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.

Silver Lake is in talks to take Workday private for up to $43B (4 minute read)

Reuters reported that Silver Lake is prepared to pay up to $43 billion to take Workday private, sending the stock up about 18% to close at $206.45 and a roughly $51.1 billion market cap, with trading halted several times during the session. Workday serves more than 11,500 customers with human resources, finance, payroll and planning software, and had lost roughly 15% year to date on worries that AI would erode traditional enterprise software. What this means for your stack: if Workday Financial Management or Adaptive Planning is in your stack or on your shortlist, add ownership change to your risk register now, because private equity ownership historically means tighter list price discipline, slower non-revenue roadmap items and harder renewals. Axios makes the counterargument worth holding onto, which is that large systems of record are the least AI-disruptable software you own, and that is exactly why they are being bought. Discuss with ChatGPT →

Second source, with the Dayforce and Electronic Arts comparables, at Axios.

Stripe and Advent bid $60.50 a share for PayPal, valuing it near $53B (3 minute read)

The Wall Street Journal reported that Stripe and private equity firm Advent have offered $60.50 a share for PayPal, valuing it at roughly $53 billion, with talks described as advanced and a deal possible within weeks. PayPal chief executive Enrique Lores, who arrived from HP in March, is midway through a turnaround that includes cutting about 20% of the workforce over two to three years. What this means for your stack: between this and OpenRouter, Stripe is in the market for both ends of the transaction, the consumer rail and the machine that decides what to spend, and if you route merchant payments through PayPal or Braintree it is worth asking your rep what the integration roadmap looks like under new ownership before your next renewal. Discuss with ChatGPT →

Ordway raises $20M to add AI agents to quote-to-cash (3 minute read)

Ordway, a billing and revenue management platform covering subscription management, usage-based billing and revenue recognition, raised $20 million in equity and debt led by Harbert Growth Partners with debt from Western Alliance Bank, and says recurring revenue doubled over two years while staying profitable. The roadmap doubles research and development spending to build agents that update billing and rebuild key performance indicator reports after a contract change, plus cash flow and churn forecasting, on top of already shipped MCP access, cash reconciliation and contract data abstraction. Discuss with ChatGPT →

Stripe is buying OpenRouter for more than $7B to own AI model routing (3 minute read)

Bloomberg reported that Stripe agreed to acquire OpenRouter for more than $7 billion, about five times the $1.3 billion valuation it carried after a $113 million Series B in May led alongside Sequoia, Andreessen Horowitz, Menlo Ventures and CapitalG. OpenRouter sits between an application and more than 400 models and picks which one runs each request based on capability and price, claiming 8 million users. What this means for your stack: the reason a payments company paid $7 billion for a router is that the router is where the meter lives, and PYMNTS makes the point plainly, which is that AI consumption is turning into a spend category that wants per-model limits, ceilings and approval thresholds rather than one line called AI. On the editor's question of whether right-sizing solves compute demand long term, the honest answer is that routing reliably cuts a bill and does not reliably cut aggregate demand, because cheaper inference has so far produced more inference rather than less. Discuss with ChatGPT →

Second source, on why a payments company wants the metering layer, at PYMNTS.

Only 11% of companies can forecast their AI spend within 10% (6 minute read)

Mavvrik's 2026 State of AI Cost Governance report, covered by Chris Middleton, found that just 11% of organizations can forecast AI spending to within plus or minus 10%, down from 15% a year ago, while 62% saw cost surprises force a material business change, 40% escalated to the board, and one in four delayed or cancelled an AI initiative. The report names task complexity, retry loops and orchestration overhead as the real cost drivers, and notes that 98% of organizations run agentic workloads while only 36% include agents in cost reporting. What this means for your stack: AI cost tracking software or processes are important for tracking and controlling spend on AI tools, particularly if you don’t have caps in place or usage varies widely between teams, projects, and time horizons. Discuss with ChatGPT →

One in four executives says they can govern AI at enterprise scale (3 minute read)

A survey of more than 100 C-suite and technology executives in the US and Canada by HFS Research and TCS found that 35% say AI consistently delivers business outcomes while keeping regulators comfortable and staying under control, 25% believe they have adequate governance and controls to scale AI across the enterprise, and about 17% trust autonomous AI for critical business functions. What this means for your stack: the 17% figure is the one to quote in a steering committee, because it says the people closest to these deployments are not yet willing to remove the human from anything that matters, which is a reasonable position and a much easier one to defend than enthusiasm. Discuss with ChatGPT →

