Blog
Written by

Naman Mathur
Published on

Key Takeaways
Segregation of duties holds when the agent has preparer rights only, a named accountant reviews, and the platform blocks anyone approving their own work.
In our reconciliation test, general-purpose agents gave different answers to the same data in 22 of 24 cases, so rerunning the same match is the most revealing vendor test.
Checklist tools track the close without doing any of it, and Provision 29 declarations from 2026 raise the bar for approval evidence.
Enterprise resource planning (ERP) add-ons with embedded artificial intelligence (AI) suit a group on one ERP. Multi-ERP groups need agentic close controls that span every ledger.
Control in an agentic close comes from who approves the work, and from a system that refuses to post anything until they have. The standard advice, keep a human doing the preparation, keeps the bottleneck and adds no control a reviewer couldn't provide faster. That approval design is what we mean by agentic close controls.
If you want the basics first, start with our controller's definition of agentic close. Below, we cover control, one task step by step, segregation of duties, agent washing, adoption drivers and AI-native platforms versus ERP add-ons.
How Do Finance Teams Maintain Control and Oversight When AI Agents Handle Close Tasks?
Finance teams keep control by letting agents prepare and people approve, with the system enforcing the split and logging every action. The agent drafts the reconciliation or journal. A named accountant reviews it, and nothing reaches the general ledger (GL) until they do.
Most guidance offers a human-in-the-loop close as design advice. Written as a system rule, it survives the last two days of close, when someone would otherwise approve in bulk to hit the deadline.
Three mechanisms make the rule hold. The first is role separation: the agent is a user with preparer rights and nothing more, so it cannot sign off its own output. The second is a posting gate, which stops a journal reaching the GL until the required approvals and supporting evidence are attached.
The third is evidence captured as the work happens. Each match, edit and approval is logged with the person, the change and the time, so the audit file builds up during close.
Most teams can't show this yet. KPMG's 2026 AI in Finance survey found: "Fewer than half of organizations (42 percent) are fully assurance-ready for AI-enabled finance processes." In our experience, the gap is wider for agents.
If you can't evidence who approved an AI output, your auditor can't rely on it. Before you pilot agents, write down which approvals your policy requires for each task type. That list becomes your posting gate configuration.
What Does an AI-Powered Close Process Actually Look Like in Practice?
An AI-powered close process in practice is one task moving through six controlled steps. The agent does the preparation, and a person owns the approval. Take a month-end bank reconciliation for a UK operating entity.
This is how Stacks accounting agents handle it. Use the same six steps to test any platform you evaluate. The table shows who acts and what an auditor can test.
Step | Actor | Action | Control | Evidence |
|---|---|---|---|---|
1 | System | Syncs bank and ERP balances | Source and time logged | Sync record |
2 | Agent | Proposes matches with reasoning | Preparer rights only | Reasoning per match |
3 | Agent, then named owner | Routes unmatched items | Named exception owner | Exception log with owner |
4 | Reviewer | Accepts, edits or rejects matches | Reviewer differs from preparer and configurer | Decision and reason recorded |
5 | System | Blocks posting until cleared | GL posting gate | Approvals and evidence attached |
6 | Controller | Signs off and locks the period | Automatic reopen on balance change | Sign-off and reopen record |
Source: Stacks close management workflow, 2026.
Steps 1–3: Data Lands, the Agent Prepares, Exceptions Escalate
The task starts when data lands. The ERP sync records its source and time, so a reviewer can see which ledger balance the reconciliation was built on. A late journal after that sync then shows up as a changed balance.
The agent proposes matches, each with its reasoning: the reference, the amount, the counterparty and how close the dates sit. Underneath, deterministic rules handle the exact matches, so the agent's judgement goes on the ambiguous items. That layering matters because an exact match should give the same answer every time.
Anything the agent can't match goes to a named owner for that account. A bundled receipt, a short payment or an unexplained bank fee lands with the person responsible. Ask vendors who owns an exception on day two of close, and expect a name.
Steps 4–6: Review, Posting Gate, Period Lock
The reviewer opens each proposed match and accepts, edits or rejects it. Every edit or rejection records a reason, and that reason is what an auditor samples when testing the review control.
Posting is blocked until the approvals and evidence your policy requires are present. The adjusting journal waits at the GL posting gate, and the agent has no way to push it through. The audit trail for AI agents shows the preparer, the reviewer and the time of approval on the same record.
Sign-off locks the period. If someone posts a late accrual and the ERP balance changes after sign-off, the task reopens and goes back to the reviewer. That closes the gap where a signed reconciliation no longer agrees to the ledger it was signed against.
