Accrual automation is moving from spreadsheet macros to AI agents that read the documents behind each estimate. Done well, it means every accrual starts from an open purchase order, a goods receipt or a contract term rather than last month's number rolled forward. This article explains how agents build those proposals, how prepayments follow the same pattern, and how reviewers approve, adjust or reject what the agent puts forward. It also covers the evidence auditors will expect and the estimates that are not worth automating.
1. What is accrual automation, and what does it change?
Accrual automation is the use of software to identify, calculate and propose period-end estimates for costs incurred but not yet invoiced, and for amounts paid in advance that need releasing over time. Older tools only did the arithmetic: a recurring journal template, a percentage uplift, a reversal flag. The judgement stayed with an accountant who emailed budget holders, chased replies and keyed figures into a workbook. AI agents change where that judgement starts. They read the purchase order, the receipt history, the contract and the supplier's invoicing pattern, then propose an accrual with the reasoning attached.
The accountant's job shifts from compiling to reviewing. That sounds minor, but it rewrites the month-end accruals timetable. Instead of waiting for budget holders, the reviewer starts day one with a list of proposals, each carrying a confidence level and links to the source documents. Questions go to budget holders only where the agent cannot support an estimate from the evidence. In a well-run setup, the accrual agent works as one of several finance AI agents, each with a defined role and decision rights that stop short of posting anything without a person's approval.
2. Where accrual and prepayment estimates go wrong today
The problems are familiar to any controller. Goods received not invoiced balances build up because receipts are booked against purchase orders that were never closed, so the GRNI report overstates the liability and someone cleans it by hand. Service costs are harder still: there is often no receipt at all, only a contract and a budget holder's memory of what was delivered. The result is a pile of round-number accruals that reverse next month and get rebooked with little scrutiny.
Prepayments fail differently. An annual software licence or insurance premium is paid, coded to a prepayment account, then released from a schedule kept in a workbook that one person understands. When that person is on leave, the release is missed or doubled. Take a distribution group with entities in Dubai and Riyadh: each entity may keep its own schedule in its own format, so the group balance sheet review turns into a reconciliation exercise instead of a genuine challenge of the numbers.
3. How AI agents propose accruals from open purchase orders and receipts
The starting point is the open purchase order book. For each line with a receipt but no matching invoice, the agent values the received quantity at the order price, checks currency and tax treatment, and looks for signs the receipt is stale: months old with no invoice, a supplier that normally bills within days, a line already part-invoiced at a different price. Stale items are proposed for write-back or sent to the buyer, never accrued blindly. The logic is close to that used for three-way match exceptions, and both processes draw on the same data.
Orders without receipts need a different test. Where an order covers a service delivered over time, the agent compares order value, delivery dates and any timesheets or milestone confirmations to estimate what has been consumed by period end. It then checks invoice history: if this supplier usually bills a month in arrears, a missing invoice is expected and an accrual is justified. Every proposal shows the calculation, the documents relied on and the accounts and cost centres it would hit, so the reviewer sees exactly why it exists.
4. Contracts, service accruals and the prepayment amortisation schedule
Some of the largest estimates sit outside purchasing altogether: rent, utilities, professional fees, maintenance agreements and commissions. Here the agent reads the contract to extract billing frequency, fee basis, escalation clauses and service period. A facilities contract billed quarterly in arrears gives a predictable monthly accrual. A legal engagement billed on hours worked needs input from the firm or the matter owner, and the agent should say so rather than invent a figure. Knowing which estimates it cannot support is as valuable as producing the ones it can.
Prepayments run the same logic in reverse. When an invoice covers a future period, the agent reads the service dates from the invoice and contract, proposes the prepayment coding and builds the prepayment amortisation schedule at the point of entry. Each month it drafts the release journal and checks the remaining balance against the remaining service period. Cancellations and early renewals are flagged for review. Moving schedules out of personal workbooks into a governed system is often the most useful result of the whole exercise.
5. How reviewers approve agent-proposed accruals
Review design decides whether accrual automation earns trust or simply adds a second layer of work. Group proposals by risk, not by supplier. High-confidence, low-value items backed by a receipt and a stable invoicing pattern can be approved in batches. Material items, judgemental service estimates and anything that moves beyond an agreed tolerance from the prior month go to a named reviewer one at a time, with source documents open alongside. Setting those human-in-the-loop approval thresholds is a controller's policy decision, and it should be written down before the first live close.
