The finance systems & AI brief for controllers and CFOs
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.
🏛️ Systems & Stack
Claude now works inside Google Docs, Sheets and Slides (4 minute read)
Anthropic released a Claude extension inside Google Docs, Sheets and Slides, where it can write formulas, build pivot tables and charts, run Python to clean data and write results back, and build slides from a deck's existing layouts; it is in public beta on all paid Claude plans, with separate connectors that let you create and edit Google files directly from the Claude chat. What this means for your stack: by default every edit shows up as an approval card before it lands, so keep that setting on for any shared workbook that feeds reporting, and on Team and Enterprise plans decide which groups get the add-on before people install it themselves, since access follows existing Google sharing. Discuss with ChatGPT →
Stuut, whose AI agents chase unpaid invoices by email, text and phone, log into customers' payment portals and match incoming cash to open invoices, raised a $52.5 million Series B led by Insight Partners with a16z and Microsoft's M12, bringing total funding to $93 million; it says 81.7 percent of its outbound collections work runs without a person and 95 percent of incoming payments match automatically across more than 150 customers, including Honeywell and ZoomInfo. What this means for your stack: before any agent contacts your customers about money, write down which accounts it may never touch (strategic accounts, anything in dispute or with legal) and make sure every message and cash match lands in a log your auditors can read. Discuss with ChatGPT →
Ramp reportedly raises $1.85 billion at a $60 billion valuation (2 minute read)
Ramp is raising about $1.85 billion at a $60 billion pre-money valuation led by Thrive Capital and Dragoneer, according to Bloomberg, up from $44 billion in June and $32 billion last November; the spend management company serves more than 70,000 businesses and launched in the UK in September. Discuss with ChatGPT →
Avalara ships its tax agent hub and sets EU e-invoicing dates (4 minute read)
Avalara made Aviator, its hub of tax AI agents that map products to tax codes, check exemption certificates and log each action for audit, generally available, and said it is approved for France's e-invoicing mandate and expects to be ready for Germany in January 2027 and Spain in October 2027. What this means for your stack: if you invoice customers in the EU, line your billing system up against these country dates now, because each mandate needs its own approved connection and cross-border B2B e-invoicing across the EU follows from 2030. Discuss with ChatGPT →
Chase puts Gusto-powered payroll inside its business banking (3 minute read)
Chase launched Chase Payroll, powered by Gusto, so small business owners can run payroll, payroll taxes and employee payments inside Chase Business Online next to their banking, with next-day processing; in Chase's own survey, 37 percent of owners said meeting payroll squeezed cash flow in the past six months. What this means for your stack: for a small subsidiary or new entity, payroll in the operating bank simplifies cash timing, but confirm how payroll journal entries reach your ERP (the core accounting system) before you move off a dedicated provider. Discuss with ChatGPT →
🤖 AI in Finance
OpenAI's Decisions API returns typed answers, not paragraphs (3 minute read)
OpenAI opened its Decisions API to all developers in public beta: instead of free-form text it returns a typed answer (how likely a statement is true, a pick from fixed options, or a score on a scale) with probability scores, runs on GPT-6 Luna, and costs $0.10 per million input tokens with no charge for output, which OpenAI says is up to 10 times faster than its Responses API. What this means for your stack: if you are pricing an agent to code transactions or route invoices, redo the cost math at these rates and use the probability score as a cutoff that sends low-confidence calls to a person, but test on your own labeled transactions first, because OpenAI has not published the workload behind the 10x claim. Discuss with ChatGPT →
Meta, Shopify, Stripe and Walmart back a standard for AI agents (4 minute read)
Meta and Sierra, the enterprise AI company co-founded by Bret Taylor, are developing the Personal Agent Protocol, an open standard backed by Shopify, Stripe, Walmart and others that sets how a customer's AI agent identifies itself to a business, signs in, and gets read-only or write access; version 0.1 of the specification is due later in October, and OpenAI and Anthropic are not yet participants. What this means for your stack: the model here (customers grant an agent access, businesses limit what it can do) is the shape your vendor portals and approval rules will need once agents start opening support cases or paying invoices, so watch the spec before building anything custom. Discuss with ChatGPT →
EY's AI services sales jump 49 percent as audit growth lags (4 minute read)
EY's global revenue rose 4.7 percent in local currency to about $57 billion for the year ended June 2026, with sales of AI services up 49 percent, 130,000 assurance professionals now AI-enabled and Copilot licenses up from 203,000 at year end to 381,000, while assurance (audit) was the slowest-growing line at 3 percent. What this means for your stack: ask your audit partner this cycle which procedures now run on AI tools and whether that changes the evidence they request from your systems, or the fee. Discuss with ChatGPT →
Hone raises $60 million seed for agents that own ongoing work (3 minute read)
Hone, a five-month-old startup founded by former OpenAI and Cognition staff, raised a $60 million seed round led by Benchmark and Index Ventures at a $285 million valuation to build agents that take on business responsibilities for weeks or months and check in with people when needed; early work with Cognition covers sales operations such as scoring leads. Discuss with ChatGPT →
⚖️ Regulation & Reporting
At its October 7 meeting the FASB (the US body that sets accounting rules) added no new revenue projects to its agenda and asked staff to research three narrower questions: how revenue is presented for assets carried at fair value, how guarantees of a company's own performance interact with revenue recognition, and how to judge whether a payment to a customer buys a distinct good or service. Discuss with ChatGPT →
⚡ Quick Links
FIS launches an agentic cash forecasting suite (3 min read)
Blends ERP, treasury and credit data; simulates late-paying customers.
