The Continuous Close, Honestly Assessed

The most useful finding on close automation is an absence: one named, self-reported case in the entire category. This guide sets out what can actually be evidenced, why the binding constraint is evidentiability not accuracy, and the approval-gate design pattern and checklist that make a faster close defensible.

The most useful finding in this topic is an absence.

We went looking for named enterprises with an independently verified, quantified close-cycle reduction attributable to agentic AI.

We found one self-reported figure and nothing else.

Hewlett Packard Enterprise reports a 40% faster financial reporting cycle from its internal agent "Alfred", built with Deloitte and NVIDIA, piloted in 2025 and scaling through 2026 across weekly operational performance review, AP/AR, and credit and collections (CFO Dive). It is self-reported, it is a reporting-cycle figure rather than a close-cycle figure, and it is the strongest named data point available. Every other quantified case in this category — BCG's, the close vendors' — is anonymized, and most are vendor-authored.

That absence is the finding. The financial close is the most heavily marketed finance AI use case in the market. If verifiable named references existed at any scale, you would have seen them by now, because the vendors selling into this category have every commercial incentive to publish them.

Gartner's own guidance quietly confirms the scarcity. Analyst Mike Helsel advises that "CFOs should insist on industry-specific features, transparent pricing, and reference-able customer adoption for AI tools" (CFO Dive). You do not need to tell buyers to insist on reference-able customers in a market where reference-able customers are easy to find.

This guide sets out what the close automation category can actually evidence today, why the binding constraint is not accuracy but evidentiability, and the design pattern and checklist that make a faster close defensible.

What is marketed versus what is evidenced

| Claim you will hear | What the evidence actually supports |
|---|---|
| "Continuous close — recognize and reconcile as transactions occur" | Reaching limited production at leaders. A genuine real-time close is a 2029–2030 proposition on BCG's own stated horizon (BCG, June 2026) |
| "30% faster close" | A Gartner forecast for 2028, not a measured outcome — see below |
| "Autonomous journal entry creation" | Real capability, but BCG's own published design pattern requires human approval for entries above $25,000 |
| "Our agent does reconciliation" | Gartner estimates only ~130 of thousands of self-described agentic vendors are genuine; reconciliation tools have a 20-year RPA and rules-engine heritage (Gartner, 25 June 2025) |
| "Proven at enterprise scale" | One named, self-reported case. Everything else anonymized |

The Gartner number, labeled properly

Gartner predicts that embedded AI in cloud ERP applications will drive a 30% faster financial close by 2028, with AI-enabled tools rising to 62% of cloud ERP spending by 2027, from 14% in 2024 (Gartner, 24 February 2026 | CFO Dive).

This is a forecast. It is not a measured result, and it is not a benchmark. It describes what Gartner expects the market to achieve two years from now, on average, from capability embedded in ERP software rather than from bespoke agent builds. It is being quoted in vendor decks as though it were an observed outcome. It is not.

Read alongside Helsel's own caution and it becomes something more useful than a target — an argument about route. If the 30% arrives through embedded ERP capability, then for most mid-market organizations the close-acceleration question is substantially a question about when the ERP gets upgraded, not about when AI gets "adopted". That reframing saves a lot of money.

What is genuinely measured

Two categories of real evidence exist.

Sector survey data. In banking, 76% of respondents report moderate or significant close-efficiency improvement and 78% report error reduction (KPMG Global AI in Finance 2026, n=1,013 senior finance leaders across 20 countries at companies above US$250m revenue, fielded March 2026). This is self-reported perception of improvement across a large, well-constructed sample — genuinely useful, but not the same thing as a measured cycle-time reduction in a named company.

Capacity measurement. BCG puts general accounting and close at a 20–30% capacity unlock today, with a further 20–30% available long term (BCG, June 2026). Note that this is capacity, not cycle time. Freeing 25% of a controller's month does not necessarily shorten the close by 25% — it depends entirely on whether the freed capacity sits on the critical path.

