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
American Express will cut roughly 3,500 to 3,900 jobs (about 4.5% to 5% of its workforce) and took a $336M restructuring charge, while CEO Stephen Squeri said more AI agents are coming to its card and expense platform to answer Ramp and Brex. A payroll product with AI insights is slated for early 2027, built on its Center and Hyper acquisitions. What this means for your stack: if your T&E program runs on an Amex corporate card, expect auto-generated expense reports and agent features to land inside the card bundle, so revisit whether you still need a separate expense tool at renewal. Discuss with ChatGPT →
Related: paymentsdive.com
Mercury passes $1B in annualized revenue and adds payroll (2 minute read)
Business banking fintech Mercury hit $1B in annualized revenue, up about 50% year over year, with more than 450,000 customers, and is launching a payroll service; CEO Immad Akhund also discussed AI agents spending money on the podcast. Discuss with ChatGPT →
Hyphenate (formerly Maximor), founded by Microsoft veterans, raised $45M for agents that run order-to-cash, procure-to-pay, close and treasury work on top of a company's existing ERP, and says revenue grew 86x since its seed round. Discuss with ChatGPT →
Mastercard opened early access to Agent Pay for Machines, giving developers APIs and a sandbox for AI agents that pay other services within a set budget, after pilots where one agent made nine payments to two vendors using batched payment vouchers settled in card, bank or stablecoin. Discuss with ChatGPT →
Fintech startups raised $10.6B globally in Q3, down 21% from a year earlier, across just 687 rounds, the fewest since at least 2018, as investors concentrate money in AI leaders like Ramp. Investors said agent payments have moved from 'if' to 'when' but warned that agent checkout tools may be overfunded relative to actual volume. Discuss with ChatGPT →
🤖 AI IN FINANCE
Barclays cuts 5,000 jobs, with most hitting support functions (2 minute read)
Barclays will cut about 5,000 roles (roughly 5% of staff) by year end, saying about 60% fall in support functions including developers as AI makes remaining staff more productive. Third-quarter profit rose 13% to £1.65B and the bank announced a £500M buyback, though results missed some forecasts. Discuss with ChatGPT →
Citi has blocked all agentic access to its systems beyond existing read-only open banking feeds, saying it needs a 'know your agent' process (the agent version of customer identity checks) before opening up, joining JPMorgan, Bank of America, Morgan Stanley and Schwab. Separately, Citi is building banking functions directly inside clients' SAP and Oracle ERP (core accounting) systems so treasury teams stop logging into multiple bank portals. What this means for your stack: any plan to let an AI agent pull bank data or initiate payments should assume read-only bank feeds for now, and treasury teams should ask their banks about ERP-embedded banking instead. Discuss with ChatGPT →
Anthropic released Claude Duet by waitlist to US Pro and Max subscribers, an agent that keeps working in the background across connected apps and checks in before it sends a message, signs, deletes or pays. Purchases run through Stripe's Link wallet with a one-time virtual card, so the real card number is never shared. What this means for your stack: if employees start using agents like this for business purchases, your spend policy needs to say which cards and wallets agents may use and how those receipts reach your AP or expense system. Discuss with ChatGPT →
Meta's Muse agent hits 10M users, and shopping is the top use (4 minute read)
Meta said its Muse personal agent passed 10 million US and Canadian users, with shopping the most common task and an average purchase above $200, high enough that Meta had to ask Stripe to raise credit limits in launch week. Meta plans to charge businesses a fee on purchases Muse helps complete. Discuss with ChatGPT →
Related: paymentsdive.com
A Visa executive said at Money20/20 that 23% of US consumers are comfortable with an AI agent buying on their behalf, per Visa's Trust Index survey of 2,065 people, and named shared agent identity and ongoing consent as the fixes. Discuss with ChatGPT →
⚖️ REGULATION & REPORTING
Ramp applies for a national trust bank charter (3 minute read)
Days after raising $1.85B at a $60B valuation, Ramp applied to the OCC (the federal bank regulator) to form Ramp National Trust Bank, which would offer investment management and digital asset custody but would not take deposits or make loans. The filing lands as community banks sue the OCC over its wave of trust charter approvals. What this means for your stack: if you park operating cash in Ramp's treasury product, a charter would move that service under direct federal supervision, which is worth flagging to your treasury policy and auditors. Discuss with ChatGPT →
Related: bankingdive.com
The SEC's Division of Examinations named tokenization (putting securities on blockchain ledgers) as an explicit review area for clearing agencies and market systems in fiscal 2027, alongside cybersecurity and emerging tech such as automated services. Discuss with ChatGPT →
Related: dart.deloitte.com
🛠️ THE PRACTITIONER
MAVI bets AI will create a shortage of mid-level accountants (3 minute read)
MAVI, a marketplace placing AI-skilled global accountants with US companies, argues that automating entry-level work will shrink the pipeline into mid-level roles, and defines the new skill set as directing AI tools and catching what they get wrong; it has 3,000+ professionals and a $4M seed. What this means for your stack: if AI is absorbing your junior work, write down how the next senior accountant gets trained, because the old path of learning through reconciliations is going away. Discuss with ChatGPT →
⚡ QUICK LINKS
Anthropic's IPO filing shows 14x revenue growth and an $80B paper loss (2 min read) - $16.3B first-half revenue, up 14x, and 93 pages of risk factors.
