LEDGER

The finance systems & AI brief for controllers and CFOs

TOGETHER WITH CFOLAYER

Bespoke consulting for finance tech stack strategy, selection, & implementation

Get a personalized roadmap for your finance tech stack. CFOLAYER partners with CFOs and controllers to design, select, and implement the right systems for your scale and business model - eliminating vendor confusion and slow time-to-value.

Ramp launches Router, an AI model gateway wired to its spend platform (3 minute read)

Ramp launched Router, a single application programming interface that sends each AI request to whichever model can handle it most cheaply, with a dashboard showing token spend, cost, latency and fallback attempts per request. It supports OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI and Z.ai, is free to route through the end of 2026 with a $26 launch credit, and Ramp says it cut the company's own model costs 30% and cuts customer inference bills 40% on average. Ramp was valued at $44 billion after its $750 million round in June and says Router already moves more than 2.75 trillion tokens a month. What this means for your stack: this is the same layer Stripe reportedly paid more than $7 billion for two weeks ago, being given away free by a company that already sits on your expense data. If your engineers are calling four model providers on four invoices, the reason to look at this is not the 40% claim, it is that per-request cost lands on the same platform as your card spend, which is the first time AI consumption has a chance of arriving pre-coded rather than as a mystery on a corporate card. Read the retention terms before you route production traffic, since the default is one year with personally identifiable information stripped, and it is United States only for now. Discuss with ChatGPT →

Second source: router.com. The ramp product page at router.com carries the 40% savings claim and the token volume figure. TechCrunch also covered the launch on August 20.

Rillet hits a $1B valuation with a $100M round aimed at NetSuite (3 minute read)

Rillet, an AI-native general ledger built for mid-market finance teams, raised a $100 million Series C led by ICONIQ with Sequoia and Andreessen Horowitz returning, reaching a $1 billion valuation and more than $200 million raised in total. The company says it now serves more than 600 customers, doubled annual recurring revenue in the past three months, and closed the round in under 48 hours. Chief executive Nicolas Kopp frames the pitch as the general ledger becoming the operating system for finance rather than only a system of record. What this means for your stack: the number to weigh is not the valuation, it is 600 customers, because that is the reference-call pool you now have if Rillet is on your shortlist against NetSuite or Sage Intacct. A year ago the honest objection to an AI-native ledger was that nobody your size had run a year-end on one, and that objection is getting weaker every quarter. The EY alliance announced in April is the other signal worth chasing, since an implementation partner with audit-side familiarity is usually the gap that sinks a migration off a legacy ledger. Discuss with ChatGPT →

Airwallex raises $320M and moves into automated bookkeeping (4 minute read)

Airwallex raised a $320 million Series H led by Addition at an $11 billion valuation, up from $8 billion in December, and now serves more than 675,000 businesses at over $1 billion in annualized revenue while remaining EBITDA positive. President Lucy Liu described a new product called T:0 as automated bookkeeping that runs the finance function the way assisted driving runs a car, alongside an agentic consumer wallet called Airi, and said the company intends to be ready to list by year end while noting it is not the best time to go public. What this means for your stack: if you already route cross-border payments through Airwallex, a bookkeeping product from your payments provider is worth a hard look at the boundary rather than the feature list. The value is that the entity that moved the money can code the entry without a bank feed guessing at it. The risk is the same one every embedded finance product carries, which is that switching payment providers later starts to mean switching your ledger too. Discuss with ChatGPT →

Banks are buying insurance against their own AI getting it wrong (4 minute read)

Testudo, a Lloyd's of London backed startup, launched what it describes as the first AI liability insurance product in January, writing policy limits up to $10 million at annual premiums typically between $10,000 and $20,000, and says banks have asked for cover as high as $500 million that it cannot yet write. Testudo underwrites on historical litigation data rather than on any audit of how the model performs, and the article carries genuine dissent, with Patelco Credit Union's chief technology officer saying the credit union has not deployed customer-facing AI at all and an Ocorian executive comparing the marketing wave to Y2K. What this means for your stack: an insurer pricing this risk without inspecting the model is the tell. It means the market currently has no accepted way to measure how likely your AI is to be wrong, so it is pricing your legal exposure instead, which is a proxy for how litigious your customers are rather than how good your controls are. Treat a quote as a data point on what your downside looks like to an underwriter, not as evidence that the deployment is safe, and note the gap between a $10 million limit and the $500 million banks are asking for. Discuss with ChatGPT →

A CFO is building an AI ledger to check his real ledger (4 minute read)

Luke Pritchett, chief financial officer at spend management company PEX, is building what he calls a shadow ledger, a parallel AI-maintained version of the books that runs alongside the official general ledger to catch errors, and he wants it 80% to 90% complete by the end of 2026. He is using Anthropic's Claude under an enterprise license to build agents that connect the corporate card platform, human resources system, bank feeds and accounting system, including an expense policy agent that chases missing receipts and handles lost receipt affidavits without blocking the transaction. Pritchett says individual AI use tends to produce 20% to 30% productivity gains and expects agents working with each other to break past that ceiling. What this means for your stack: a second ledger you reconcile against the first one is an old control idea wearing new clothes, and it is a more defensible way to introduce AI into the close than letting it post. The agent proposes, the difference gets investigated, and the official books stay human-owned, which means you can show an auditor what the AI touched and what it did not. Discuss with ChatGPT →

