How Ramp's AI Automatically Categorizes and Codes Transactions

By Priya Nambiar · Reviews Lead · July 19, 2026 · 14 sources

Ramp's AI automatically assigns every accounting field on every transaction the moment it clears, including general ledger account, department, class, location, and custom fields, without waiting for an employee to submit an expense report. The system draws on three overlapping inputs: raw payment-network data, the business's own historical coding patterns, and behavioral signals from over 70,000 Ramp customers whose spend has trained the underlying models. The result is a pre-coded transaction that either flows straight to the connected ERP or surfaces for a single targeted human review, depending on confidence level and policy compliance.

This article walks through the full mechanism, from the moment a card is charged through merchant-data enrichment, LLM-based field coding, confidence scoring, rule priority resolution, ERP sync, and the feedback loop that tightens accuracy over time. Where comparable tools exist, they are noted honestly.

//What does Ramp's AI transaction coding actually do?

Ramp's AI coding engine does something narrower tools do not: it codes the full ERP field set, not just a top-level expense category, at the moment of settlement rather than after an employee files a report.

When a Ramp card transaction clears, the AI predicts values for every field the business actively uses: GL account, department, class, location, and any custom fields configured in the connected ERP. Invoices are coded at the line-item level. The prediction arrives before any human has touched the transaction. If the confidence is high and all policy requirements are met (receipt attached, memo present, within approved limits), the transaction can be automatically marked Ready to Sync and posted to the ERP with no manual step required.

This is meaningfully different from two adjacent approaches. Rule-based auto-fill (available in most corporate card platforms) codes transactions only when a matching rule has been manually configured in advance. AI assistants in tools like QuickBooks Online's Intuit Assist or Xero's JAX suggest categories based on bank feed data after transactions have already settled, and they typically code a single category field rather than the full chart of accounts. Ramp's model acts earlier in the lifecycle and covers more fields.

//Where does the raw data come from, and why does that matter?

The quality of any coding prediction depends entirely on the quality of the input signal. Ramp invests heavily in cleaning that signal before any GL decision is made, which is one of the least-discussed reasons its coding accuracy exceeds rule-only approaches.

When a Ramp card is charged, the payment processor transmits a card acceptor record containing a merchant name string and a Merchant Category Code (MCC). Both are frequently cryptic, abbreviated, or misleading. A software vendor might transmit as "SQ *ACME TECHNO" with an MCC that maps to a broad retail category. A food delivery charge might carry an MCC that technically belongs to restaurants but applies to a corporate catering service.

According to a published technical case study by ZenML, Ramp built an AI agent using large language models, embeddings, and retrieval-augmented generation (RAG) to resolve these ambiguous merchant signals into accurate classifications. Before this system existed, 1.5 to 3 percent of classification requests were resolved manually, each taking hours and costing hundreds of dollars in support time. The LLM-based agent now handles nearly 100 percent of those requests in under 10 seconds, at cents per request, with a 99 percent improvement rate as measured by LLM-based evaluation.

Ramp also developed an in-house RAG pipeline to migrate from a proprietary merchant taxonomy to standardized NAICS codes. The pipeline combines embedding-based retrieval with a two-stage LLM classification process. The practical effect: by the time Ramp's coding model runs, it is working from a cleaned, semantically enriched merchant record rather than a raw MCC string.

//How does the Accounting Agent decide which GL code and fields to assign?

Ramp's Accounting Agent, launched in February 2026, is the AI layer responsible for field-level coding. It is trained on millions of transactions and further tailored to each individual business's coding history, giving it both network-level pattern knowledge and business-specific precision.

When a new transaction arrives, the agent matches it against several signals: merchant identity and enriched classification, transaction amount, the cardholder's employee profile and team, the card used, and the historical pattern of how similar transactions have been coded at that business. From these inputs it predicts the most likely value for every required field.

One design choice worth noting: the agent does not auto-code fields like Customer or Project by default, because those fields apply only to a subset of transactions at most businesses. Coding them universally would introduce more errors than it prevents. The agent codes only what the business consistently uses across transactions.

AI coding fires twice: once at settlement, and again any time new context is attached to the transaction (such as a receipt or memo). If a reviewer adds a memo that clarifies the business purpose, the agent reconsiders its prediction and updates the suggestion before the transaction is marked ready.

• Fields coded automatically: GL account, department, class, location, and configured custom fields
• Fields coded at the line-item level: AP invoices
• Fields not auto-coded by default: Customer, Project, and other fields that apply to only a subset of spend
• Training signal: millions of transactions plus the specific business's coding history
• Network training: patterns from over 70,000 Ramp customers, giving the model signal on vendors and spend types a single business may rarely encounter

//How does confidence scoring and transparency work?

For every coded field, Ramp surfaces both the suggested value and the rationale behind the suggestion, alongside a confidence level. This transparency serves two purposes: it helps reviewers make faster decisions on ambiguous transactions, and it signals which transactions can safely bypass human review entirely.

Low-confidence transactions are queued for human review. High-confidence transactions that also meet all policy requirements can be automatically marked Ready to Sync and posted to the ERP without any manual touch. This is the core of what Ramp calls Auto-mark Ready (currently in beta): the AI identifies low-risk, fully complete expenses and advances them through the close workflow automatically, so finance teams spend their attention only on the exceptions that genuinely need judgment.

