Most AI ERP evaluations fail at the same point: the buyer agrees agents are interesting, agrees the ledger stays deterministic, agrees the pilot was fine, and then cannot name what the system should do on Monday morning. A platform without a sequenced backlog is a demo. The question is which workflows get handed over first, in what order, with which control around them, and what number proves the handover worked.
The order below assumes a controller at a mid-market company with a reasonably clean general ledger, a bank feed, and an AP inbox. It moves from workflows where the ledger consequences of an agent error are small and reversible to workflows where the agent is reasoning across periods and entities. It is also the order in which the evidence a buyer should demand gets harder to fake.
Week One Belongs to Bank Reconciliation
Cash reconciliation is the right first workflow for three reasons. It runs daily, so the agent accumulates feedback inside a week rather than a quarter. The source of truth lives outside the ledger, so a bad match is caught by the next day's statement rather than buried in a journal. And the pain is already measured: 50% of finance teams take six or more business days to close, with cash reconciliation named as the single most time-consuming activity, and teams spending 20 to 50 hours per month on reconciliation alone across three to five systems.
The mechanism is narrow. The reconciliation agent ingests the bank feed, pulls the open items from the cash GL account, proposes matches on amount, date window and counterparty, and posts nothing. A human clears the proposed matches in bulk; the agent writes to a staging table that the ledger pulls from on confirmation. The control is the staging boundary. The metric is auto-match rate against a baseline, measured weekly.
Benchmarks exist, so judge the vendor against them. Mid-market deployments typically reach 80 to 85% initial auto-match rates, climbing to 90 to 95% after the first 90 days as the matching rules learn vendor and payment patterns. A platform that cannot show its own match rate on your statement in a two-week pilot is not ready for a longer commitment. For how to structure that pilot, the piece on a two-week AI ERP pilot covers the scoping discipline.
Week Two Adds AP Invoice Coding and Three-Way Match
Invoice coding is the second workflow because it has the clearest extract → code → check → post loop and the clearest control surface. The AP agent reads the invoice, extracts header and line data, proposes a GL account and cost centre from vendor history, checks the invoice against the PO and receipt, and routes for approval. It never posts above a threshold a human has signed off on.
Adoption is already past the tipping point. 73% of AP departments now use some form of automation for invoice processing, up from 56% in 2022, and the cost spread between leaders and laggards is wide: best-in-class organisations process invoices for $2.36 each versus $10.89 for all others. The gap is almost entirely automation depth.
The control that makes an AP agent safe inside a SOX environment is not the agent itself, it is the approval threshold ladder behind it. Low-value invoices from known vendors with a clean three-way match post straight through. Everything else routes to a reviewer whose approval limit is set in the ledger, not the agent. The write to the ledger is executed by the ledger, not the model. The approval threshold design is where most of the audit conversation lives.
Three metrics prove this one worked: touchless invoice rate (target above 60% by week eight), average coding accuracy on a sampled audit (target above 97%), and cost per invoice compared to the opening baseline. The last of these is the number the CFO cares about.
Week Four Hands Over AR Follow-Up and Cash Application
By week four the AP and reconciliation agents have produced enough daily signal that the team trusts the staging-then-post pattern. The right next workflow is AR, specifically dunning and cash application, because the working capital story is loud. Global average DSO reached roughly 62 days in 2024, up two days year on year, and $1.7 trillion is trapped in excess working capital across U.S. companies, with DSO degrading for a second consecutive year.
The AR agent runs three loops. It drafts and sends dunning messages against an aging policy the controller signs off once. It applies incoming payments to open invoices using remittance text, bank memo and amount, writing proposed applications to the staging table the cash agent already uses. And it drafts collection notes against a confidence score, so the human collector opens the day with a prioritised queue rather than a spreadsheet.
The control here is tone and policy. Dunning templates are reviewed by the controller before the agent is allowed to send unattended, and any message to a top-20 customer routes to a human regardless of value. The metric is DSO movement at 30, 60 and 90 days post-rollout, plus unapplied cash as a percentage of receipts. Both should trend down inside a quarter.
Week Eight Takes On Expense and Accrual Workflows
Expense coding and month-end accrual drafting are the right week-eight workflows because they both require the agent to reason about period, not just transaction. An expense is coded against a trip or project, not just a vendor. An accrual is estimated against a service already consumed but not yet billed. These are the workflows where a bad agent produces a journal that looks right and reconciles wrong three months later.
The mechanism for accruals is the one place the extract → code → check → post pattern gets a fifth step: reverse. The agent proposes an accrual journal, the controller approves, the ledger posts, and the reversal is scheduled automatically for the first day of the next period. For expense, the agent codes receipts against policy, flags out-of-policy items for the manager and never approves its own exceptions.
The control here is sampling. A controller who samples 10% of agent-coded expenses weekly for the first month and 2% thereafter has a defensible audit position. The structural limits on what agents can do inside the ledger matter more here than in week one, because the posting is now closer to judgment than to data.
Week Twelve Earns the Close Agent
The month-end close is the last of the first-ninety-day workflows, not the first. The reason is sequencing: the close agent is useful only if the four prior workflows are producing clean daily signal. If bank reconciliation is still a monthly scramble, closing in three days is a schedule problem the agent cannot fix. The piece on closing faster as a reconciliation problem argues this at length.
The close agent orchestrates rather than posts. It runs the close checklist, flags open items by owner, reconciles subledger to GL for AP, AR and cash, and produces the flux analysis draft against prior period and budget. The controller signs the close; the agent assembles the evidence. The metric is days to close, measured consistently against the baseline, and the quality metric is post-close adjusting entries, which should fall as the agent catches reconciling items earlier in the cycle.
Why This Order and Not Another
Three forces drive the sequence. The first is reversibility: a bad reconciliation match is caught tomorrow, a bad accrual is caught next quarter. The second is signal density: workflows that run daily train the agent and the team faster than workflows that run monthly. The third is data quality exposure. 49% of executives cite data inaccuracies as a barrier to agentic AI, and Gartner estimates poor data quality costs organisations $12.9 million per year on average. Starting with cash exposes vendor master and bank reference issues cheaply. Starting with the close exposes them in front of the audit committee.
Two anti-patterns are common enough to call out. Handing an agent the close in week one skips the reconciliation signal that would have made it accurate. Handing an agent purchase requisitioning before AP coding is stable means the agent is now making spending decisions on a chart of accounts it has not yet learned. Both show up in reference calls, which is why the reference call questions worth asking focus on sequencing more than feature lists.
What a Buyer Should Demand Before Signing
The backlog above is also a scoring rubric. A vendor pitching an AI ERP should be able to demonstrate, on your data in a bounded pilot, the week-one and week-two workflows end to end: match rate on your bank feed, touchless rate on a sample of your invoices, and the audit trail each agent produces. Everything past week four is a design conversation, not a demo. The underlying platform matters more than any single agent, because the staging boundary, the approval ladder and the audit log are platform concerns.
Controllers who sequence the backlog this way end the first quarter with four workflows running under daily measurement and a close that is faster because the inputs are cleaner, not because the close itself was automated. That is the order of operations. Agents earn the harder work by clearing the first problems first.