Finance leaders put guardrails around 'shadow finance' (3 minute read)

CFO Alliance roundtables in Houston, Dallas and Austin found finance leaders building governance around what they are calling shadow finance, meaning analysis and reporting work being done in AI tools outside the systems of record, where adoption has moved faster than oversight or any measurement of return. What this means for your stack: shadow finance is shadow information technology with worse consequences, because the output is a number somebody puts in a board deck rather than an app nobody sanctioned. The cheapest first control is not a policy, it is a rule that any figure reaching an external or board audience has to be traceable to a system of record, which quietly forces the tool question without having to police which tools people open. Discuss with ChatGPT →

Alice raises $140M to stress-test other companies' AI models (3 minute read)

Alice raised a $140 million round led by Apax Digital at a valuation between $700 million and $800 million, bringing total funding to $280 million, and is approaching $100 million in annual recurring revenue. The company tests models for weaknesses during training and deployment, simulates malicious prompts and jailbreak attempts, monitors inputs and outputs in real time and sets safety guardrails, and says it protects eight of the top ten foundation model labs, with Anthropic, Google, Nvidia and Cohere named as customers. What this means for your stack: the reason this belongs in a finance brief is that a business testing AI safety reached roughly $100 million in recurring revenue, which tells you the labs themselves do not consider their own models self-evidently safe. When a vendor tells you their AI has guardrails, the follow-up question is who tested them and against what, and whether you can see the result. Discuss with ChatGPT →

Second source: Bloomberg. Bloomberg broke the story on August 25 and is paywalled. Calcalist carries the valuation range and full investor list.

Google ships finance-specific AI agents with citations and audit trails (3 minute read)

Google Cloud launched Gemini Enterprise for Financial Services in preview, led by a managed Financial Research Agent that returns confidence scores, source citations and audit trails with every answer, plus more than 50 finance-specific skills covering credit risk, portfolio monitoring and know-your-customer work and 13 data connectors including LSEG, S&P Global, Moody's, FactSet and SEC Edgar. CME Group and Deutsche Bank are named early users, it runs inside both Google Workspace and Microsoft 365, and Accenture, Deloitte, KPMG and PwC are named implementation partners. What this means for your stack: this one is aimed at capital markets and corporate banking rather than the back office, so the feature list will not help your close. The part to steal is the packaging, because a confidence score, a citation and an audit trail attached to every answer is exactly the evidence pack you should be demanding from the finance AI vendors who are selling to you, and it is now a shipped product rather than a nice idea. Discuss with ChatGPT →

New audit standard adds a fraud lens and asks about your whistleblower line (3 minute read)

The AICPA Auditing Standards Board approved SAS No. 151, which requires auditors to apply what it calls a fraud lens during risk assessment, to understand any whistleblower program the company operates, and to document and communicate more when fraud is identified or suspected, while keeping the standing presumption that revenue recognition carries fraud risk. It takes effect for audits of periods ending on or after December 15, 2028, with early adoption permitted. What this means for your stack: the whistleblower requirement is the one that lands on you rather than on your auditor, because they now have to understand a program you may run informally or not document at all. Two years is enough time to fix that cheaply, and the work is unglamorous, meaning write down how a report gets made, who sees it, how it is tracked and what happened to the last few. If that record does not exist in a system today, the audit is when its absence becomes visible. Discuss with ChatGPT →

CFOs are being told to document exactly how AI touched the numbers (3 minute read)

Ian Schnoor, executive director of the Financial Modeling Institute, argues that finance leaders need policies that answer one question in writing, which is exactly how AI was used in a given process, and says most companies still have no strong disclosures or guidelines around it. He also makes the case that AI raises rather than lowers the modeling skill required, because someone has to keep human judgment on assumptions like inflation and interest rates and run their own checks and stress tests. What this means for your stack: the practical version of this is a column, not a policy document. Add a field to your close checklist and your model documentation that records whether AI was used in a step and at what level, meaning drafted, checked or fully automated, and who reviewed the output. It costs nothing this quarter and it is the artifact you will be asked for when your auditor or your audit committee gets around to the question. Discuss with ChatGPT →

AICPA updates its digital assets practice aid on stablecoins (2 min) The AICPA published updated guidance in its digital assets practice aid covering stablecoin accounting, mining revenue recognition and how current auditing standards apply, landing in the same window as the FASB proposal on whether qualifying stablecoins can sit in cash equivalents.

Fintech funding hit $361M across 16 deals this week (3 min) Payments and AI led the week's fintech funding, with Ingenico taking €150 million from a PIMCO-led consortium, Rillet's $100 million Series C, Rezolv raising $12.5 million for AI-native lending and Natural securing a $100 million credit facility for AI agent payments infrastructure.

