Finance teams have stopped asking whether AI belongs in the back office. A Gartner survey found 58% of finance functions already using AI in 2024, with intelligent process automation and anomaly detection as the leading use cases. Yet when Gartner polled 183 CFOs and senior finance leaders again in 2025, adoption had crept up a single point, to 59% — and 91% of respondents said their initial AI efforts delivered only low or moderate impact.
That plateau is not a technology story. It is a placement story. AI pays off quickly in a handful of transaction-level workflows where the unit economics are easy to measure — and it disappoints almost everywhere it is sold as blanket transformation. Knowing the difference is the whole game.
Follow the cost of a single invoice
The strongest business case in finance operations lives in accounts payable, because every improvement multiplies by invoice volume. Ardent Partners’ 2024 State of ePayables study puts the average cost of processing one invoice at $9.40. Best-in-class AP operations handle the same invoice for $2.78 — roughly 70% below that average.
Speed follows the same curve. The industry average is 9.2 days to process a single invoice; best-in-class teams do it in 3.1 days — about two-thirds faster — according to the same Ardent Partners research. Gaps that size change payment timing, early-payment discount capture, and supplier relationships — not just headcount math.
The Billentis e-invoicing report reaches a similar conclusion from a different direction: switching to electronic, automated invoicing lets companies cut invoice processing costs by 60–80%, with ROI arriving in six to 18 months. That is a payback window a CFO can underwrite without a leap of faith. When clients ask us where to start, this is why the answer is almost always the invoice pipeline — capture, coding, matching — rather than something more glamorous.
Why so many pilots stall between planning and production
The adoption numbers hide a sharp split between experimenting and operating. In a Deloitte and IMA survey of more than 900 finance and accounting professionals, only 9% said they were actually using generative AI, another 8% were in an early phase of adoption, and 60% put adoption at least two years away.
In our experience the stall point is rarely the model. It is the operational last mile: processes that were never redesigned for automation, vendor master data too inconsistent to match against, and no clear owner for the exceptions the system kicks out. Buying software is a procurement exercise; getting invoices to flow through it untouched is an operations discipline. That last mile — not tool selection — is where most of the 91% reporting modest impact are stuck.
The exception queue is a permanent job, not a phase
Here is the number vendors rarely lead with: even best-in-class AP organizations still run a 9% invoice exception rate, against 22% for typical teams, per Ardent Partners — and only 32.6% of B2B invoices move via straight-through processing today. The same research found 31% of AP teams using some form of AI — a share the study projected at the time would reach 45% by the end of 2024. Adoption is real; full autonomy is not.
The practical conclusion: human-in-the-loop review is not a transitional cost you automate away later. It is a permanent operating layer. The model that works pairs AI capture and matching with trained reviewers who:
- own the exception queue and work it against defined SLAs;
- hold document-level output to a 99%+ accuracy target behind a dedicated QA layer;
- feed every correction back into rules and templates so the exception rate keeps falling.
Seen that way, the people in the loop are the reliability layer of the system — not the cost the system exists to remove.
Where the promises outrun the evidence
Two areas deserve open skepticism. The first is the month-end close. APQC benchmarking across roughly 2,300 organizations shows top-quartile performers wrap up the monthly close in 4.8 days or less, the median organization needs 6.4 calendar days, and bottom-quartile performers take 10 or more — figures that have barely moved despite years of automation spend. The close is a coordination and judgment problem, not a keystroke problem. AI earns its keep in the transaction layer that feeds the close — matching, coding, reconciliation prep — but a vendor promising to automate the close itself is overpromising.
The second is generative AI anywhere near the ledger. KPMG’s 2024 global report on AI in financial reporting found that 21% of companies using AI in financial reporting cite hallucinations as a significant concern, and the concern climbs as organizations move from general AI into generative applications. The design rule is simple: generative models may draft, suggest coding, and flag anomalies; deterministic checks and a human sign-off decide what posts. Every AI output in a financial workflow needs a traceable source, because your auditors will ask for one.
Measure documents, not fields
Vendor accuracy claims are the last trap. The benchmark most often cited as the ceiling for acceptable manual data entry is a 1% error rate, notes Conexiom — roughly what unaided human keying produces. Benchmarks compiled by DigiParser put manual entry errors at 1–4% of fields even for trained staff, spiking to 18–40% under time pressure; for AP invoices specifically, manual error rates of 3.5% compare with 0.05% for automated capture, and each error costs $50–$150 once investigation and correction are counted.
Automated capture wins that comparison decisively — but read the fine print. Accuracy is usually quoted per field, and an invoice carries many fields: a system that is 97% accurate per field still corrupts roughly one in four invoice records once you are capturing ten fields per invoice. Vendor benchmarks are also measured on clean test sets, not your production mix of scanned PDFs and supplier quirks. The metrics that matter are document-level: the share of real invoices that flow through untouched, the share that hits the exception queue, and how quickly exceptions clear.
Key takeaways
- The clearest AI ROI in finance is per-invoice: $9.40 average processing cost versus $2.78 best-in-class, and 9.2 days versus 3.1, per Ardent Partners.
- Adoption has plateaued near 59% (Gartner) because the bottleneck is process redesign, data quality, and exception ownership — not tool selection.
- Even best-in-class AP runs a 9% exception rate; human-in-the-loop review is a permanent operating layer, not a phase.
- Keep generative AI out of autonomous posting: draft and flag with AI, validate deterministically, and require human sign-off before anything reaches the ledger.
- Judge systems by document-level touchless and exception rates on your production invoice mix, not per-field accuracy claims.
The pattern is consistent: AI pays off where the work is high-volume, rule-adjacent, and measurable — and it pays off fastest when one team owns the process end to end, exceptions included. That pairing of automation with accountable human review is exactly how we build our AI automation in finance engagements at Apex. If you want to see what the per-invoice math looks like on your own volumes, we will run it with you — no platform pitch, just the numbers.
Sources & further reading
- Ardent Partners, The State of ePayables 2024 — Bottomline
- Billentis Report 2024 — SupplyOn
- Gartner: 58% of Finance Functions Use AI in 2024 — FutureCFO
- CFOs’ AI adoption slows as challenges mount: Gartner — CFO Dive
- Only 9% of finance leaders are using generative AI tools (Deloitte/IMA survey) — CFO.com
- Metric of the Month: Cycle Time for Monthly Close (APQC) — CFO.com
- AI Hallucinations in Accounting and Audit (KPMG 2024 report) — Trullion
- What’s a Good Data Entry Error Rate? — Conexiom
- Manual Data Entry Error Rate statistics — DigiParser