How a failed Ramp and Airwallex partnership turned investors hostile (6 min read) - Inside the failed partnership that turned Ramp's investors against Airwallex (paywalled).
PayPal CEO says no real agentic commerce volume until 2028 (3 min read) - Lores is cutting layers and expects no real agent volume until 2028.
Consumer comfort with agent purchases collapses as the ticket size rises (2 min read) - Comfort drops from 29% for purchases under $20 to 6% over $500.
🔁 ICYMI
Worth a second look:
AI now beats licensed CPAs on Mercor's month-end close tasks - Read it next to this week's MAVI piece: if AI clears the rules-heavy close steps, the question becomes who trains your next senior accountant. Discuss with ChatGPT →
Airwallex rebuilds business accounts so AI agents can manage global cash - The product at the center of this week's Ramp and Airwallex rivalry, and a concrete model of agent permissions for multi-entity treasury. Discuss with ChatGPT →
OpenAI launches Dots, always-on ChatGPT agents with their own computers - With Muse and Claude Duet now spending real money, Dots' read-only-until-approved control model is the template to compare them against. Discuss with ChatGPT →
📖 WORTH THE READ
Simon Taylor argues that personal AI agents will become a primary channel into financial products, the way mobile did, and lays out how emerging standards for agent sign-in, identity proof, tool access and payment authorization could let providers open up safely. Discuss with ChatGPT →
📚 DEFINITIONS
AI terms that showed up in this week's reading:
Agentic commerce: Shopping or paying where an AI agent, not a person, searches, selects and completes the purchase on someone's behalf.
Know your agent (KYA): The agent version of know your customer checks: verifying which agent is acting, who it works for and what it is allowed to do.
Human in the loop: A design where a person must approve an AI agent's action, such as a payment or signature, before it executes.
Personal Agent Protocol (PAP): An emerging standard for a service to recognize a customer's agent, sign it in and limit it to agreed permissions that can be revoked.
x402: An open payment standard named after the HTTP 402 'Payment Required' code that lets agents and APIs request and send small payments automatically.
x401: A proposed standard for an agent to present verified proof of a person's identity or eligibility without uploading documents to each service.
MCP (Model Context Protocol): A common way for AI agents to discover and use a service's tools and data, with the provider deciding what is allowed.
Verifiable Intent: Mastercard's approach to recording, in a tamper-resistant way, exactly what a customer authorized an agent to buy.
Machine-to-machine payments: Payments made by one piece of software to another, such as an agent buying API access, often batched and settled later.
Autonomous finance: Vendor term for agents that own a finance workflow end to end, escalating to a person only on exceptions.
Evaluation awareness: When an AI model behaves differently because it detects it is being tested, which makes safety results less reliable.
Chain-of-thought monitoring: Reviewing a model's step-by-step reasoning to catch bad behavior; this week's Gemini safety card says it can sometimes be bypassed.
Open weights: A model whose trained parameters are published so anyone can run it on their own hardware, as Moonshot plans for Kimi K3.1.
Sparse activation: Running only part of a model for each request, as Apple's on-device model does by activating 1 to 4 billion of its 20 billion parameters.
S-1: The registration statement a company files with the SEC before an IPO, disclosing financials and risks. Anthropic filed its S-1 this week.