PwC says AI audit time savings are modest so far (4 min read)
Tools flag missing support; busy season barely shorter yet.
🔁 ICYMI
Worth a second look.
John Collison: AI agents will rewire internet commerce (3 minute read)
Stripe co-founder John Collison argued that agents, not people, will make a growing share of online transactions and that Stripe is rebuilding for that; per the thread, Stripe processed $1.9 trillion in 2025, up 34 percent, and its 2026 Atlas class of new startups is tracking at five times the revenue of last year's.
Why now: Published October 5, just outside this week's window. Read it with the Personal Agent Protocol story above, and with the IPO item below: Stripe sees startup formation at a high just as the exit window narrows. Discuss with ChatGPT →
IPO postponements nearly doubled in the third quarter (3 minute read)
Renaissance Capital counted seven IPOs of $50 million or more postponed or pulled in the third quarter, up from four in the second, after Oura shelved a $2.2 billion offering when shares looked set to price at the low end; analysts pointed to worries about AI spending, bond yields at a 19-year high and rising rates.
Why now: AI is helping founders start companies faster (see Collison's Atlas numbers), but it is not offsetting the market headwinds at the exit, and worry about AI spending is one of those headwinds. Discuss with ChatGPT →
Amex puts expense software and AI agents on its corporate card (4 minute read)
American Express launched an Amex Corporate card for companies with $10 million to $300 million in revenue that bundles cashback with built-in expense management and AI agents for policy questions and spend analysis.
Why now: Ramp's reported $60 billion valuation and Chase's move into payroll this week show the banks and the fintechs converging on the same bundle: card, spend controls, payroll and payments in one place. Discuss with ChatGPT →
📖 Worth The Read
a16z: the next CFO designs the operating system (10 minute read)
a16z's Seema Amble and colleagues argue the scarce CFO skill is now designing the company's operating system of data, workflows, agents and controls, citing Anthropic's finance team running more than 70 finance-specific AI skills (one drafts a 90 to 95 percent complete monthly review in about 30 minutes) and lean teams aiming for finance at about 2 percent of headcount, down from about 5 percent. Most of the tools it names, including Rillet and Stuut, are a16z portfolio companies. What this means for your stack: try the authors' rule of thumb on your close checklist, where a task done twice becomes a documented workflow and a workflow repeated twice becomes a candidate for an agent, with finance owning the approval rules. Discuss with ChatGPT →
📚 Definitions
AI terms that came up in this issue, in plain language.
Typed output
An AI answer returned in a fixed format, such as yes or no, one option from a list, or a score from 1 to 5, instead of a paragraph. OpenAI's Decisions API works this way. Your systems can act on it directly without someone, or something, reading and interpreting text.
Probability score and threshold
A number showing how confident the model is in its answer, paired with a cutoff you choose. Above the cutoff the system acts on its own; below it, a person reviews. Set the cutoff by testing on your own past transactions, not the vendor's demo.
Input and output tokens
Tokens are the small chunks of text AI models read and write, and most AI pricing is per million of them. Input is what you send in; output is what comes back. The Decisions API charges only for input, which is why it is cheap for classifying large volumes.
Agent protocol
A shared set of rules for how AI agents identify themselves, sign in and act across different companies' systems. The Personal Agent Protocol from Meta and Sierra is one attempt. Without a standard, every vendor connection is a custom build.
OAuth
The standard that lets you give one app limited access to another without handing over your password, as in 'Sign in with Google.' The Personal Agent Protocol uses it so a customer can grant an agent read-only or write access to an account.
Approval mode
A setting that decides whether an AI tool waits for a person to accept each change or applies changes on its own. Claude's Google Workspace add-on defaults to asking before every edit. For anything that feeds reporting, keep approvals on.
Computer use
An AI model operating software the way a person does, by looking at the screen, clicking and typing. John Collison calls it a bridge to the old web, since it works on systems that have no API, such as a customer's payment portal.
Finance engineer
a16z's term for finance staff who build their own automations, dashboards and agents with AI coding tools instead of waiting for IT. Think of the analyst who used to own the macros, now building small apps.
Cash application agent
Software that matches incoming customer payments to the right open invoices without a person keying them in. Stuut says its agents match 95 percent of payments automatically. Exceptions still need a human, and so does the policy for what counts as a match.