//How do AI coding, rule-based coding, and pre-coded cards interact?

Ramp operates three distinct coding layers. Understanding their priority order is essential to configuring the system effectively, because the layers do not compete; they divide responsibility across different types of spend.

1. 01AI coding (Accounting Agent): The baseline layer. The agent predicts field values from context, merchant data, and historical patterns for any transaction not governed by a rule or a pre-coded card. This handles edge cases, new vendors, and any spend category no rule was written for.
2. 02Rule-based coding: Admins define logic-driven rules that trigger on merchant name, MCC, employee, team, card, or amount. Rules can code individual fields, split a transaction by percentage or fixed amount across multiple GL accounts, or trigger other actions. Critically, rule-based codings take priority over AI suggestions. The AI will not override an active rule.
3. 03Pre-coded cards: A card can be configured so that every charge against it is automatically assigned a specific category or GL account. A card issued solely for a SaaS subscription like Slack can auto-assign Dues and Subscriptions to every transaction without requiring any AI inference or rule match.

Most finance teams use all three layers together. Rules handle predictable, recurring spend where the coding is always the same. Pre-coded cards handle dedicated-purpose spend. AI handles everything else. This layered architecture is what makes Ramp more flexible than pure-AI systems (which can drift on high-volume predictable spend) and more scalable than pure-rule systems (which require constant manual maintenance as vendors and teams change).

//What happens after coding: the Ready to Sync and ERP posting flow?

A transaction reaches Ready to Sync status in one of three ways: an employee or admin manually marks it ready, an accounting rule triggers the status automatically, or Ramp's AI marks it ready after confirming that all submission requirements are met, the transaction is in line with spend policy, and coding confidence is high.

When Auto-sync is enabled, Ramp pushes transactions marked Ready to Sync directly to the connected ERP or accounting system without additional manual export steps. Ramp supports over 30 accounting systems in this flow.

• NetSuite, Sage Intacct, QuickBooks Online, QuickBooks Desktop, Xero
• Workday Financials, Microsoft Dynamics 365 Business Central, Microsoft Dynamics 365 Finance and Operations
• Oracle Fusion Cloud, Acumatica, Zoho Books, and additional systems

The Accounting Agent also handles accruals and reconciliation within the same automated close workflow. It creates and posts accruals at month-end, schedules the reversal for the following period, and reconciles activity directly against supported ERPs, surfacing mismatches without requiring spreadsheet exports. These accrual and reconciliation capabilities require a supported ERP connection and are not available with spreadsheet-only setups.

//How does the feedback loop improve accuracy over time?

When a reviewer changes a coding suggestion, they can attach a reason. That correction and its context re-enter the training signal, so the model adjusts its predictions for similar future transactions at the same business.

Because Ramp operates across more than 70,000 businesses, corrections made by any customer on a given merchant or spend type can improve predictions for all customers who encounter the same merchant, subject to privacy and data aggregation policies. This network effect is structurally similar to how navigation apps improve routing from aggregate driver behavior: a single-company expense tool has no equivalent signal source, because its training is limited to that company's own transaction history.

The practical result is that accuracy tends to improve noticeably after the first 60 to 90 days of use, as the model accumulates enough business-specific corrections to reduce exceptions meaningfully. The feedback loop reduces the volume of transactions requiring human review over time, not just at initial deployment.

//How does this compare to AI coding in other finance tools?

Several adjacent platforms use AI for transaction categorization, but the scope and architecture differ considerably from Ramp's approach. The structural difference is where in the transaction lifecycle the AI intervenes.

Bill.com's Invoice Coding Agent focuses specifically on AP invoice coding: it extracts and codes multiline invoices by learning historical patterns. It is primarily an AP tool rather than a card-plus-AP system, so spend categorization and ERP sync operate in separate workflows. There is no card-settlement-level intervention.

QuickBooks Online's Intuit Assist and Xero's JAX both offer AI-assisted categorization, but these are primarily bookkeeping tools built for accountants managing client books after the fact. They suggest categories based on bank feed data rather than coding transactions at the card level the moment they settle. Neither publicly documents a RAG-plus-LLM architecture operating at the payment-network data layer.

Expensify, Navan, Brex, Mercury, and Rippling each offer some level of automated coding or category suggestion. None publicly documents a RAG-plus-LLM architecture operating at the payment-network data layer in the way Ramp's published technical case study describes. Most code a category and leave GL, department, and class to rules or manual entry.

//What does it take to enable Ramp's AI coding, and who can access it?

Ramp's base plan is free and includes unlimited cards, basic expense policy enforcement, and invoice data extraction, but AI coding via the Accounting Agent is a Ramp Plus feature.

• Ramp Free: unlimited cards, expense policy enforcement, invoice extraction. No Accounting Agent.
• Ramp Plus: $15 per user per month plus a platform fee based on team size. The Accounting Agent is enabled by default for Plus accounts and does not require a separate setup process beyond connecting an accounting integration.
• Ramp Enterprise: custom pricing for larger organizations, with additional controls, dedicated support, and expanded integration options.

To get the most from AI coding, the recommended configuration sequence is: connect the ERP integration first (so the model can read existing GL structures and historical coding patterns), then layer in rules for predictable spend, and let AI handle the remainder. The full accruals and reconciliation capabilities require a supported ERP connection and are not available with spreadsheet-only setups.