Worth a second look

Ramp gives finance teams one dashboard for AI spend Ramp built the dashboard first and the router second. Read together, the strategy is obvious: own the meter, then own the thing being metered. Resurfaced because Router, this issue's lead, only makes sense next to it. Discuss with ChatGPT →

Visa and Airwallex target 42-day freight payment cycles One month before the $320 million round, the same company was pushing into freight working capital. The bookkeeping product is not a pivot, it is the next square. Resurfaced alongside this issue's $320 million Airwallex round. Discuss with ChatGPT →

PCAOB opens comment on its 2026 to 2030 strategic plan Comments close September 4. The technology and data oversight goal is the one that decides what your external auditors will ask you to produce about AI-assisted work, and it is still open for comment. Resurfaced because the comment window closes nine days after this issue lands. Discuss with ChatGPT →

The case that the money is in the application layer, not the model (16 minute read)

Dave Yuan of Tidemark argues that foundation models will not build the unglamorous vertical software that actually gets work done, because vertical markets lack pre-training data, have unclear success metrics and require workflow change customers resist, and he points at the roughly $10 billion the labs have put into services businesses as evidence they know it. The piece assembles the supporting data carefully, citing MIT's Project NANDA finding that 95% of organizations got no return from $30 billion to $40 billion of enterprise generative AI spending, and an Anthropic case study where accuracy went from 21% with no hand-built skills to 95% with them, then drifted back to 65% within a month once the engineering stopped. What this means for your stack: the 95% to 65% drift is the most useful number in this issue for anyone running a finance AI pilot. It says an AI workflow is not a project that finishes, it is a control that degrades, and the degradation is silent because the output still looks like an answer. Whatever you deploy, budget for someone to re-test it on a schedule and decide now what accuracy level makes you turn it off. Discuss with ChatGPT →

Terms that turned up in this issue.

Temperature. A dial on a model that controls how much randomness goes into choosing each next word. Near zero it returns its most likely answer nearly every time, and higher settings let it wander, which is what people are describing when they call a model creative. For finance work you want it low, and you want to know what your vendor set it to. The same question asked twice should return the same answer, so if your AI tool gives a different figure on Tuesday than it did on Monday, temperature is the first thing to ask about. It is also one of the easiest things for a vendor to change quietly in a product update.

Model card. A short standard document published alongside a model stating what it was built to do, what it was tested on, how it scored, and where it is known to fail. This is the closest thing the industry has to a datasheet, and it belongs in your procurement checklist. If a vendor embedding a model in a finance product cannot produce the card, they either do not know which model version they are running or they would rather not say, and both answers tell you something. The limitations section is the part to read, because it is where the honest ones write down what not to use it for.

Hallucination. When a model produces something fluent, specific and wrong, and delivers it with exactly the same confidence as everything else. It is a property of how these systems generate text rather than a defect awaiting a patch. This is the subject of this issue's insurance story. The working control is not better prompting, it is requiring the model to show its source, because a claim you can trace is a claim you can check and a claim you cannot trace has no business in a workpaper.

Guardrails. Rules wrapped around a model that limit what it can do or say, sitting outside the model rather than inside it. In finance tools this covers refusing certain requests, blocking access to certain data, or requiring human approval before an action executes. Ask two questions of any vendor guardrail: what happens when it fires, and who is able to turn it off. A guardrail that logs and continues is a monitoring tool, and a guardrail any end user can disable is a suggestion.

Red teaming. Deliberately attacking your own AI system to find where it breaks, using hostile prompts, edge cases and jailbreak attempts, before someone outside does it for you. It is penetration testing pointed at a model. This is the business Alice built to roughly $100 million in recurring revenue, which is the number in this issue. For your purposes it is a line on the vendor questionnaire: has this been red teamed, by whom, and may we see the findings?

Fallback. What a routing layer does when the model it picked fails, times out or refuses: it quietly retries the request on a different one. Ramp's Router counts these per request on its dashboard. Worth tracking for two reasons. Fallbacks cost you twice for one answer, and they change which model actually produced the output. If you ever need to evidence a run, which model answered is not recoverable from the request log unless fallbacks were recorded.

Deterministic workflow. A process that produces the same result every time from the same input, meaning ordinary software rather than a model. Most of the work inside a finance system is deterministic and should stay that way. The useful frame from this issue's Tidemark piece is that a working system is AI reasoning plus deterministic automation plus a human handling exceptions. The expensive failures happen when someone lets a model do a job that arithmetic already did correctly.

System of action. Software that decides when to act and then acts, as distinct from a system of record that waits to be told. The difference is who initiates. Your general ledger is a system of record. The pitch underneath Rillet's round and Airwallex's bookkeeping product is that it becomes a system of action. Before buying that, write down which entries you are willing to have posted without anyone asking you first, because that list is the real scope of the change.