Google Sheets can now turn a range into an app from a prompt (3 minute read)

Google launched Sheets Canvas, which uses Gemini to turn spreadsheet data into an interactive mini-app described in plain language from the Ask Gemini side panel, with two-way syncing so edits in either the app or the underlying sheet update the other. It is rolling out globally in English to Google AI Pro and Ultra subscribers and to eligible Workspace Business and Enterprise Standard and Plus plans. What this means for your stack: the interesting part is the two-way sync, because it means the shadow finance app your FP&A analyst builds on Friday is writing back into the same sheet your used to reviewing or working in. Discuss with ChatGPT →

FASB proposes letting qualifying stablecoins sit in cash equivalents (4 minute read)

The FASB, the US body that sets accounting rules, issued a proposed update that would let certain digital assets including stablecoins be classified as cash equivalents if the holder has an on-demand contractual right to redeem them for cash directly with the issuer, and the issuer holds reserves of at least one to one in short-term highly liquid investments. The proposal also requires every entity presenting cash equivalents to disclose the significant components, naming Treasury bills, commercial paper, stablecoins and money market funds, with comments due November 19, 2026. What this means for your stack: the disclosure requirement applies to everyone presenting cash equivalents, not only to companies holding digital assets, so even if you will never touch a stablecoin you will likely be breaking out the components of that line. Check now whether your general ledger tags cash equivalents by instrument type or lumps them into one account, because that is the change that lands on the close calendar. Discuss with ChatGPT →

The SEC moved separately the same week with Regulation Crypto Assets, covered at CFO Dive.

Hillary Salo will chair the FASB from July 2027 (3 minute read)

Hillary H. Salo, currently FASB vice chair and a former KPMG audit partner, will succeed Richard Jones as chair on July 1, 2027 for a seven year term running to June 2034. Salo started her career as a FASB postgraduate technical assistant, rejoined as technical director in 2020, and became vice chair in July 2024. Discuss with ChatGPT →

How to give an AI agent database access without giving it the database (9 minute read)

Pamela Fox lays out four designs for letting an agent reach a PostgreSQL database, from free-form SQL, which she treats as prototype only, through read-only SQL with guardrails, templated queries with safe parameter substitution, and destructive actions gated behind an explicit confirmation dialog. The defence is layered rather than single, combining query parsing that rejects anything other than a SELECT, a database role with minimal permissions, read-only transactions, query timeouts and row limits. What this means for your stack: the pattern transfers to finance systems directly, but read the gap honestly, because this article is about preventing unintended writes and runaway queries and says nothing about who is allowed to see what. Card numbers, bank details and payroll are a classification problem, not a mutation problem, and a read-only role that can still SELECT the full card column is exactly as compliant as no control at all. Ask for column-level permissions or a masked view, scope the database role to the smallest set of tables the agent needs, and treat any MCP server that can reach cardholder data as in scope for PCI DSS rather than as a developer tool. Discuss with ChatGPT →

A Fields Medalist maps where LLM maths ability stops (11 minute read)

Timothy Gowers argues that recent AI mathematical results cluster in a specific place, which is finding counterexamples and building proofs out of standard techniques, because models can carry enormous breadth of known method and brute force through search trees a human would prune. Where they still struggle is the judgement work, meaning knowing which branch is worth following, deciding what to leave unspecified while a proof takes shape, and diagnosing why a failed approach failed, and he offers the 2016 cap-set solution as a benchmark for the kind of conceptual leap that has not happened yet. What this means for your stack: this is the follow-on to last issue's Riemann zeta item and it sharpens the same point. Models are strong where the method is known and the search space is large, which describes reconciliation exception hunting, contract clause extraction and variance driver enumeration, and weak where the answer depends on knowing which of six plausible explanations is the real one. Staff accordingly. Discuss with ChatGPT →

Fintech funding hit $808M across 12 deals this week (3 min) Financial infrastructure took five of twelve deals, with Fazeshift, an accounts receivable automation platform that raised a $22 million Series A in May, taking an undisclosed follow-on from Amex Ventures, and Fisent Technologies raising a $4.3 million Series A for generative AI process automation in regulated enterprises.