And one uncomfortable structural finding. KPMG's headline conclusion from those 1,013 leaders is that "the strongest improvements are in decision-making quality, forecast accuracy and responsiveness. These are judgment-heavy areas, not transactional processes" — transactional processes improved "at smaller margins." The close is the most-marketed use case in finance AI and sits on the lower-return side of that finding. That does not make it not worth doing. It makes it worth doing with realistic expectations and a controlled budget.

The binding constraint is evidentiability, not accuracy

Most close automation programs are scoped as if the risk were the agent getting the answer wrong. That is a real risk, but it is not the one that stops you.

The one that stops you is being unable to demonstrate how the answer was produced.

Only 42% of organizations are strongly assurance-ready, rising to 60% among agentic AI leaders. And KPMG flags "limited audit trail documentation" as a key vulnerability, observing that most organizations track outcomes but not evidence-production methods (KPMG, March 2026). That is precisely the gap an auditor will find.

The regulators have already closed the obvious escape route. The UK Financial Reporting Council's generative and agentic AI factsheet, published 27 March 2026 identifies three risk categories (misuse of output; deficient output; non-compliant methodology) and four mitigation categories (system design and development; certification; staff education and governance; human oversight and review). Critically, it confirms that firms and engagement partners retain full audit quality responsibility under ISQM (UK) 1 and ISA (UK) 220 (FRC). The key sentence: "Auditors who use AI cannot blame the tech for mistakes." (Compliance Week).

FINRA's 2026 Regulatory Oversight Report names auditability explicitly among seven risk categories for autonomous AI systems, warning that multi-step agent reasoning makes outcomes "difficult to trace or explain" (FINRA).

Meanwhile your internal auditors are already moving: 76% of internal audit teams now test their AI models (Workiva Executive Benchmark 2026, n=1,497).

A faster close you cannot evidence is not a faster close. It is an unsupported assertion that will be unwound during fieldwork, at which point your cycle time goes back up and your credibility does not come with it.

There is an upside to taking this seriously. Assurance-ready organizations outperform the rest by 3x to 6x across metrics, and on error reduction 33% report significant improvement versus 6% — a 5.5x difference (KPMG, March 2026). The controls are not a tax on the benefit. They correlate strongly with getting the benefit at all.

The approval-gate design pattern

The pattern that wins is already settled, and it has three parts.

1. Agent proposes. Human approves above materiality. Every action leaves a trail.

2. Set dollar thresholds, not principles. "Human in the loop" is not a control; it is an aspiration. BCG publishes the emerging standard: human approval required for journal entries over $25,000, human sign-off on material account movements and unusual variances, and human sign-off on financial statements described as "nonnegotiable" (BCG, June 2026).

Do not adopt $25,000 because BCG published it. Derive your own thresholds from your own materiality, and set them at multiple levels — a single threshold is almost always wrong. A workable structure:

| Tier | Example rule | Control |
|---|---|---|
| Tier 1 — routine | Recurring entries and reconciliations below the threshold, matching a known pattern | Agent posts. Sampled review monthly. Full log retained |
| Tier 2 — threshold | Entries above your dollar threshold, or unusual account combinations | Agent proposes, named human approves before posting |
| Tier 3 — judgment | Accruals, provisions, impairment inputs, manual top-side entries, anything affecting a key judgment | Agent may prepare supporting analysis only. Human authors and owns the entry |
| Tier 4 — non-negotiable | Financial statements, disclosure narrative, going-concern assessment | Human sign-off, always. No exceptions configured |

Write the thresholds into system configuration, not into a policy PDF. A threshold enforced by a document is a threshold that will be missed under close pressure at 9pm on working day three. Review them quarterly, and log every override with a reason.

3. Log evidence production, not just outcomes. Your minimum record per agent action: the inputs consulted, the model or system version, the rule or reasoning path invoked, the human who reviewed it, the timestamp, and what changed on review. If your vendor cannot produce that record, the tool cannot be used above Tier 1. Make it a procurement requirement rather than an implementation discovery.

