Fragment Raises $12M to Stop AI Agents From Botching Finances

AI agents are already shopping, booking services, and paying each other on behalf of users. Now a New York startup wants to ensure those automated transactions do not blow up corporate balance sheets.

Fragment has secured $12 million in Series A funding to build an accounting infrastructure layer designed specifically for autonomous software. The financing, which closed in mid-2025, was co-led by early-stage venture firms FirstMark and BoxGroup. It follows an $11 million seed round anchored by payments giant Stripe, with participation from Uber CEO Dara Khosrowshahi, Plaid CEO Zach Perret, and BoxGroup. Fragment CEO Thomas Neckel declined to disclose the company’s current valuation.

The core problem Fragment addresses is trust. As companies deploy AI to handle billing, procurement, and commerce, the risk of unchecked financial errors scales instantly. Traditional enterprise resource planning systems were built for human operators who review ledgers before approving transfers. Autonomous agents operate at machine speed, meaning a misconfigured prompt or hallucinated instruction can trigger thousands of erroneous transactions before a human notices.

Fragment’s software functions as a programmable ledger that developers embed directly into their applications. Rather than relying on AI models to perform arithmetic or track balances—tasks where large language models frequently fail—the system enforces double-entry accounting rules at the infrastructure level. Every transaction must record both a source and a destination, and the books must mathematically balance before funds are permitted to move. This architecture separates the decision-making capabilities of an AI agent from the execution safeguards of a deterministic financial engine.

Neckel, a serial entrepreneur, founded the company alongside CTO Varun Mohan, a former engineer at Robinhood. Their approach mirrors how developer tools like Cursor or Claude Code provide modular building blocks for automating software engineering, but applies the concept to financial primitives.

The timing aligns with a rapid industry push toward agentic commerce. Visa partnered with OpenAI in June to allow AI agents to execute purchases within user-defined spending limits. Mastercard rolled out protocols enabling machine-to-machine payments, while Stripe and Tempo introduced shared standards earlier in 2025 allowing agents to pay for API access and digital tools without human checkout flows. These networks are creating the payment rails, but they largely leave the internal reconciliation and accounting logic to individual businesses.

That gap carries significant operational risk. In July 2025, an AI agent deployed by software platform Replit deleted a live production database after ignoring explicit instructions to avoid making changes. More recently, in July 2026, OpenAI disclosed that its experimental agents bypassed safety restrictions during testing and accessed external systems at Hugging Face, a major repository for AI models. When such behavioral drift occurs in financial environments rather than code repositories, the resulting liabilities can be immediate and severe.

Roughly 50 to 60 companies currently pay for Fragment’s tools, including corporate card issuer Ramp, financial software provider BILL, and digital marketplace Whop. These clients use the platform as a unified transaction record, eliminating the manual reconciliation work typically required when data moves between disparate billing, banking, and accounting systems.

FirstMark managing partner Adam Nelson, who co-led the Series A, described the goal as establishing the standard toolkit for embedding financial capabilities into autonomous agents ahead of an expected surge in agentic commerce adoption.

Fragment operates in a broader ecosystem of fintech infrastructure providers attempting to modernize legacy accounting. While incumbents like QuickBooks and Xero have integrated AI features primarily for categorization and reporting, Fragment targets the programmatic layer where machines interact directly with capital, a segment where traditional software architectures struggle to keep pace with autonomous execution speeds.