Test this in any demo. Sign off a reconciliation, post a journal to that account in the ERP and watch what happens. Our approvals and audit trail checklist lists the other checks worth running.
How Is Segregation of Duties Maintained When AI Agents Execute Close Tasks?
Segregation of duties (SoD) is maintained by treating the AI agent as a user with preparer rights only. It can draft, match and propose. It cannot review, approve, post or lock.
Two rules follow. The agent can't approve its own output, and whoever configures an agent can't approve what that agent produces.
The second rule is the one we see teams miss most often. If the person who set an agent's matching tolerances also reviews its matches, they are checking their own work at one remove. The Committee of Sponsoring Organizations of the Treadway Commission (COSO) addressed this directly in February 2026.
COSO's GenAI internal control guidance recommends: "Separate the ability to configure AI settings from the authority to approve or review outputs." The role matrix below applies that principle to the close.
Role | Prepare | Review | Approve and post | Configure agent | Lock period |
|---|---|---|---|---|---|
AI agent | Yes | No | No | No | No |
Reviewer | No | Yes | Yes | No | No |
Agent administrator | No | No | No | Yes | No |
Controller | No | Yes | Yes | No | Yes |
Source: Stacks implementation practice, 2026; COSO generative AI (GenAI) guidance, February 2026.
Then watch for control drift. An administrator takes reviewer rights for one busy month, and nobody removes them. Run a quarterly access review of who can configure, review, post and lock, and treat any overlap as a deficiency to remediate.
How Can Finance Teams Tell if an AI-Powered Close Tool Is Genuinely Agentic or Just Automated Workflow?
You can tell whether an "AI-powered" close tool is genuinely agentic by testing what it does with real work. Agent washing is common, and many automated workflows have simply been relabelled.
Gartner's agent-washing warning from June 2025 is blunt: "Gartner estimates only about 130 of the thousands of agentic AI vendors are real." Run these five tests on your own data before you sign.
Does it complete the task or just tick it? The reconciliation should arrive prepared, with matches proposed, before anyone opens it.
Does it handle an exception it wasn't scripted for? Feed it a short payment or a bundled receipt and see whether it proposes an answer.
Does it give the same result twice? Rerun the same match on the same data and compare every line.
Does it show reasoning per item? Each match should state why, in terms a reviewer can check against the source document.
Is it bound by your controls? Its output should hit the same posting gates and approvals as work prepared by a person.
Repeatability is where general-purpose tools fall down. In our reconciliation reliability test, three general-purpose agents ran eight real reconciliation datasets ten times each. In 22 of 24 cases, they did not give the same answer twice.
On the hardest dataset, one agent produced 7–302 wrong matches depending on the run. False matches can still balance to the cent, so a reconciliation that ties out tells you nothing about whether the matching was right.
Robotic process automation (RPA) is a separate comparison, covered in how agents differ from RPA. The five tests here ask a narrower question: does the tool fit your control environment?
What's Driving Finance Teams to Adopt Agentic AI Instead of Traditional Close Management Software?
Finance teams adopt agentic AI over traditional close management software because checklists track work and do none of it. A checklist tells you a reconciliation is overdue. Someone still has to prepare it.
Entity growth makes that gap expensive. In our experience, each acquisition adds bank accounts, intercompany pairs and statutory filings, and hiring rarely keeps pace. Review gets squeezed into the final days.
UK regulation also raises the bar for evidence. The Financial Reporting Council (FRC) sets the UK Corporate Governance Code. Provision 29 asks boards of companies reporting under it to declare whether material controls were effective. FRC's Provision 29 mythbuster confirms the declaration applies to periods commencing on or after 1 January 2026.
A board will struggle to make that declaration on the back of spreadsheet sign-offs nobody can reconstruct. Agentic close controls give the board an approval record it can rely on.
Customer results show what changes. Epidemic Sound went from contract on 29 August 2025 to a fully live close on 6 October 2025, across 70+ bank accounts and 13 entities, with 713 tasks at 100% completion. The team now saves about 5 hours per close on approvals.
In 2025, Nivoda cut its close from 20+ days to 12 and its journal posting from 8+ hours to 15 minutes, and now runs 95% of its reconciliations automatically.
That is why we built agentic AI to work inside your processes. Before any pilot, measure your current close length, review hours and unmatched lines, so you can compare like for like afterwards.
How Do AI-Native Close Platforms Compare to Add-On Automation Modules From Legacy ERP Vendors?
An AI-native close platform works across every ledger you run, while an ERP add-on module works best inside its host ERP. Which fits depends on how many systems hold your numbers.
