AI-native ERP startups raised more than $350M in 2025. Campfire pulled $100M across a Series A and B in four months. Rillet raised roughly $95M in disclosed rounds inside a year. DualEntry came out of stealth with a $90M Series A. Numeric added $51M; Puzzle added $30M. The category is real, the capital is concentrated, and the founders are not bluffing.
And yet, across a set of interviews with European mid-market heads of finance this year, the same answer kept recurring: they are not switching their ERP. Not "haven't yet" — staying put. One CFO at a fast-growing Finnish software company on Business Central put it in a single line: "We won't switch ERP, but we'd buy the layer."
Both things are true at once. That is not a contradiction so much as the shape of a market splitting in two. This piece maps the split, the evidence on each side, and who each motion appears to fit.
Disclosure: I'm building Artifi, an AI tooling layer for finance teams, so I have a stake in this market. What follows is an attempt to map it straight, including the parts that complicate my own position.
The funding is real
Campfire, founded in 2024, raised a $35M Series A led by Accel in June 2025, then a $65M Series B in October — $100M in a single year (TechCrunch). Its customers include Replit, Decagon, Trust & Will, Coder, and Flex, and within nine months of going to market it had companies with 100+ employees moving off NetSuite (Campfire).
Rillet shows the same shape: a $25M Series A led by Sequoia in May 2025, then a $70M Series B co-led by a16z and ICONIQ ten weeks later — roughly $95M in disclosed rounds inside a year, on top of an earlier seed (Rillet). The pitch — "built by accountants, replacing 20th-century accounting software" — is landing with nearly 200 customers (GlobeNewswire).
DualEntry launched from stealth in October 2025 with $90M from Lightspeed, Khosla, GV, and Contrary, publishing aggressive win-rate claims against both incumbents and rivals (DualEntry). Numeric, focused on close automation, raised a $51M Series B led by IVP (PR Newswire); Puzzle added $30M for AI-native accounting aimed at startups (Puzzle).
That is more than $350M of fresh capital across named 2025 rounds — over $400M cumulatively across these companies. The thesis is consistent across the decks: legacy GLs were built for a world where humans typed journal entries, AI changes the unit economics of a finance team, and the right way to capture that is a new system of record, not a feature on an old one. The data below suggests that thesis fits a specific buyer well, and much of the mid-market less well.
Who is switching
The named logos point to a clear profile. Replit, Decagon, Windsurf, Postscript, Slash, Coder, Flex, Trust & Will — tech-native companies, mostly post-Series B, mostly software, mostly single-entity or simple multi-entity, mostly on QuickBooks or early NetSuite, usually with a one-to-three-person finance team carrying a heavy manual close. DualEntry's stated range runs from "$5M revenue to NYSE-listed," but the customer it highlights is Slash, a $100M-ARR fintech running on a finance team of one (DualEntry).
That is the wedge: greenfield or near-greenfield, high-trust software companies where the alternative to a young vendor is hiring more senior accountants. The replacement is often less "rip out NetSuite" than "intercept the QuickBooks-to-NetSuite upgrade before a six-month implementation begins." The buyer profile is narrow and consistent — under ~500 employees, simple operational footprint, little localized statutory reporting, US-headquartered, comfortable taking platform risk. For that buyer, the funding data shows AI-native ERPs already winning.
Who is keeping their GL, and why
The rest of the mid-market looks different. Microsoft Dynamics 365 Business Central crossed 50,000 cloud customers in late 2025, up from 30,000 two years earlier, and now exceeds NetSuite's customer count; mid-sized firms make up just over half of that base, and the on-prem BC/NAV install base pushes the total past 100,000 companies (MSDynamicsWorld).
The switching math explains the inertia. The Panorama Consulting 2025 ERP report puts average implementation cost around $450,000 (3–6% of annual revenue), with frequent budget overruns and, in complex sectors like discrete manufacturing, high rates of operational disruption (Panorama). Most mid-market CFOs have either lived through a project like that or watched one end a career.
A €60M Finnish industrial group I spoke with this year illustrates the operational reality. It runs multiple legal entities, three currencies, a chart of accounts customized over twelve years, a Finnish payroll-bureau integration, an Estonian e-invoicing pipeline, and a depreciation schedule the founder personally signed off on. The pain is real — the accountants spend about ten days a month on what is essentially data movement. The CFO's read was blunt: "Even if it works perfectly, the cost to switch is bigger than the savings for the next five years." For a company with 8–12 years of customizations, statutory reporting tied to local tax authorities, and an undocumented web of integrations, replacing the GL is an operational decision, not a software one — and on current numbers the expected value of that decision is often negative.
Microsoft is building for exactly this group: the Business Central 2025 release waves ship Copilot features — AP agents, sales-order agents, plain-language analytics, AI bank reconciliation — designed to keep customers on the system of record and add AI on top (Microsoft).
The "layer" pattern — an on-ramp, not a ceiling
The "add a layer instead of replacing the system" motion is not new. Brex became the first corporate card with a "Built for NetSuite" certification and shipped connectors to QuickBooks, Sage Intacct, and Xero; Ramp went further with bi-directional sync into QuickBooks, NetSuite, Sage, and Oracle Fusion (House Blend). Numeric runs close management on top of any GL; Tabs does it in AR; Rippling layered onto payroll before it became a platform.
The important detail for anyone reading "buyers only want a narrow tool" as a ceiling: in these cases the narrow tool was the entry point, not the endpoint. Brex and Ramp began with a single workflow — corporate spend — and expanded outward into reconciliation, bill pay, accounting sync, and treasury. The pattern is land-and-expand: a specific, low-risk workflow earns the integration and the trust that later make a broader footprint viable. Read that way, "we'd buy the layer" is less a rejection of comprehensive systems than a statement about sequencing — start with one workflow, widen from there. (The buyer-side evidence for that preference is in my notes on 100+ CFO conversations.)
What this means for each group
For vendors: in the available data the two motions correlate with different buyers — replacement with greenfield, tech-native companies; the layer with the installed base. As part of the outreach described in those conversations, I also A/B-tested the two framings on European mid-market CFOs; the "keep your GL, add the layer" framing drew roughly twice the reply rate of "replace your ERP." It's a small sample, and a reply rate is not a purchase, but the direction was consistent.
For finance teams: the relevant question the data raises is less "AI-native or legacy" than "where does the recurring cost actually sit." For the companies keeping their GL, the cost concentrated in data movement around it — reconciliation, AP coding, consolidation, statutory reporting — none of which strictly requires switching the system of record.
For investors: the AI-native ERP category appears real but concentrated — upper-SMB and lower-mid-market in tech-native geographies, rather than the full ERP market. The layer motion addresses the larger installed base at lower switching cost. Both can be true; conflating them is where the analysis tends to go wrong.
The open question
The unresolved part is whether the two motions stay separate. The land-and-expand pattern cuts both ways: layer entrants may widen "up" toward full systems, while AI-native ERP players may push "out" into the installed base. If the on-ramp leads to the same destination from two directions, today's clean split is a phase, not an equilibrium. The capital is flowing toward replacement; the near-term demand in these interviews flowed toward the layer. Watching which one expands into the other is the more interesting question than picking a winner now.