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
Visa, BILL and Billtrust let an AI agent pay supplier invoices (4 minute read)
In what the three call the first live pilot of its kind, a BILL Payment Agent registered through Visa's Trusted Agent Protocol pays buyer-approved invoices directly through Billtrust supplier portals using tokenized credentials instead of shared portal logins; the agent executes only payments the buyer already approved and does not decide what to pay, and the companies did not disclose the pilot's size. What this means for your stack: this is the controlled version of agentic accounts payable (AP), so before relying on it confirm the agent cannot pay anything outside an approved invoice, that each payment ties back to one invoice in your AP ledger, and who is liable if a misconfigured agent pays the wrong supplier, since an agent can move money faster than a clerk. Discuss with ChatGPT →
At Workday Rising, Workday made its Sana assistant the starting point for HR and finance work, made Agent Harness (for building custom agents on Workday data) generally available on Flex Credits consumption pricing, and said Sana is rolling out inside Claude, Gemini Enterprise, Microsoft Copilot, Teams and Slack, with policy checks, permissions and approvals staying in Workday. What this means for your stack: if Workday is your system of record, outside agents will reach your finance data through Workday's rules rather than around them, so review security groups and business process approvals now, because those settings become the agents' permissions too. Discuss with ChatGPT →
Ramp opens Glass, its internal AI workspace, to other companies (4 minute read)
Ramp co-CEO Eric Glyman opened early access to Glass, the workspace Ramp built to turn one employee's AI workflow into a skill the whole company can reuse, with company knowledge connectors, model routing and per-project spend meters and caps; Ramp says moving its own work onto its model router cut its AI bill 30 percent, after it spent $1.5 million on tokens in a single week in May. Two days later Ramp confirmed a $1.85 billion raise at a $60 billion valuation and said annualized revenue tops $2 billion, up from $1.5 billion in June. What this means for your stack: the part worth copying is the spend meter, whatever tool you use; give each AI project a budget owner and a cap, and treat vendor savings claims carefully, since Implicator notes Ramp's figures are self-reported with no baseline. Discuss with ChatGPT →
Also: Ramp's Glass announcement · Implicator: the skeptical read · TechCrunch: the $60 billion round
FIS opens its Claude-built AML agent to every US bank (3 minute read)
FIS made its anti-money-laundering (AML) investigations agent, built with Anthropic, available to any US bank including non-FIS customers; every output is a recommendation an analyst must confirm, each statement links to its source record, and design partner Amalgamated Bank projects more than 90 percent less time on first-level alert reviews, a projection rather than a measured result. What this means for your stack: the control design travels to any finance agent you deploy: the agent recommends, a person closes the item, and the agent's reasoning and the person's decision are logged together for audit. Discuss with ChatGPT →
Causa Prima, founded this year by Taulia founder Maex Ament and two partners, raised a $20 million seed led by Base10 Partners for software that lets a company's AI agent negotiate payment terms and resolve billing disputes with a supplier's agent, taking a cut when it arranges early payment. Discuss with ChatGPT →
🤖 AI in Finance
Stripe says agent-driven shopping went vertical in September (3 minute read)
Stripe President John Collison said purchases made by AI agents on Stripe went "vertical" starting in September, tying the jump to products launched at its May Sessions event, though he gave no volume figures; Payments Dive notes that agents paying for things still trails agents shopping, that liability for chargebacks is unresolved, and that Amazon has already blocked Meta's Muse shopping agent from its site. What this means for your stack: long-running agents are new for most consumers and many more users are coming, so revenue from agent customers may grow while the cost of serving them grows faster; agents are patient, read every policy and are already finding refund and credit paths where humans hadn’t previously found for make time to look for. So tighten refund and credit rules in your billing and support systems, tag agent-initiated orders and disputes so they show up in your order-to-cash reporting, and track refunds, credits and chargebacks as a share of revenue, not just revenue growth. Discuss with ChatGPT →
JPMorgan said outside AI agents may view some retail account data read-only but may not send money or pay bills, with Jamie Dimon warning on the bank's third-quarter earnings call that it could face thousands of rogue agents; Citi and Bank of America followed with their own limits on agent access, according to the Financial Times. What this means for your stack: if a vendor pitches an agent that pays bills by logging into your bank portal with your credentials, expect the bank to block it; ask instead whether it runs on sanctioned rails such as bank APIs, network tokens or your AP platform, which is the route the Visa, BILL and Billtrust pilot in Systems & Stack takes. Discuss with ChatGPT →
Anthropic put OpenTelemetry trace export in public beta for Claude Managed Agents, the Messages API and its SDKs, so each model call, tool use and connector call an agent makes can flow into tools such as Datadog, Splunk, Grafana or Databricks with timings, token counts and request IDs; FloQast says it uses these traces to answer auditors' questions about what its accounting agents did, and Ramp says traces helped it cut token use 10 to 20 percent on some agents. What this means for your stack: a trace shows what an agent did, not whether it should have, so pair traces with written approval rules, and check what you export first, because full-content traces can carry customer data and credentials into tools your security team and auditors have not reviewed. Discuss with ChatGPT →
On DualEntry's 101-task accounting benchmark, Google's Gemini 4 Argon scored 91.1 percent overall, ahead of GLM-5.3 Flash and Grok 4.7 at 88.1 percent and Claude Fable 5.1 at 85.1 percent, yet it managed only 75 percent on bank reconciliation, where Meta's Muse Wonder led at 91.7 percent; DualEntry sells accounting software, and its post gives no detail on how tasks are built or graded. What this means for your stack: no model wins every accounting task, so if you are picking one for reconciliations or close work, test it task by task on last month's real data instead of choosing the overall leader. Discuss with ChatGPT →
⚖️ Regulation & Reporting
AICPA pilots a CPA-led assurance report for AI agents (3 minute read)
The AICPA (the US accounting profession's main body) and the Artificial Intelligence Underwriting Company are piloting a single CPA-led engagement that pairs a SOC 2 report with an AIUC-1 certification, which live-tests agent behavior against risks such as hallucinations, data leaks and prompt injection; several CPA firms are already delivering it, with a full launch expected in early 2027. What this means for your stack: add it to vendor due diligence now; when an AI vendor hands you a SOC 2, ask whether it covers how the agent behaves or only where it is hosted, and whether an AIUC-1 is coming, since that certification requires quarterly retesting. Discuss with ChatGPT →
IRS final rules allow card payments directly to the IRS (2 minute read)
The IRS issued final regulations (T.D. 10060) letting taxpayers pay by credit or debit card directly to the IRS instead of only through third-party processors, effective for payments made on or after Federal Register publication on October 19; the IRS may cover processing fees only if it recovers them in full from card payers, and those fees are not refundable. Discuss with ChatGPT →
⚡ Quick Links
TokenFlow adds AI model budgets and attribution for finance.
