence do they leave?
Limits versioned. A flag cites the limit in force on the purchase date. Evidence: the limit version.
Groups explained. Every flag shows the items and why they were linked. Evidence: the link reasons.
Decision recorded. Approver decisions on flagged groups are kept. Evidence: decision and reason.
Patterns escalated. Repeated groups from one requester go to internal audit automatically. Evidence: the escalation.
What should you measure?
- Groups flagged before approval against groups found afterwards. More of the first means the preventive check is working.
- Flags confirmed as splits by approvers, against flags approved as separate needs.
- Repeat requesters, which may point to a training need or a limit that is set too low for the work.
- Purchases sent back for quotation, and the savings found where the quote changed the price.
Where does it usually go wrong?
The window is too short. Splits across a month boundary are missed because the window resets. Use a rolling window.
Suppliers vary. The requester spreads the purchase across two related suppliers. Group by supplier family where you can, and by description.
Card and invoice data are separate. Half the purchase is on a card and half invoiced. Read both.
Limits are simply too low. If a team splits every month because their work needs bigger purchases, the limit is the problem. That is a policy change, not a detection fix. Detection belongs with the preparation work that keeps period-end clean, a theme in a faster close comes from preparation, not automated approval.
Questions people ask
Is this the same as duplicate detection?
No. Duplicates are the same charge twice; splits are one need presented as several charges. The grouping logic is similar, and the duplicate side is covered in can an AI agent detect duplicate invoices.
Can split purchases indicate fraud?
They can, especially combined with other signals such as a little-known supplier. Patterns like that go to internal audit, alongside checks such as detecting fictitious suppliers.
Should requesters see the flag?
The approver sees it first. Whether the requester is told, and how, is the approver's call.