Reviewers need three actions: approve, adjust with a reason, or reject with a reason. The reasons matter more than the decisions. An adjustment because a delivery slipped is information the agent should use next period; an adjustment because the reviewer prefers a round figure is a habit worth challenging. Only approved proposals become accrual journal entries, posted through the ERP's normal workflow so segregation of duties still applies. Each journal keeps its documents, calculation and approval together, which is exactly what auditors ask for when they test an estimate.
6. Accrual reversal, true-ups and tightening each period
Most accruals reverse on the first day of the next period, and that is where the learning sits. When the real invoice arrives, the agent compares it with its proposal and records the difference by supplier, cost type and reviewer adjustment. Persistent over-accrual on a category, such as agencies that bill below purchase order value, should change how that category is estimated. Persistent under-accrual on utilities might point to a tariff change nobody updated. After a few closes, the controller has a measured view of estimate quality instead of an impression.
Accruals also feed everything downstream: the trial balance, flux review and variance commentary all wait on them. On AITHENTIC's agentic month-end close project, the published result is month-end close cut from 5 days to 1 on Oracle AI Agent Studio, and a timetable like that only holds when accrual proposals are ready early. As for where to begin, start with GRNI and recurring service contracts, leave litigation and restructuring provisions with people, and widen scope only once reviewers trust the evidence they are given.
Key challenges
- GRNI balances inflated by stale receipts and purchase orders that were never closed.
- Service accruals based on a budget holder's recollection with no supporting document.
- Prepayment schedules held in personal workbooks, in different formats across entities.
- Round-number accruals that reverse and get rebooked each month without scrutiny.
- No systematic measure of how accurate last period's estimates turned out to be.
The framework
A workable operating model runs in four stages: propose, evidence, approve, learn. In the propose stage, agents scan open purchase orders, receipts, contracts and prepayment schedules through the final days of the period and draft accruals and releases with calculations attached. In the evidence stage, every proposal links to its source documents and states its confidence and the reason for any gap. In the approve stage, proposals are routed by materiality and risk to batch or individual review, and only approved items post through the ERP journal workflow under existing segregation of duties. In the learn stage, actual invoices are compared with reversed estimates, and the controller uses the results to adjust estimation rules and review tolerances.
How to implement it
- Clean the open purchase order book and close stale receipts before go-live so GRNI proposals start from a credible base.
- Write the accrual policy as rules the agent can apply, covering materiality, tolerance against the prior month and categories that always need a named reviewer.
- Load contracts and prepayment invoices for recurring services so the agent can extract billing terms and service periods.
- Run the agent alongside the manual process for two or three closes and compare its proposals with what the team actually booked.
- Switch to agent-first review for GRNI and recurring service accruals while keeping judgemental provisions with people.
- Review estimate accuracy by category each quarter and adjust rules and thresholds based on the evidence.
Frequently asked questions
What is accrual automation?
Accrual automation is the use of software to identify, calculate and propose period-end accruals and prepayment releases from source data. Current approaches use AI agents that read open purchase orders, goods receipts, contracts and invoice history, then draft each accrual with its calculation and evidence. An accountant approves, adjusts or rejects every proposal before anything posts to the ledger.
Can AI calculate accruals without a purchase order?
Yes, for many recurring costs. The agent reads the contract or agreement to find the fee basis, billing frequency and service period, then estimates what has been consumed by period end. Where the cost depends on information outside the documents, such as hours worked by an external adviser, it should flag the gap and request an estimate from the budget holder rather than guess.
How do auditors view AI-generated accruals?
Auditors treat them like any other accounting estimate: they want the basis, the approval and evidence of how accurate past estimates were. If each accrual carries its source documents, the agent's calculation, the reviewer's decision and a comparison with the eventual invoice, the evidence is often stronger than for manual estimates. Human approval and posting through normal ERP controls remain essential.
How are prepayments amortised automatically?
The agent reads the service period from the invoice and contract when the invoice is booked, codes it to prepayments and builds the amortisation schedule at that point. Each month it proposes the release journal for approval and checks that the remaining balance matches the remaining service period. Cancellations, renewals and changed contract dates are flagged for review rather than adjusted silently.
Which accruals should not be automated?
Genuinely judgemental estimates are poor candidates for accrual automation: litigation provisions, restructuring costs, disputed amounts and one-off settlements. They depend on legal views and management intent rather than documents an agent can read. Better early targets are goods received not invoiced balances and recurring service contracts, where evidence is structured and volume makes manual review slow.