Worth a second look

OpenAI cut its cheapest model's input price by 80% OpenAI dropped GPT-5.6 Luna to $0.20 per million input tokens and $1.20 per million output tokens, an 80% reduction, and cut Terra 20% to $2 and $12, while leaving Sol unchanged and adding a Fast mode at 2.5 times the speed for twice the price. OpenAI says Luna matches frontier-class quality from a year ago at a fraction of the cost per task, and cites Notion measuring Terra at GPT-5.5 quality for half the cost per task in 60% less time. What this means for your stack: this is the price ladder that makes routing worth doing, because a ten to one gap between the cheap tier and the expensive one is only a saving if something is deciding which tier each request deserves. Resurfaced here because it is the number behind this issue's OpenRouter story. Discuss with ChatGPT →

Mistral OCR 4.1 adds confidence scores per block, at $4 per 1,000 pages Mistral released OCR 4.1, a document reading model that returns paragraph-level bounding boxes, structural block labels and a confidence score for each block, priced at $4 per 1,000 pages or $5 per 1,000 with structured annotations, with batch processing supported. What this means for your stack: the confidence score per block is the part worth caring about, because it turns document capture from a yes or no into a queue you can threshold, routing anything below your cutoff to a human instead of into the ledger. That is the difference between an extraction tool you can put in front of an auditor and one you cannot. Discuss with ChatGPT →

AI-enabled breaches cost $6M each, and only 37% encrypt data properly IBM's 2026 Cost of a Data Breach study, covering 602 organizations, found a quarter of malicious breaches were AI-enabled, a 56% year over year rise, costing about $6 million each against a global average of $4.99 million, while more than 20% of breaches targeted AI models or applications directly. Only 37% of organizations encrypt sensitive data both at rest and in transit. What this means for your stack: pair this with the MCP piece in The Practitioner, because the encryption figure is the quiet answer to whether an agent can safely reach cardholder data. If the data is not encrypted at rest today, adding a query interface on top of it does not create the exposure, it just makes it easier to reach. Discuss with ChatGPT →

Terms that turned up in this issue, plus two readers asked for.

Fine tuning. Taking a general purpose model and training it further on your own examples so it behaves the way you want on your specific task. You are not teaching it new facts so much as new habits, such as your revenue recognition memo format or your expense policy exception language. Most finance teams do not need it. Fine tuning is worth considering when you need a consistent output shape across thousands of runs and prompting keeps drifting, and it is the wrong tool when what you actually need is for the model to see your current data, which is a retrieval problem. It also creates a version you now own, maintain and revalidate every time the base model changes.

System prompt. The standing instruction sitting above every conversation with a model, invisible to the user, that sets its role, rules and boundaries. It is what makes one assistant answer as a cautious auditor and another as an eager sales engineer using the identical underlying model. Treat a vendor's system prompt as a control document and ask to see it. It is where the instruction not to give tax advice, or the instruction to always cite a source, either exists or does not, and it is also the thing most likely to change silently in a product update. If a vendor will not show it to you, you cannot evidence what the tool was told to do.

Model routing, or an AI gateway. A layer sitting between your application and many models that picks which one handles each request, usually trading quality against price and speed. This is the entire product Stripe reportedly paid more than $7 billion for. The finance analogy is payment routing, where you send each transaction down the cheapest rail that will actually clear it. The saving is real and the control question is the same one you would ask of any routing engine, which is who sets the rules and who can see what they cost.

Token. The unit models are billed in, roughly three quarters of a word. Input tokens are what you send, output tokens are what comes back, and output is usually five to six times the price of input. This is why AI cost lines behave like a utility bill rather than a subscription, and why 43% of the companies in this issue's cost report named tokens as their top source of unexpected spend.

Inference. Running a trained model to get an answer, as opposed to training, which is building the model in the first place. Inference is the recurring operating cost, training is the capital event.

Post-training. Everything done to a model after the main training run, including fine tuning and reinforcement learning, to make it better at particular kinds of work. Increasingly this is where capability gains come from rather than from bigger base models.

Agentic workload. Work where a model runs a multi-step task on its own, calling tools and retrying, rather than answering one question. It is the reason cost forecasting got harder, because a single request can quietly become forty. Retry loops and orchestration overhead are the invisible part of the bill, and 98% of organizations run these workloads while only 36% track them separately.

OCR, or optical character recognition. Reading text out of an image or a scanned document. Modern versions also return where on the page each block sat and how confident they are about it. The confidence score is the finance-relevant part, because it lets you set a threshold and route anything below it to a human rather than into the ledger.

MCP, or Model Context Protocol. The standard way an AI agent connects to a system such as your general ledger, spend platform or database. It defines how the connection works, not what is allowed through it. Which is the whole point of this issue's Practitioner item. The protocol is not a permission model, and whatever the agent's database role can read, the agent can read.