What AI-Native ERP Actually Means: A CFO/CTO Evaluation Framework

The vendor demo I watched in April had the phrase “AI-powered ERP” in the title, the footer, and the first three slides. By slide four it was clear the AI in question was a natural-language query bar on top of a 2015 database architecture. The bar could accept “show me AP aging over 60 days” and return a filtered table. That is not AI-native. That is a search box with a better interface.

The taxonomy matters because $313M in Series A and Series B funding has landed on three startups in the past 12 months — DualEntry ($100M+, Lightspeed + Khosla), Campfire ($100M, Accel + Ribbit), Rillet ($108.5M, a16z + ICONIQ) — all claiming to be the AI-native alternative to NetSuite’s $4 billion mid-market franchise. The claims are real. The differentiation between them is real. But the vocabulary has gotten loose, and CFOs and CTOs making a platform decision in 2026 need a way to evaluate claims that does not rely on vendor demos.

The two-tier taxonomy

Architecture-native means the model is in the loop at the data layer. Journal entries are categorized, reconciled, and anomalies flagged before a human reviews them. The AI is not responding to queries about data — it is acting on data as it flows in. An architecture-native system can book a journal entry, match a bank transaction, flag a duplicate, and surface a variance without a human copying anything. The model’s access is structural, not conversational.

AI-bolted-on means a model layer sits on top of a conventional relational database. Users interact through natural language. The underlying data architecture is unchanged; the model is a translation layer between human language and SQL, or a summarization layer on top of existing reports. This is not useless — query interfaces save real time — but it does not compress the close cycle. The reason is architectural: the model can answer questions about data after a human has processed it, but it cannot process the data itself. Categorization, matching, and reconciliation still run on human time.

A practical test: can the system act on data without a human first copying it to a prompt? If a user must type “show me AP aging over 60 days” and the system queries to answer, the architecture is bolt-on. If the system surfaces AP aging outliers proactively, without a query, the model is in the data layer. This single question cuts through most vendor AI feature marketing.

Most legacy ERP vendors shipping “AI features” in 2025-2026 are in the second category. The question for the three new entrants is whether they have cleared the first bar.

The three entrants applied to the taxonomy

Campfire has the most credible architecture-native claim. Its LAM (Large Accounting Model) is a proprietary model trained exclusively on accounting data, achieving over 95% accuracy on reconciliations and variance analysis. Campfire reports close cycles cut by five times — 144 days reclaimed annually per customer. The benchmark customers (PostHog, Replit, Decagon) are SaaS companies with complex billing stacks, and the data suggests the automation is real, not presentation-layer. The distinguishing architectural fact: LAM acts on records directly rather than in response to queries.

Rillet has the clearest ICP precision: SaaS companies. Native ARR, MRR, and NRR calculated directly from the general ledger — not configured, not imported. Two hundred native integrations to the standard SaaS finance stack (Stripe, Ramp, Brex, Rippling, Salesforce, HubSpot). Aura AI handles flux analysis and accruals automatically. The Sequoia and a16z backing signals conviction about the SaaS-first positioning. Current customers — Windsurf, Bitwarden, Decagon — track exactly to the ICP. If you are a SaaS company with a standard billing and payroll stack, the case is narrow and strong. The narrow ICP is also the honest limitation: Rillet’s differentiation is largely irrelevant for non-SaaS companies, and its coverage outside the SaaS-first use case is less documented than DualEntry’s broader-market claims.

DualEntry is the broadest claim. Multi-entity consolidation, capital management, SaaS metrics, revenue recognition, billing — the platform attempts to cover more surface area than Campfire or Rillet. The Altis diligence report notes it trails those two in SaaS-specific modules and finds the automation depth may be lower than marketed. The 24-hour deployment claim is the signature differentiator — implementing NetSuite takes 6-18 months and $100-500K in consulting fees; DualEntry claims same-day go-live. The honest assessment: strongest for greenfield mid-market multi-entity companies that want the fastest path to any modern ERP, not necessarily the deepest AI in any specific workflow.

What NetSuite is doing

Oracle has been shipping AI features under the NetSuite Next branding: natural language reporting, automated reconciliation suggestions, AI-assisted data entry. These are genuine improvements. NetSuite’s revenue has grown 18% year-over-year and the platform holds roughly 34% of mid-market cloud ERP share. It is not standing still.

The harder question is architecture. NetSuite’s data layer is a 2000s relational database with AI features sitting on top of it. Shifting AI from the query layer to the data layer is a significant architectural migration, not a feature sprint. The entrants built from that starting point; NetSuite would have to migrate toward it without breaking a platform running the finance operations of 40,000+ companies.

Whether NetSuite closes the AI architecture gap before the entrants reach enterprise feature maturity is the timing question every CFO and CTO is implicitly betting on.

Three-question decision heuristic

Before evaluating any vendor, answer these three questions against your last close.

1. What fraction of your finance close time is rote work? Map your last close: how much time was categorization, matching, reconciliation, and variance formatting versus judgment calls, stakeholder communication, and review? If the rote fraction is above 60%, the AI-native automation is a material close-time lever. If it is below 40%, you are buying a cleaner interface for work that is already judgment-heavy. The automation value is real in the first case; in the second it is incremental.

2. Are you a SaaS company with a standard billing stack? If yes, Rillet’s architecture — where ARR and MRR are first-class GL outputs, not configured dashboards — gives you something that does not exist on NetSuite without significant configuration or a BI layer on top. If you are not a SaaS company, the SaaS-specific differentiation is less relevant and the broader feature-and-reliability comparison matters more.

3. What is your multi-entity situation? Single-entity SaaS companies: the conversation is between Campfire and Rillet. Multi-entity mid-market companies with subsidiaries, intercompany elimination, and currency consolidation: DualEntry’s architecture is designed for this, and Campfire’s coverage is less documented at that complexity. NetSuite has the deepest multi-entity track record, and that track record is worth something for complex operating structures.

If your answers point toward one of the entrants, the next step is a structured pilot — not another vendor demo. Take your last month-end close, run it with real transaction data in the target platform, and measure close time against your baseline. Three days of real work tells you more than any demo. If your rote-work fraction is genuinely above 60%, the automation delta will be visible immediately.

The honest framing

None of these platforms replace NetSuite for a $300M+ multi-entity enterprise with 15 years of configuration baked in. What they replace is the mid-market implementation decision — the company currently evaluating NetSuite, or currently on NetSuite for two to three years with a finance team that spends most of its close on rote work.

The $313M in funding is a bet that AI-migration tooling has made switching costs low enough to reopen that decision. Published analysis of AI-assisted ERP migration programs puts the reduction at approximately two times cost and duration — not free switching, but half-price switching. Whether half-price is cheap enough depends on how much pain your finance team has with the current close cycle.

Run a pilot close. Three days against your last month-end, real data, real workflows. That is the evaluation, not the demo.