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's CEO pitches Ramp as a productivity company, not a fintech (2 minute read)

In a conversation with Simon Taylor, Eric Glyman says 70% of Ramp's R&D team is building unreleased products, its non-engineers wrote 7,000 pull requests (code changes) last quarter using AI, and AI is becoming the third major category of business spend. Discuss with ChatGPT →

Connor's take: Ramp keeps shipping products it says it built for itself first, and the pace is hard to ignore: Router, Ramp Glass, an accounting agent, procurement agents, AI spend management, receivables, and a bank charter application, all this year. For a 50-person company that is close to a full finance stack behind one login. For a large org with an ERP, a procure-to-pay (P2P) suite, and a treasury system already wired together, I still struggle to see the case for ripping any of that out. But Glyman calling Ramp a productivity company is interesting. Large orgs may never run Ramp as their accounting system of record but still buy it as the layer employees, and increasingly their agents, actually touch.

Intuit bets accounting firms will trade billable hours for advisory work (6 minute read)

At Intuit Connect, Intuit previewed Agentic Books Close, AI agents that run month-end prep using a firm's own procedures, and said firms on its Accountant Suite save more than nine hours per client each month. Analyst firm Futurum notes the close agent is still a pilot and will become a paid add-on. What this means for your stack: if your outside accountants bill by the hour for close work, ask now how they plan to price it once agents do the prep. Discuss with ChatGPT →

Also see: futurumgroup.com

Connor's take: The real news here is Intuit treating accounting firms as customers instead of a sales channel. Once an agent does the close prep, an hourly bill for that prep either shrinks or turns into an awkward conversation, so firms have to reprice or watch margin evaporate. Vendors are moving the same way from the other side: UiPath will charge per document its agents process, not per implementation project. If you use an outside firm, ask two things this quarter: how will they price work their agents now do, and who reviews what the agent produced? I would push for a fixed fee on close support with advisory priced separately, because paying by the hour for work a machine did is the worst of both worlds.

Natural lets AI agents choose ACH, wire, or instant payment rails (3 minute read)

Natural's new Direct product lets AI agents pick among ACH, wire, real-time payments, and soon paper checks when they execute a payment. What this means for your stack: if an agent can choose the rail, your AP approval rules need to be set by rail and amount, not just by vendor. Discuss with ChatGPT →

Aion launches an agent that does finance work and waits for approval (3 minute read)

Aion's new agent sorts bills, chases invoices, prepares transfers, and issues cards with limits the team sets, while a person approves anything that moves money. Discuss with ChatGPT →

UiPath will price its invoice agents per document, not per project (3 minute read)

UiPath's ready-made procurement and invoice-processing agents arrive in early 2027 and will be billed per document processed instead of as a software implementation project. What this means for your stack: per-document pricing turns AP automation into a variable cost, so model it against your invoice volume and seasonality before signing. Discuss with ChatGPT →

Instinct gives AI shopping agents their own wallet with a spending cap (3 minute read)

Instinct, a personal AI assistant, now lets US users load a Stripe Link balance that only their agent can spend, and any purchase under a per-purchase limit the user sets goes through without asking for approval. The money sits in USDC (a dollar-pegged stablecoin) and pays merchants through a normal prepaid Visa card, so stores do not need to change anything. What this means for your stack: if employees start delegating purchases to agents, your expense policy needs a line on who is liable for an agent purchase that was within the cap but outside what the person intended. Discuss with ChatGPT →

Also see: stablecoininsider.org

Also see: runtimewire.com

Connor's take: Agent payments just moved from demo to wallet. Instinct lets you park money only your agent can touch and set rules on what it buys and how much, which kills the approve-every-purchase problem and creates a new one. When the agent buys something inside the cap that you never actually wanted, who eats it: you, the agent maker, the wallet, or the merchant? RuntimeWire's read of Link's terms is that the risk lands on the user, and card dispute rules were written for stolen cards, not for a bot that followed instructions too literally. If your company lets employees hand spending to agents, write the answer into your expense policy before your first disputed charge writes it for you.

AI agents will pick whichever payment method pays them the most (4 minute read)

Plaid's head of payments says agents will choose a payment method based on the goals their owner sets, such as maximizing card rewards or discounts, which could shift the mix away from cards toward bank payments as agents negotiate with merchant-side agents. What this means for your stack: an agent that hunts every promo code will find the ones your team forgot to expire, so audit who can create discount codes and how they post to revenue. Discuss with ChatGPT →

Connor's take: This is the part of agentic commerce almost nobody in the CFO's office is modeling. A person forgets the promo code at checkout; an agent never does, and it will stack the best card reward, the cheapest rail, and every discount it can find on every single order. Personally, I would point an agent at my points strategy and let it run. Corporately, flip the lens: if a stale promo code, an over-generous partner discount, or a forgotten loyalty credit is sitting in your stack, agents will find it and use it at scale. Know who can create discount codes, when they expire, and whether they land as a reduction of revenue or get buried in marketing. The FASB proposal in Regulation & Reporting makes that last question more than academic.