The honest case for the add-on is strong. You deal with one vendor, and your auditor already knows the system. Gartner's cloud ERP close prediction, from February 2026, forecasts a "30% faster financial close by 2028" for cloud ERP users with embedded AI assistants.
For a group on one ERP, that makes embedded AI a credible route. The limits show up in group structures. In our implementations, many European groups run several ERPs after acquisitions, such as Xero in a UK subsidiary, Fortnox or Visma in a Nordic one, and a tier-one suite at group level.
Subledgers often sit outside the ERP entirely, in Stripe, Adyen, Chargebee, Coupa, Pleo or payroll. In our experience, an add-on's controls stop at its own ledger. Intercompany eliminations, cross-ERP reconciliations and subledger tie-outs fall outside its reach, so the riskiest work stays in spreadsheets.
Factor | AI-native close platform | ERP add-on module |
|---|---|---|
Ledger coverage | Every ERP in the group | The host ERP only |
Subledgers outside the ERP | Reconciled in the same close | Often left to spreadsheets |
Control scope | One control set across entities | Controls scoped to one ledger |
Vendor relationship | An additional vendor to assess | Your existing ERP vendor |
Auditor familiarity | A new system to walk through | Already known to your auditor |
Best fit | Multi-ERP or acquisitive groups | Single-ERP groups with subledgers inside |
Source: Stacks implementation practice, 2026.
For the wider comparison, read agentic versus legacy close platforms. The practical test is quick: list every system that feeds your trial balance (TB), then check which option covers all of them.
Why Standard Advice on AI Close Controls Fails
Two pieces of standard advice dominate AI agent governance in finance, and both fail. The first is "keep a human doing it", which puts your reviewer back in the preparer's seat.
You keep the fatigue and lose the speed, and the person checking the work is often the person who did it.
The second is "bolt a large language model (LLM) onto the existing process". An LLM wired to a spreadsheet produces plausible matches with no repeatability, no reasoning record and no posting gate.
Both fail for the same reason. They put accountability in a policy document, and a policy document can't stop a journal posting at 11pm on day three.
Audit regulators have already drawn this line. Published on 30 March 2026, FRC's agentic AI guidance states that "the human auditor is always accountable". Finance should hold the same line, with a named controller owning every agent output.
Accountability belongs in the system. Name the owner on each task, block posting without their approval and keep the log where your auditor can sample it.
Why Choose Stacks for Agentic Close Controls
Stacks builds agentic close controls into the close. Stacks close management software enforces segregation of duties through a preparer/reviewer workflow. Preparers prepare, reviewers review, and nothing posts to the GL until it has cleared the controls your policy requires.
Sign-off locks the period, and a task reopens if the ERP balance changes afterwards. Every action sits in a continuous audit trail your auditor can follow.
Agents prepare reconciliations, matches and journals as data lands, and your team handles the exceptions and the sign-off. AI matching typically starts at 80–85% of transactions in the first month and reaches 95–97% within three to four months, where rules-based matching tends to stall at 75–80%. Workspace Agents let your team configure its own recurring processes, such as accrual preparation, inside the same close.
For your audit committee, Stacks holds System and Organization Controls (SOC) 2 Type II and ISO 27001 certification, with Sarbanes-Oxley (SOX)-ready controls. It connects to the ERPs you already run, so there is nothing to migrate and nothing to replace.
FAQs for Agentic Close Controls
Do AI Agents Post Journals Without Approval?
Not in a controlled close. The agent prepares the journal, and it waits at the GL posting gate until the approvals and evidence your policy requires are attached.
Can an AI Agent Approve Its Own Work?
No. The agent holds preparer rights only, and the person who configures it shouldn't approve its output either.
Will Auditors Accept Agent-Prepared Work?
Auditors test the review control, so they need evidence of who approved each item, what changed and why. Agent-prepared work with that trail can be tested in the same way as work prepared by a person.
Is Generative AI the Same as Agentic AI?
No. Generative AI drafts answers on request, while an agent completes a task inside your process. We cover the difference in how generative AI is used in finance today.
Do I Need to Replace My ERP to Use AI Agents in the Close?
No. An AI-native close platform connects to the ERPs you already run, and each ERP stays your system of record.
See How Stacks Maps to Your Close Process
Book a 30-minute walkthrough. We'll discuss your close process, show you which tasks run automatically, and outline the next step from there.
"If you want to achieve an agile, transparent, and significantly faster close without adding headcount, Stacks is the way to go."
Jessica Lindholm, VP of Accounting at Epidemic Sound
Trusted by fast-growing companies including: Datavant, Future, Cleo, Motorway, Nivoda, Yoto and Epidemic Sound.