If anyone on your team thinks this is over-engineering, point at the firms selling AI transformation. Deloitte Australia delivered an AU$440,000 report to the Department of Employment and Workplace Relations containing a fabricated quote from a federal court judgment and citations to academic papers that do not exist. Deloitte agreed to repay the final contract instalment and republished with a retroactive disclosure that Azure OpenAI GPT-4o had been used (CFO Dive | OECD.AI Incident Registry).

Engage your auditor in week one

Not at year-end planning. Not during fieldwork. Week one.

The reason is structural. If you design the close, deploy the agents and then discover that your external auditor will not place reliance on the evidence your system produces, you rebuild the control environment at three times the cost and you lose a reporting cycle doing it. If you agree the evidence specification with them before you build, you build once.

What to put in front of them in that first meeting:

  • The process map of the close as it exists today, with the intended agent-assisted steps marked and the critical path identified
  • The agent inventory — what each agent does, what data it touches, what it may decide unaided
  • The threshold table above, populated with your own materiality figures
  • A sample evidence record for one agent action
  • Your exception and override policy, including who can override a gate and how that is logged
  • Your failure log — only 29% of organizations formally track where AI adoption fails (KPMG, March 2026)

Then ask them directly: "What would you need to see to place reliance on this?" Write the answer down and build to it. It is the cheapest de-risking available in the whole program.

The assurance-ready close checklist

Work through this before, during and after implementation. Anything unticked is a finding waiting to happen.

Before you automate anything

  • [ ] The close is value-stream mapped, with the critical path identified
  • [ ] Non-value-adding steps have been eliminated or simplified before automation is considered
  • [ ] Baseline captured and signed: working days to close, by entity; number of manual journals; reconciliation exception rate; post-close adjustment count; restatement history
  • [ ] Materiality thresholds agreed with the Controller and documented at each tier
  • [ ] The external auditor has seen the design and stated what they need to place reliance
  • [ ] Vendor tested against Helsel's criteria — industry-specific features, transparent pricing, reference-able customer adoption

As you build

  • [ ] Approval gates enforced in system configuration, not in a policy document
  • [ ] Evidence record specified and demonstrated for one full cycle: inputs, version, reasoning path, reviewer, timestamp, change on review
  • [ ] Agent inventory published, with authority level and named human owner per agent
  • [ ] Override path defined, restricted to named individuals, and logged with reasons
  • [ ] Segregation of duties re-tested — an agent with both preparation and posting rights has quietly collapsed a control you spent years designing
  • [ ] Rollback procedure tested: what happens when an agent posts a run of incorrect entries on working day two
  • [ ] Failure log created, with the expectation stated that it will contain entries

Before you claim a result

  • [ ] Cycle time re-measured against the signed baseline, by entity, over at least three consecutive closes
  • [ ] Post-close adjustments and error rates compared to baseline — a faster close with more adjustments is not an improvement
  • [ ] Internal audit walkthrough completed end to end, with findings closed or scheduled
  • [ ] External auditor has confirmed reliance, in writing, on the evidence the system produces
  • [ ] Underperformance documented alongside the wins, in the same pack
  • [ ] Freed capacity redirected to named work with named people, not banked as an unallocated saving

The honest bottom line

Close automation works. The capability is real, the capacity release is measurable, and the sector-level improvement data is credible. But it is the most over-marketed use case in finance AI, it sits on the lower-return side of KPMG's central finding, its headline "30% faster" number is a forecast for 2028, and it has almost no named, verifiable references.

Buy it accordingly: modest budget, embedded-in-ERP first, thresholds and evidence trail designed in week one, auditor engaged before the build, and your own baseline as the only benchmark you trust.

EZ Consulting delivers this as the Assurance-Ready Close engagement — a Lean Six Sigma close diagnostic that value-stream maps the close and eliminates before automating, followed by either a remediation plan or a managed close service with agreed cycle-time SLAs and a documented evidence trail. The two-day "Lean Six Sigma for the Financial Close" workshop and the Assurance-Ready Close Checklist are available separately.

Sources

Version August 2026 · Last reviewed July 31, 2026 · https://ez.consulting/ai-lab/library/continuous-close-guide