The phrase 'AI-native general ledger' shows up in every Series B pitch and ICONIQ memo on this category. It also shows up in vendor copy that does not always agree on what the term means. This piece is a controller's read on what an AI-native GL actually is at the architecture level, written through the lens of Rillet (the company most associated with the phrase) and contrasted with Campfire's L.A.M. approach.
What is an AI-native general ledger?
An AI-native general ledger is a system of record where the GL data model and runtime were designed from inception around machine learning, with classification, matching, reconciliation, and anomaly detection treated as built-in capabilities of the platform rather than features added by integration. The category emerged in 2024 to 2025 with Rillet and Campfire as the two most-funded vendors, drawing more than $200M in combined venture capital from Sequoia, Andreessen Horowitz, ICONIQ, Accel, Ribbit, and Foundation Capital.
The defining architectural property is that data flows into structured GL records at ingest time, not into intermediate databases later normalized by ETL. Once the data lives in the ledger as structured records, AI can act on it directly. This is the distinction that separates AI-native from bolt-on AI in legacy ERPs.
What does an AI-native GL automate?
Across the category, the automation surface is consistent. The vendor implementations differ in technique but the workflows targeted are largely the same:
- Transaction classification
Inbound transactions from banking, billing, payroll, and expense feeds are coded against the chart of accounts based on vendor history, account patterns, and prior decisions.
- Payment matching
Invoices, payments, and remittances are matched programmatically, including high-cardinality matching against vendor name variations and payment-cycle batching.
- Bank reconciliation
Bank statement lines are matched to GL entries, with discrepancies flagged for review before they reach the close package.
- Revenue recognition
ASC 606 schedules generate from contracts and billing data, supporting subscription, usage-based, tiered, and milestone revenue models.
- Anomaly and variance detection
Spend patterns, vendor behavior, and account-level variances are surfaced continuously rather than at month-end. The reduced surprise factor at close is one of the most consistent customer outcomes reported across the category.
- Flux commentary drafting
AI generates first-draft variance narratives that the controller refines for board and audit reporting.
The architecture: real-time GL vs batch ETL
This is where AI-native GLs make their loudest architectural claim. The traditional pattern in enterprise accounting is batch ETL: data accumulates in source systems, periodic jobs extract and load it into the ERP, and reconciliation happens against the resulting snapshots. AI-native vendors argue that the latency in that pattern is what makes the month-end close painful.
The AI-native alternative is real-time native ingestion. Source systems push transactions into the GL as they occur, the records are structured at write time, and the AI layer operates against the live ledger continuously. The practical effect for finance teams is that close becomes a review-and-approve exercise on classifications the model has already proposed, rather than a from-scratch coding exercise on a fresh batch.
Customer outcomes consistent with this architectural claim: Postscript, with over $100 million in ARR, closes books in three days using Rillet. Windsurf runs its entire finance function with two people (source: Rillet blog, August 6, 2025). The 5-times-faster close figure Campfire publishes from its customer base reflects the same underlying pattern.
Bolt-on AI in legacy ERPs vs native AI: what is actually different?
The framing of 'bolt-on bad, native good' is the venture pitch. The honest engineering view is more interesting.
Legacy full-suite ERPs (Oracle NetSuite, SAP, Microsoft Dynamics 365, Oracle Fusion) have all made substantial AI investments. Oracle launched NetSuite Next with the SuiteAgents framework at SuiteWorld 2025. SAP shipped Joule across the suite. Microsoft Copilot is embedded across Dynamics 365 modules. These are real, substantive capabilities, not marketing veneer.
The architectural difference is at the data model layer, not the AI capability layer. A platform whose schema was designed in 1998 to handle multi-entity GL transactions for the enterprise will have different latency characteristics than a platform whose schema was designed in 2023 to handle structured transaction ingestion for AI agents. Neither is universally superior. They are different bets for different buyers.
The trade-offs:
- Legacy ERPs with AI add-ons offer depth and breadth. 25+ years of vertical features (life sciences, manufacturing, professional services), multi-subsidiary maturity, the largest partner ecosystems, and the most extensive SuiteApps/equivalent marketplaces. For companies whose complexity exceeds what AI-native vendors are built for today, the depth wins.
