Blog
Written by

Naman Mathur
Published on

Key Takeaways
Your close is ready for artificial intelligence (AI) agents when every task has a named owner and you have a baseline to judge them against. Clean data can wait.
Run the proof of concept on your one or two busiest entities, in parallel with two live closes and connected read-only, so nothing touches the ledger.
A scale test belongs inside the pilot: add an entity partway through, because 13 entities create 78 intercompany pairs where five create 10.
Before day one, map how each role changes and write go, extend and stop criteria, so the sponsor decides on evidence rather than impressions.
Most close automation pilots prove nothing. They run as demos on a clean sample month, where every match lands and the steering group leaves with an impression instead of evidence.
You have to decide whether to commit your group close to agentic financial close software, and only your own ledger can settle that. Gartner's 2025 finance AI survey of 183 finance leaders found 25% of finance organisations are unsure how to move from planning to piloting.
Our view is that a proof of concept (PoC) should run as a sequence: readiness, scope, parallel run, scale test, change plan and decision gate. Below, we answer how to judge readiness for agents, structure a PoC, run the pilot, test scale by entity and plan the change.
How Do Finance Leaders Decide Whether Their Close Process Is Ready for AI Agents?
Your close is ready for AI agents when every task has a named owner and you know the measures you would judge an agent on. An agent here is software that prepares reconciliations and matches for a person to review. Perfect data is not the bar.
Teams often wait for a data clean-up that never finishes. KPMG's 2026 Global AI in Finance report, a survey of 1,013 senior finance leaders, found 36% cite data quality as their biggest barrier. A pilot on live data finds the weak spots sooner, because exceptions point straight at the accounts and feeds that need fixing.
Check | Why it matters | Ready if |
|---|---|---|
Close calendar documented | Agents need task order and dependencies | Every task has a working-day deadline |
Named preparer and reviewer per task | Segregation of duties (SoD) must survive the change | No task has one person in both roles |
Baseline metrics captured | A pilot without a "before" figure proves nothing | Recent closes measured on the pilot definitions |
Reconciliation policy written down | Agents apply rules, and habits are not rules | Tolerance and evidence set per account type |
Materiality thresholds set | Thresholds define what counts as an exception | Agreed by account, in amount and percentage |
Enterprise resource planning (ERP) and bank access available | The pilot needs live data on day one | Read-only ERP user and bank feeds approved |
Subledgers mapped | Much of the close sits outside the ERP | Stripe, Adyen, Coupa and payroll mapped to general ledger (GL) accounts |
Intercompany counterparties listed | Eliminations fail on unmapped pairs | Every entity pair and its accounts listed |
A sponsor who owns the decision | Someone has to sign go or stop | One named finance leader with the decision date in their diary |
Security review started | In our experience, security reviews stall more pilots than integration does | Vendor questionnaire returned before setup begins |
Source: Stacks implementation practice, 2026.
If two or more checks fail, fix those before you book the pilot. A messy ledger is missing from this list on purpose.
What's a Good Way to Structure a Proof-of-Concept for Financial Close Automation Software?
Structure the PoC around your one or two busiest entities over two full live closes. Connect read-only and agree the success measures before it starts.
Element | What to set | Example |
|---|---|---|
Scope | Entities and processes in the trial | Your one or two busiest entities, covering bank reconciliation and intercompany |
Period | The live closes you run in parallel | Two consecutive month-ends, ideally including a quarter-end, outside audit fieldwork |
Data | Connection and sources | Read-only ERP, bank feeds, Stripe and payroll |
Roles | Who prepares and reviews, and who decides | Current preparers and reviewers, plus one sponsor |
Measures | Baseline figures agreed upfront | The six measures in the next section |
Duration | Setup, then live closes | Setup weeks plus two full closes |
Source: Stacks implementation practice, 2026.
The read-only connection is the first technical step, and planning the ERP integration covers the access and mapping it needs. Time the pilot so it spans a quarter-end, with its late accruals and foreign exchange (FX) revaluation. A quiet month hides the work you are buying the platform for.