Anthropic adds reports showing what Claude spending actually buys (3 minute read)

Anthropic released smart reports, employee surveys, and pooled team budgets for Claude Enterprise, grouping usage into workstreams with the spend, finished deliverables, and recurring problems for each. What this means for your stack: pooled budgets give you a way to allocate AI spend to departments in your planning model instead of carrying it as one IT line. Discuss with ChatGPT →

DoorDash has seen only several hundred orders from AI agents (2 minute read)

Against roughly 3 million deliveries a day, DoorDash said AI agents have placed only a few hundred orders, though it is building its own agents and looking at ad opportunities inside agent platforms. Discuss with ChatGPT →

Amazon tells engineers token counts are a bad measure of AI value (3 minute read)

Amazon published a memo calling token usage, AI tool adoption rates, and code-change counts deeply flawed measures of AI work and warned against tokenmaxxing, burning AI usage just to look productive. What this means for your stack: if your AI enablement dashboard tracks logins or prompts, add an outcome metric like days to close or exceptions cleared. Discuss with ChatGPT →

Anthropic's amended IPO filing adds Q3 revenue, losses, and compute bills (1 minute read)

Anthropic's updated S-1 (IPO registration filing) shows nine-month revenue of $33.24B, a nine-month net loss of $104B, and about $895B of computing commitments. DealBook reported the company stopped recognizing revenue from customers violating its usage policy and that five customers make up 25% of revenue. Discuss with ChatGPT →

FASB proposal clarifies when payments to customers reduce revenue (3 minute read)

The FASB (the US accounting rule-setter) proposed an update to ASC 606 (the revenue standard) clarifying when payments made on a customer's behalf, or to a customer's customer, count as a reduction of revenue; comments are due November 23. What this means for your stack: if you fund promos, rebates, or discounts that reach end customers through a partner, check whether you are booking them as marketing expense or as a revenue reduction. Discuss with ChatGPT →

Alphabet's quarterly filing now warns AI could cause catastrophic incidents (4 minute read)

Alphabet expanded its AI risk factor (the risk disclosures in its quarterly filing) to say its agents acted outside their intended scope in testing and that misuse could lead to severe or catastrophic incidents, and AI-agent mentions across Alphabet, Meta, and Microsoft risk sections rose from 5 to 31. What this means for your stack: if you file with the SEC and run agents in finance, expect your auditors and disclosure committee to ask whether your risk factors mention them. Discuss with ChatGPT →

AI output is overwhelming the people who have to review it (6 minute read)

Linear's CEO argues that AI lets everyone produce more work, but all of it lands on the same limited capacity to review and decide, and output is not proof of value. What this means for your stack: before adding agents that draft reconciliations or flux commentary, size the reviewer time they will consume. Discuss with ChatGPT →

Shopify takes Shop Pay into personal AI agents (16 min). Shop Pay now works inside personal agents and off-Shopify brands.

Why AI adoption stats range from 20% to 90% (4 min). McKinsey, Ramp and Census measure three different things.

AI can cut the value of a job without eliminating it (7 min). AI reprices tasks well before it removes roles.

Cisco hands one engineer an unlimited AI budget to rewrite legacy code (2 min). One engineer, unlimited tokens, 98% of legacy code, 24 months.

Glasser lets agents buy premium data one call at a time (2 min). Agents can now buy premium data per API call.

SEC hosts a roundtable on machine-readable filing data and AI (1 min). Public SEC roundtable Nov. 9 on machine-readable filings and AI.

Worth a second look:

Meta, Shopify, Stripe and Walmart back a standard for AI agents. With agents now holding their own wallets, the Personal Agent Protocol's model of customers granting an agent scoped access is the clearest blueprint for who authorized what. Discuss with ChatGPT →

a16z: the next CFO designs the operating system. A useful companion to this week's billable-hour debate: if agents do the prep, the finance leader's job becomes designing the workflows, controls and approval rules around them. Discuss with ChatGPT →

Avalara ships its tax agent hub and sets EU e-invoicing dates. Germany's e-invoicing date is January 2027, about two months out, so this is the week to check your billing system's readiness. Discuss with ChatGPT →

The $81B Anthropic loss is mostly an accounting mark, not cash burn (9 minute read)

Jamin Ball explains that Anthropic's $109B of convertible notes are old investor notes marked to the current share price, and nearly all of the $81B first-half loss is that fair-value (mark-to-market) change, while the company has burned only about $7B of the roughly $141B it raised. Discuss with ChatGPT →

DeepMind essay on when AI ability turns into economic change (15 minute read)

Argues that AI already matches specialists on self-contained deliverables but lags on coordination across people and systems, and that wages and job definitions will show the change first. Discuss with ChatGPT →

Resisting cognitive surrender to AI output (16 minute read)

Describes cognitive surrender, accepting whatever an agent produces without checking it or considering alternatives, and how to keep judgment in the loop. Discuss with ChatGPT →

AI terms that showed up in this week's stories.

Agentic commerce: Shopping where an AI agent finds, compares and pays for things on a person's or company's behalf, often without a human clicking checkout.

Per-purchase limit: A cap the owner sets on a single agent purchase; anything under it goes through without approval, anything over it asks first. It limits each transaction, not total spend.

Stablecoin (USDC): A digital token designed to hold a steady value, usually one US dollar. USDC is the most common one used for agent payments.

Payment rail: The network money moves on, such as card networks, ACH, wire, or real-time payments. Each has different cost, speed and reversal rules.

Agentic close: AI agents that run month-end close prep (reconciliations, accruals, schedules) against a team's own procedures, leaving people to review and approve.

Tokenmaxxing: Running AI tools on low-value work just to push usage numbers up, usually because a dashboard or leaderboard rewards token counts instead of results.

Tokens: The chunks of text an AI model reads and writes. Most AI services bill by tokens, so token counts drive the AI bill.

Cognitive surrender: Accepting whatever an AI produces without checking it, questioning it, or considering alternatives.

Recursive self-improvement (RSI): AI systems that help build more capable AI systems. Anthropic's amended S-1 says it may soon meet its own internal definition of it.

Distillation: Training a cheaper model by feeding it another model's outputs. Anthropic stopped recognizing revenue from accounts it said were doing this against its usage policy.

Mark-to-market (fair value): Revaluing an asset or liability at its current market price each period. It explains most of Anthropic's headline loss: its convertible notes rose in value with the share price.