Every month end, finance drops everything to close. Bank recs, intercompany matching, accrual reversals, explaining why an account jumped. By the time it is done the numbers are already stale, and the cycle starts again. Meanwhile the business keeps adding entities, markets and policies, and the team does not grow with it.
A playbook from Nominal, a US finance software company, puts a name to a different way of working: Agentic Performance Management, or APM. The idea is to have AI agents execute the in-between work of accounting while people set the rules and approve whatever falls outside them. Below is our read in plain operations language, and what we take from it if you want agents working in your finance function.
Where this comes from
This is vendor material, so the case study figures are the ones the vendor reports. The logic does not depend on any particular tool, though, and that is what we care about here.
What changes: from system of record to system of action
The ERP stores the entries. EPM and FP&A tools structure the reporting. Neither executes: matching, chasing discrepancies, preparing the adjusting entry and following it through is still people’s work. AI copilots help with pieces, but they stop at a suggestion.
The playbook describes a layer of agents on top of the ERP you run today, with no migration. Agents take the transaction data, match, eliminate, prepare entries and flag exceptions; clean numbers flow on to planning; and the results are written back to the general ledger.
The gain is not doing the same work faster. It is doing work nobody could do before, like reviewing every entry instead of a sample.
The four principles
- Execution, not suggestions. The agent finishes the workflow; the controller moves from doer to reviewer and saves judgment for the calls that need it.
- A continuous close. “Done” becomes a daily state, not a month-end moment. The books stay close-ready in near real time.
- No system swap. It works across multiple ERPs, entities and currencies, like a control tower over what is already there.
- Human in the loop. Rules and approval thresholds written in plain language, escalation defined by the team, and full documentation at any time.
Seven use cases
The playbook groups the work into seven areas. Each follows the same shape: a problem that repeats, an agent that executes, and a person who approves the exceptions.
- Continuous bank reconciliation. Cash matching runs all month; suggested adjustments go for approval and, once approved, post to the ledger.
- Transaction patrol. Scanning for missing, misclassified, duplicate or split entries meant to dodge approvals, including repeated under-billing of a customer.
- Continuous close. Matching, flux analysis and policy checks happen every day; only real issues get escalated.
- Trigger-based GL automation. The agent “listens” for ERP events, like a vendor bill that calls for an accrual reversal, and proposes the next step right away.
- Intercompany at scale. Intercompany activity matched from plain-language instructions, with context from both sides when it does not tie out.
- Flux analysis on every account. Instead of asking the team to explain “what changed,” the agent traces each variance back to the transactions behind it.
- Audit and acquisition-ready. Every agent action and every human approval is logged; you can drill from any balance to its source in a few clicks.
The numbers the vendor reports
The playbook includes three customers. These are Nominal’s reported results, but they give a sense of scale:
In the playbook’s comparison, the close goes from 8–12 days to 1–3 days, and team focus shifts from 85% execution to 75% strategy.
- Leanpay (buy now, pay later, five countries): month-end close 20–25% shorter and weeks saved on year-end consolidation.
- Green Street Power Partners (solar, 280+ project entities): more than 60 hours saved per month.
- Kunai (IT consultancy preparing for a sale): 75% of anomalies resolved automatically, and the acquisition closed without a new ERP, on top of QuickBooks and Excel.
Trust comes first
In finance, controls are not optional. The design the playbook describes is the same one we use in any area: the agent prepares within rules and thresholds, executes what falls inside them, sends exceptions to a person, and everything lands in the audit trail.
For a CFO, this is the internal-controls language they already speak: segregate the preparer from the reviewer. The difference is that the preparer is now an agent, and review can cover the full population.
How to tell real agents from “AI-powered” tools
“AI-powered” can mean anything. The playbook suggests three questions for any demo, from any vendor, us included:
- Who does the work? Does the tool only suggest actions for someone else to take, or does it finish the workflow and post to the ledger with approval?
- Do you have to switch systems? Does it work with the ERP, spreadsheets and entities you run today?
- How is it governed? Can your team configure the rules? Can every action be traced and audited? Is the AI preparer clearly separated from the human reviewer?
And three red flags: a rules engine dressed up as AI that only routes tasks; a copilot that cannot write to the ledger, so someone still touches every transaction; and a heavy reliance on consultants for every new entity or policy.
What this means in practice
The lesson goes beyond finance: the model is the easy part. The work is in connecting the right sources, writing the rules, setting approval thresholds and keeping the audit trail. And you can start small, with a single workflow, without replacing your ERP.
- The AI Analyst maps where the close gets stuck, which reconciliations repeat, and where sampling lets revenue slip through.
- The Forward Deployed Engineer builds the agent inside your operation: ERP and spreadsheet connections, plain-language rules, thresholds, human approval and a log of every action.
- Training teaches the team to review instead of execute, which is the role change the playbook describes.
Reconciliations, accruals and intercompany are just examples. If a task repeats, in finance or anywhere else, it can be automated.
The close does not have to be a month-end event. With agents doing the work and people reviewing it, it becomes a daily state.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.