Capture the baseline in working days from period end, the same unit as every measure in the next section, so the before-and-after comparison holds.
Keep the pilot clear of audit fieldwork, when your reviewers have no spare hours. A pilot competing with auditor sample requests gets leftover attention and produces thin evidence.
How Should Finance Teams Pilot a Close Automation Platform Before a Full Rollout?
Pilot the platform in parallel with your live close, connected read-only, so nothing touches the ledger. A parallel close means the platform works the same period as your team, and you compare its output with the books you sign. Whichever platform you test, your team signs the real close the usual way until the pilot numbers earn trust.
What to Track Each Period
Record each measure for the baseline closes and again for every pilot close. Use identical definitions in both, or the comparison means nothing. Auto-match rate and exception volume start with the pilot, so compare them across the two pilot closes.
Measure | Definition | What it tells you |
|---|---|---|
Auto-match rate | Share of transactions matched without a person | How much routine work the agent absorbs |
Exception volume | Items the agent routes to a reviewer | Whether tolerance rules are set correctly |
Open reconciliations at day 3 | Accounts not reconciled by working day 3 | Where the backlog sits |
Days to first trial balance (TB) | Working days from period end to first TB | Progress before consolidation starts |
Reviewer hours per close | Time reviewers log on pilot entities | Whether work moved or only changed hands |
Reconciliations signed off by target day | Share signed off by the working day you set | Fit with your close calendar |
Source: Stacks implementation practice, 2026.
Then break it on purpose. Force a short pay and an intercompany balance outside your FX tolerance. Both should land in the exception queue with a stated reason.
Export the audit trail and give it to whoever runs your audit walkthroughs, since staying audit-ready through the period depends on that trail. If they cannot follow who prepared and who approved each item, the pilot has failed on controls.
How Do You Evaluate Whether a Close Automation Platform Will Scale With a Growing Entity Count?
Build a scale test into the pilot. Add another entity partway through, ideally in a different currency and ERP, and time how long it takes to go live without re-implementation.
Entity count matters because intercompany work grows faster than the group does. Possible intercompany pairs grow roughly with the square of the entity count: five entities give 10 pairs, and 13 give 78. Run these tests on the added entity:
Does the new entity inherit the close calendar, or does someone maintain a copy by hand?
Set a stricter local approval threshold for it and confirm it overrides the group default.
Connect a second ERP, such as Xero or Sage, alongside your main system and reconcile across both.
Finally, run one intercompany pair in two currencies and check the FX difference reaches the right reviewer.
For a reference point, Epidemic Sound ran its first close on Stacks across 13 legal entities and 102 bank accounts, and completed all 713 close tasks. Details are in the Epidemic Sound case study. If the entity you add mid-pilot needs a new implementation project, expect the same for every acquisition.
How Should Finance Teams Assess Change Management Needs Before Adopting Close Automation?
Assess change management by mapping who changes what, role by role, before the pilot starts. Every role in the close shifts, in a different direction.
The controller's biggest new responsibility is autoposting, which lets an agent post journals without manual approval on accounts you choose. The table shows the shift for each role.
Role | Today | With agents | What they need |
|---|---|---|---|
Preparer | Builds reconciliations and matches | Reviews agent output and resolves exceptions | Training on exception reasons and overrides |
Reviewer | Re-checks spreadsheets line by line | Approves exceptions and signs off | Evidence attached to every item |
Controller | Runs the close calendar | Owns thresholds and autoposting decisions | Authority to change tolerances, with a log |
Auditor | Tests manual controls | Walks through the changed control | A walkthrough before year-end fieldwork |
Source: Stacks implementation practice, 2026.
Sponsorship holds this together. Prosci's change management research finds projects with excellent change management are about seven times as likely to meet objectives, and names active, visible sponsorship as the biggest single contributor. In our experience, pilots stall when nobody is named to make the call.
