48% of Legacy ERP Buyers Plan to Replace Within Three Years. AI Is the Reason.

For most of the last two decades the safest thing a CFO could say about an ERP was that it worked and nobody wanted to touch it. Replacing the system of record is expensive, slow and, as one finance leader put it to Battery Ventures, one of the "career-limiting moves" of the job. That inertia has been the incumbents' best defense. New survey data suggests AI has started to wear it down, because finance teams are now measuring their core systems by a question they never used to ask: can this platform keep up with what the models can already do?

Six Months From Pilot to Default

Battery Ventures ran its second survey of finance leaders in August, polling 50 respondents who were mostly CFOs and VPs of finance, just under half at B2B software companies and most at firms of 100 to 499 employees. The headline number is the speed of adoption. 88% said AI was now deployed in their organization. Six months earlier, the figure was 17%.

What sits behind that deployment is broad rather than deep. 84% run two or more models and 28% run four or more, which means finance teams are assembling their own toolkits rather than waiting for a single vendor to deliver one. Only 8% see agents as mission-critical to finance processes so far, and adoption is highest in the simpler workflows: FP&A forecasting, management reporting and accounts payable automation. Model performance is the top buying criterion for 64% of respondents, against 20% for auditability.

The cost of running all of this is arriving as well. Only 16% actively manage token costs today, but 86% expect to within a year. That is a new variable expense line most finance budgets were not built to forecast.

Why the Incumbent Is Now the Risk

The sharper finding is what this is doing to ERP loyalty. In Battery's April survey, 77% of respondents were layering AI onto their existing ERP, and only 15% were buying AI-native systems. By August the split in sentiment between those two groups was stark. Among customers on legacy ERPs, 48% expect to replace their system within three years. Among AI-native customers, 12% do. About 30% of legacy buyers said they would switch for a compelling AI-native alternative, while none of the AI-native buyers said the same.

Confidence in the vendor's roadmap explains part of the gap. 65% of AI-native customers rated their vendor's AI roadmap delivery a 4 or 5 out of 5 in importance, compared with 15% of legacy users. Battery's authors go further, arguing that buyers now see a vendor's tenure as negatively correlated with how modern its technology is, and that vendor stability and track record rank last among buyer priorities.

None of this is new as a complaint. A year ago, Deloitte's Finance Trends 2026 survey found that 63% of finance teams had fully deployed and actively used AI, but only 21% reported clear, measurable ROI, and early-stage adopters named legacy technology (41%) as a major barrier to adoption. Only 14% had fully integrated AI agents. What has changed is that the models themselves stopped being the obvious bottleneck, which leaves the system of record as the most visible one.

The Models Passed the Test. The Ledger Has Not.

In April, Battery found accuracy and data readiness were the two main barriers. Accuracy is the one that has moved. In a study reported by Accounting Today, AI evaluation firm Mercor gave 12 CPAs, averaging about five years of experience, four month-end close tasks with a three-hour window each. Claude Opus 5 made 20 attempts across the four tasks and scored 100% on every one, each in under 10 minutes. The humans generally scored below 50% for accuracy and took 30 minutes at best, more often two or three hours. Accountants working with Claude took about 15 times longer per task than Claude working alone.

The caveats matter. The humans could not ask questions, consult colleagues or draw on client context, and Mercor said the study "ended up testing precisely what the models are best at." On Mercor's harder APEX-Accounting benchmark of 160 tasks across 10 simulated companies, The Decoder reports the best model met 61.8% of grading criteria, and no model fully solved almost 60% of the tasks. Models are now capable enough on bounded close work that the limiting factor is what they are allowed to touch and whether their output can be audited.

That is where the workflow detail in the Battery survey becomes useful. Revenue recognition and tax planning and compliance have the least penetration from general-purpose AI tools, and they are also the most audit-sensitive. CFOs told Battery that accounts receivable, close and consolidation, and accounts payable automation need to be built into core accounting systems that maintain governance and control. In other words, the further a workflow sits from an auditor, the more acceptable a bolt-on model becomes. The closer it sits, the more the core platform has to carry it.

The Playbook: Decide What Your ERP Has to Own

• Map workflows by audit exposure: Sort finance processes from forecasting and management reporting at one end to revenue recognition and tax at the other. The audit-sensitive end is where general-purpose AI tools are least used and where your core platform has to deliver controls.
• Ask for a dated AI roadmap: 65% of AI-native customers rate roadmap delivery as highly important against 15% of legacy users. Before renewal, ask your vendor which finance agents ship when, and what audit trail each one leaves.
• Test the models on your own close: Mercor's results came from simplified tasks. Run a bounded pilot on one reconciliation or accrual with your own data and measure accuracy and review time before drawing conclusions.
• Weigh auditability alongside performance: Only 20% of buyers rank auditability first. For tax, revenue and the close, a model that cannot show its work is a control gap, however accurate it looks in a demo.
• Budget for tokens now: 86% expect to be managing token costs within a year, against 16% today. Put AI consumption on the forecast as a variable cost before it lands as an overrun.
• Separate replace from extend: A full ERP replacement is not the only answer. Where a purpose-built layer for receivables or tax can integrate with the existing ledger and keep the audit trail intact, it can buy time for a cleaner platform decision.