- AI-native GLs offer speed and lean operating model. Real-time ingestion, faster close, smaller teams. For companies in the $5M to $200M revenue band with simpler entity structures, the speed wins.
Per Rillet's own framing: 'Legacy ERPs are, at their core, dumb databases. They store transactions, but the real work happens in spreadsheets and bolt-on analytics tools.' The argument is sharper than what we would write, but it captures the architectural choice accurately enough to be useful for evaluation.
Where does Rillet sit in the AI-native GL landscape?
Rillet is the platform most explicitly built around the smart-GL architectural model. The technical pattern, in Rillet's own framing: 'native integrations enable structured data to flow into their smart general ledger. AI is then applied' (source: Rillet blog, August 6, 2025).
What this looks like in product terms:
- Aura AI
Rillet's branded AI layer with three access modes: agent-assisted close workflows, conversational queries, and embedded automation across journal entry, reconciliation, and revenue recognition.
- Native integration depth
Direct connections to banking, billing platforms (Tabs, Hyperline, Paygentic for usage-based), AP automation (BILL), expense (Expensify, Float), and payroll. Data is ingested as structured records, not flat exports.
- Single source of truth for GAAP and investor reporting
The ledger is designed to serve both audit-grade GAAP outputs and the metric variants investors expect, without parallel spreadsheet sets.
- Implementation profile
Rillet customers report 4-week implementations against the 12-month baseline for traditional ERP rollouts. The brevity is a function of the native integration approach and lean configuration model.
- Accounting firm partnerships
Strategic relationships with Armanino (top-20 accounting firm) and Wiss (top-50), which extend Rillet's reach into customers whose external accountants already work with the platform.
Where does Campfire's L.A.M. take a different approach?
Campfire and Rillet are both AI-native GLs. They target overlapping ICPs (hypergrowth SaaS, AI-first companies, $5M to $200M revenue). They make different architectural bets at the AI layer.
Rillet's bet is on the integration and data layer: clean, structured data flowing in real time enables AI agents to operate effectively. Aura AI agents work on the data the platform has structured for them.
Campfire's bet is on the foundation model itself. Their proprietary Large Accounting Model (L.A.M.) is trained from scratch on accounting transaction data and fine-tuned per customer on the company's specific chart of accounts and vendor patterns. Campfire publishes a 95% accuracy claim on structured accounting tasks against a roughly 80% baseline for general-purpose LLMs (source: Campfire blog, 'Introducing Accounting Intelligence').
Both approaches are legitimate. The Rillet thesis is that with clean, structured data, off-the-shelf AI is sufficient, and the moat is the integration depth and the GL design. The Campfire thesis is that domain-specific model architecture matters, and the moat is the proprietary L.A.M. The market will likely sort these out over the next 24 months. As of 2026, both are operational platforms with hundreds of customers and roughly $100M each in raised capital.
The CFO question: is AI-native GL audit-defensible?
Yes, when the platform is designed for it. The audit-defensibility question is not really about AI. It is about provenance, attribution, and segregation of duties, all of which AI-native GLs can and do support.
The architectural attributes that determine audit defensibility:
- Attribution on every AI action. Each classification, match, and flag carries source data traceability.
- Logged and reviewable agent decisions. Auditors get the same provenance documentation for AI-generated entries as for human-generated entries.
- Human-in-the-loop defaults for material actions. Agents propose; humans approve. Autonomous workflows are configurable per role and per threshold.
- SOC 1 and SOC 2 Type 1 and Type 2 certification. Rillet and Campfire both hold these certifications.
- Data residency in the certified platform environment. Customer data is not sent to external model providers for inference.
- Granular role-based permissions. Campfire publishes 1,200+ permissions; Rillet has comparable depth.
The honest read: AI-native GLs as a category have invested ahead of where auditors are demanding. Big 4 firms have been working with Rillet and Campfire customers throughout 2025 and 2026 and the audit-trail design has held up. The pre-IPO audit risk associated with AI-native is lower today than it was 18 months ago. It is not zero, and the right disposition is verification on your specific workload, not assumed safety.
.png)
.png)