Book the auditor walkthrough before the pilot ends, because a changed control explained in advance costs far less than one discovered during fieldwork.
Why Standard Pilot Advice Fails
Most pilot advice is built for a smooth demo. The common playbook has five habits, and each hides the failure you most need to see:
Scripted demos show matches that would have cleared anyway.
With clean sample data, the exceptions that consume reviewer time never appear.
"Clean your data first" delays the pilot by months and leaves the bad feeds undiscovered.
When the team judges on impressions, the most confident presenter wins.
Pilots without exit criteria drift into an unfunded second phase.
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. A demo tests neither business value nor risk controls.
The fair counterargument is that a sandbox on sample data is lower risk. It is, but it hides your exceptions. A sandbox loaded with your own data, or a read-only connection to the live ledger, keeps that safety and shows you the real exception queue.
So spend most of the pilot in the exception queue. Reviewer time goes there, and so does your judgement of the platform.
From Pilot to Full Rollout: The Go/No-Go Decision Gate
Decide against the baseline, using criteria written down before the pilot started. Criteria agreed after the results turn the decision into a negotiation.
Criterion | Go | Extend | Stop |
|---|---|---|---|
Auto-match rate | At or above the agreed target | Rising across two closes | Flat or falling |
Open exceptions at close | Falling across the two closes, each with a reason | Not yet falling, reasons clear | Unexplained exceptions remain |
Reviewer hours per close | At or below baseline | Above baseline but falling | Above baseline and rising |
Audit trail | Auditor accepts the walkthrough | Gaps fixable before fieldwork | Preparer or approver missing |
Added entity | Live without re-implementation | Live after configuration work | Needed a new project |
Source: Stacks implementation practice, 2026.
After a go, roll out by entity or by process, whichever matches how your close team is organised. Enable autoposting account by account as confidence builds, starting with high-volume, low-risk accounts such as bank charges.
This gate tests control and accuracy. Speed comes later, and cutting close time after adoption is a separate exercise.
Why Choose Stacks for a Low-Risk Close Pilot
Stacks connects to the ERP you already run, so there is nothing to migrate and nothing to replace. This is how Stacks runs implementation: 1–2 hours of team time in week one, then about 2 hours a week in weeks two and three. Users are onboarded in a sandbox built with your own data, so the first exceptions they see are real ones.
Standard agents reach production in 4–8 weeks and custom agents take 2–4 close cycles (see our guide to enterprise close platforms). Rollout is phased by entity, so the added-entity scale test is part of the normal plan. Epidemic Sound signed on 29 August 2025 and closed on Stacks on 6 October 2025.
Stacks close management keeps your preparer and reviewer workflow, with enforced SoD. GL posting gates hold each journal until an approved reviewer releases it. Stacks also includes Workspace Agents for company-specific processes you configure yourself.
The pilot runs on your calendar. Yoto onboarded in November 2025 and went live for its February 2026 close, a timeline set by its choice to keep going through peak trading and audit season. The detail is in the Yoto case study.
FAQs for Close Automation Pilots
How Long Should a Close Automation Pilot Run?
Plan for the setup weeks plus two full closes. One close shows the platform works, and the second shows whether the results hold.
Which Close Processes Should We Pilot First?
Start with high-volume, rules-based work such as bank reconciliation and intercompany matching on your busiest entities. They produce the most exceptions, so they test the platform hardest.
Do We Need to Replace Our ERP to Run a Pilot?
No. A pilot should connect read-only to the ERP you already run, so the ledger stays untouched until you decide to go.
Is Our Data Clean Enough to Start?
Probably not, and that is fine. A pilot on live data shows which feeds and accounts need fixing sooner than a standalone clean-up project.
How Do We Keep Approvals Under Control During a Pilot?
Keep your existing preparer and reviewer assignments and leave autoposting off. Nothing should post to the ledger until a named reviewer approves it.
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, Epidemic Sound
Trusted by fast-growing companies including: Datavant, Future, Cleo, Motorway, Nivoda, Yoto and Epidemic Sound.

