F2

Agents carry the core work, spreading financial statements, synthesizing data rooms, drafting IC memos and models, and flagging risk across a loan book. Adam plans each deal as a set of tasks and dispatches sub agents across frontier models. A separate Agent answers questions across all of a firm's deals at once, drawing on market data, portfolio insight and uploaded files. The Excel engine underneath is deterministic, but it builds from what the agents extract and analyze. Remove the models and the workbooks have nothing to fill them.

Work runs in three modes. Fast Mode handles quick lookups and checks, Deep Analysis produces audit ready work, and Plan Mode, when a user asks for it, lays out assumptions, data sources, gaps and sequence for review before anything runs. Adam is sold as an autonomous director that delegates to sub agents, and portfolio monitoring flags missed budgets, shifting comps and weakening macro conditions on its own. At Decathlon, committee approval still gates the move into diligence and F2 summaries feed the IC materials, but plan review is optional rather than a required step.

Audit Mode writes every calculation into a live Excel Databook, with inline citations and precedents and dependents traceable to the source file, sheet and cell, and cited passages are highlighted in the source PDF. The spreadsheet engine is deterministic, so arithmetic runs in the spreadsheet rather than inside the model, and F2 reports a score of 95.25 percent on SpreadsheetBench Verified.

Beyond that benchmark, F2 claims Adam delivers full accuracy every time and that Audit Mode eliminates manual verification, with no error rate for its credit outputs. Its figures of 3.7 times faster, 3.3 times cheaper and 5.1 times fewer tokens name no baseline.

Golub Capital, with more than $90 billion of capital under management, invested $5 million in F2 and will deploy its technology to augment Golub's own investment processes. David Golub, one of its two chief executives, is quoted on why. Decathlon Capital Partners is the subject of a dated case study covering pre screening, data room review and checklist tracking. Five users uploaded more than 2,753 files and produced 128 reports in four months.

Certain diligence tasks fell by roughly 25 to 50 percent according to internal feedback, a range F2's summary of the study rounds up to a flat 50 percent. Alex ter Wee, a managing director at Live Oak Bank, is quoted on efficiency, and Decathlon was chosen as a design partner for Institutional Knowledge alongside firms such as Carlyle. In June 2026 F2 counted more than 100 funds and banks as users, monthly active users up 650 percent for the year and more than 15,000 deals analyzed.

Every data room document, diligence file and IC output falls under what F2 calls its Zero Data Retention Agreement, and the language models keep nothing. Client data never trains a model, and data is encrypted in transit and at rest. Customers own their F2 workspace outright, and Golub's stated reason for partnering is that its knowledge stays under its own control as models change. The privacy notice keeps service data after a relationship ends for F2's own analysis. F2 does not explain how that fits with zero retention, or how one lender's precedent base is kept apart from another's.

The privacy notice, last updated September 26, 2025, is issued by F2 AI, Inc. in New York, appoints a data protection officer and relies on standard contractual clauses for transfers from Europe and the UK. It names key subprocessors, and data is processed in the United States. Retention is open ended. Service data is kept for the life of the relationship and a period afterward for F2's own analysis and archiving, with deletion on verified request.

F2 states SOC 2 Type II and GDPR compliance and keeps its security documentation in a trust center at security.f2.ai. The report period, auditor and scope are not available outside that trust center. Data is encrypted in transit and at rest, and every model interaction falls under F2's zero retention agreement.

F2 holds no license and maps the platform to no supervisory expectation. Its commercial bank users run credit work under bank model risk and fair lending supervision, and its private credit users answer to their investors and, where they are registered advisers, to the SEC. F2 does not say how its output is documented for a bank examiner or a fund's compliance file.

Business borrowers are the subjects of these credit decisions, and Regulation B's fair lending rules reach business credit as well as consumer loans. Institutional Knowledge grounds each new deal in a firm's own history by design, which can carry past patterns of who got funded into new screening. F2 gives no account of how its screening or precedent matching is checked for skewed outcomes.

F2's public terms govern its website only. They cap liability at $50 and say nothing there should be relied on for an investment decision. Platform terms are private, so whether F2 stands behind a figure that reaches an IC memo is not known.

The privacy notice names OpenAI, Anthropic, Google Gemini, Google Cloud Platform and Auth0 as key subprocessors. In June 2026 F2 named the models behind Adam as Opus 4.8, ChatGPT 5.5 and Gemini 3.5 Flash, alongside its own tools native to Microsoft Office. It calls the platform model agnostic, able to plug in the model a client chooses. Which model handles which task is not given. The zero retention promise binds the underlying language models.

Work lands in Excel workbooks with formulas intact, IC memos, CIMs, quarterly reports and decks that teams export and edit, and a memo can be regenerated when the model behind it changes. Inputs arrive mainly as data room files, PDFs, images and Outlook emails, plus the firm's historical documents, and adding a company name and website pulls in market context. Google's financial services customers can reach the full platform through Gemini Enterprise. No named CRM, data room provider or market data connector is listed.

Enterprise Rollout comes with a dedicated implementation team, and F2 says the platform is secured in the customer's existing environment. A forward deployed team designs workflows on the platform from the first day, and Rapid Rollout is a lighter version that runs without integrations. Data is processed in the United States on Google Cloud Platform. Single tenant hosting, customer cloud deployment and regional choices are not spelled out.

F2 sells by demo and contract, with no list price. It contrasts its unlimited usage with token based tools without saying what unlimited usage costs, and the Decathlon case study mentions a signed annual contract. Enterprise Rollout comes with a dedicated implementation team, and a lighter Rapid Rollout starts with no integrations, neither with a price attached.

Private credit funds, commercial banks, private equity firms and specialty finance lenders are the users. Live Oak Bank and Western Alliance Bank are on the banking side, and Bain Capital, Golub Capital, Siguler Guff and Monarch Alternative Capital among the investors. F2 also pitches private equity deal teams on M&A diligence and buyout modeling. The work is corporate credit, from screening and underwriting company loans to monitoring them after close. Consumer lending, insurance and public markets research sit outside it.