🔁 ICYMI
Worth a second look.
Anthropic's ten finance agent templates, from close to KYC (7 minute read)
Anthropic released ten agent templates for financial services, including a month-end closer, general ledger reconciler, statement auditor and KYC (know your customer) screener, plus Claude add-ins for Excel, PowerPoint and Word. These might work right out of the box or you could copy the skills and tweak them for you specific business or processes. Discuss with ChatGPT →
Why now: FIS was a named customer at launch, and its AML agent reaching every US bank this week is one of the first production results; FloQast's audit traces above show the monitoring piece catching up.
Jamin Ball argues the AI-era control point is the layer that governs what agents know, see and do and keeps the audit trail, and that switching costs move from data to permissions. Discuss with ChatGPT →
Why now: Workday's Sana, the banks' limits on outside agents and Visa's agent registry this week are each a bid to be that clearinghouse.
Why NetSuite and Salesforce keep winning (4 minute read)
Michael Bianco, who built a NetSuite connector Stripe acquired, argues finance teams choose systems for stability, auditability, a deep consultant ecosystem and career safety, not interface quality, which is why challengers struggle to displace them. Discuss with ChatGPT →
Why now: With Ramp now selling accounting tools and Workday turning approvals into agent rules, it is a useful check on what actually makes finance teams switch. IMO, if these tools could significantly lower the cost of switching and show me something that gives me confidence their promises are true, I would be much more inclined to try new systems.
Adyen buys usage-billing platform Orb for $335 million (3 minute read)
Adyen agreed to buy Orb, whose billing engine meters real-time usage for customers such as Vercel and Replit, to tie usage-based pricing directly to payments as AI pushes more companies off flat subscriptions. Discuss with ChatGPT →
Why now: If agents become customers, they buy in small, frequent, usage-priced transactions, and the billing layer is where your O2C costs and disputes will start.
📚 Definitions
AI terms that came up in this issue, in plain language.
Trace
A step-by-step record of everything an AI agent did on one task: each model call, each tool it used, how long each step took and what it cost. Anthropic's new export sends these to your monitoring tools. Think of it as the agent's activity log, the thing you hand an auditor.
OpenTelemetry
An open standard for recording and sending traces and logs from software, used by tools such as Datadog, Splunk and Grafana. Because Claude's traces follow it, agent activity can sit in the same dashboards your IT team already watches, next to every other system.
Agent harness
The controlled environment an AI agent runs inside: what data it can reach, which tools it may use, what it remembers and what gets logged. Workday's Agent Harness lets companies build agents on Workday data while Workday's permissions and approvals still apply.
Skill (shared AI skill)
A saved, reusable set of instructions and steps that teaches an AI tool how to do one task your way, such as a flux commentary format. Ramp's Glass is built around sharing skills so one person's good workflow becomes everyone's starting point.
Tokenized credentials
Stand-in values that replace a real card or account number during a payment, so the real number is never passed around. In the Visa, BILL and Billtrust pilot, the agent pays with tokens instead of logging into supplier portals with shared passwords.
Agent registry
A directory where AI agents are registered and verified, so a website or payment portal can check that an agent is legitimate and acting for a real customer before letting it in. Visa's Trusted Agent Protocol and Agentic Directory play this role in the BILL pilot.
AIUC-1
A certification standard for AI agents that tests how an agent actually behaves, including hallucinations, data leaks and attempts to trick it, rather than only reviewing written policies. It requires annual renewal and quarterly retesting. The AICPA is now pairing it with SOC 2.
Consistency score
In AI benchmarks, the share of tasks a model gets right every time it is run, not just once. DualEntry runs each task three times. For accounting work a model that is right two times out of three still needs a reviewer every time.
Chargeback
When a cardholder disputes a charge and the card network pulls the money back from the merchant, usually with a fee on top. Who bears a chargeback when an AI agent made the purchase is still unsettled, which is why it matters for your order-to-cash costs.
Agent clearinghouse
Jamin Ball's term for the layer that sits between AI agents and company systems, deciding what each agent may see and do and keeping the record of what it did. Workday, the banks and Visa are each trying to be that layer in their corner.