LEDGER

The finance systems & AI brief for controllers and CFOs

Amex puts expense software and AI agents on its corporate card (4 minute read)

American Express launched a new Amex Corporate card for companies with $10 million to $300 million in revenue that bundles cashback with built-in expense management, setting policy by employee or department and flagging out-of-policy spend, with AI agents that answer policy questions and analyze spending, plus accounts payable automation, due later this year. The Wall Street Journal frames it as Amex's answer to Ramp, valued at $44 billion, and Brex, which Capital One bought for $5.15 billion. What this means for your stack: if you run Amex for travel and a separate spend platform for everything else, ask Amex for its ERP sync and receipt-matching details now, because consolidating cards and expense software could cut a tool, but only if the accounting integration holds up. Discuss with ChatGPT →

Also: WSJ story via Yahoo Finance (readable copy)

Trintech adds accruals, exceptions and data agents to its close platform (3 minute read)

Trintech, which sells reconciliation and close software, launched three AI agents at its Chicago user conference: one that pulls and cleans data from multiple sources, one that recommends accrual amounts from historical patterns and flags estimates outside the trend, and one that sorts and investigates reconciliation exceptions, joining its earlier flux and variance agents. What this means for your stack: an agent that proposes accrual amounts is making an estimate, so require it to show the history behind each number and keep a reviewer sign-off, the same as you would for a staff accountant's accrual. Discuss with ChatGPT →

Airwallex rebuilds business accounts so AI agents can manage global cash (3 minute read)

Airwallex redesigned its Business Accounts to show balances and activity across all of a company's legal entities under one login, with separate accounts for operating cash, yield and expenses, payouts on local rails in more than 120 countries, and support for AI agents that can rebalance liquidity, put idle cash to work and time currency conversions within permissions and approvals the company sets. What this means for your stack: if you run treasury for several entities across countries, list which moves an agent could make on its own (sweeps under a set threshold, for example) and which need two approvers, before any vendor switches this on for you. Discuss with ChatGPT →

AI now beats licensed CPAs on Mercor's month-end close tasks (5 minute read)

Mercor gave 12 licensed CPAs, averaging about five and a half years of experience, four month-end close tasks from its APEX-Accounting benchmark (a standardized test for AI); they averaged about 37 percent and took 30 to 180 minutes per task, while Claude Opus 5 scored 100 percent on all 20 attempts in under 10 minutes and GPT-5 reached about 69 percent. Claude cost $0.21 per scoring criterion met against $10.35 for the accountants, though the authors stress the tasks test detail-oriented instruction following, not the whole job. What this means for your stack: the gap is widest on tightly specified, rules-heavy steps, so pick two recurring reconciliations or accrual calculations, run an AI tool on them alongside your team for one close, and score both against the same checklist before changing who does the work. Discuss with ChatGPT →

CFOs say a quarter of finance time now goes to checking AI (4 minute read)

A Datarails survey of 270 CFOs at US companies with more than $100 million in revenue found finance teams spend 26 percent of their time verifying or correcting AI output, 75 percent name lack of auditability as the biggest problem, 56 percent have received materially different answers to the same prompt, and only 4 percent trust AI with the month-end close. Another 32 percent went more than 10 percent over their AI budgets in the past year. What this means for your stack: track review hours on AI-assisted work as a separate line for one quarter, because if checking eats a quarter of the week, that cost belongs in every AI business case next to the time saved. Discuss with ChatGPT →

OpenAI launches Dots, always-on ChatGPT agents with their own computers (4 minute read)

OpenAI's Dots are agents that keep working on assigned goals in the background, each with its own cloud computer, access to more than 4,000 connected apps plus Slack and Teams, and one is included for Pro and Business Premium subscribers. They stay read-only while researching and must clear an auto-review step before consequential actions; one early tester's agent spotted a customer he had forgotten to bill and drafted the invoice for his approval. What this means for your stack: an always-on agent that drafts invoices or handles procurement is a new kind of user, so decide now which systems it can read, which actions need a named approver, and how its activity shows up in your access reviews. Discuss with ChatGPT →

Bank AI job postings rise 49 percent as orchestration skills take off (4 minute read)

Hiring data from Draup counted 139,819 AI-related job postings at banks this year, up 49 percent, with mentions of agent orchestration (coordinating several AI agents on one process) up 1,721 percent and responsible AI postings up 657 percent at banks including JPMorgan Chase, Citigroup and Capital One. Discuss with ChatGPT →

Salesforce survey: 46 percent of finance leaders now steer company AI strategy (3 minute read)

Salesforce's CFO Priorities Report, based on 865 finance leaders across three continents, found 46 percent are now primary decision makers on their company's AI strategy and 89 percent of those using AI report a positive return, yet 60 percent still complete at least one in five workflows by hand and 51 percent call security the top blocker to expanding AI. Discuss with ChatGPT →

PCAOB sets its agenda: fraud, going concern and data and technology (2 minute read)

The PCAOB (the regulator that oversees public company auditors) published agendas with standard-setting projects on fraud, going concern, auditor independence, noncompliance with laws and comfort letters, plus research on data and technology and audit committee communications, after its first public comment period in June. What this means for your stack: the fraud and data and technology projects point toward auditors asking for more system-generated evidence, so document how data moves from source systems to the ledger before they ask. Discuss with ChatGPT →

Five US exchanges start overnight stock trading on December 6 (5 minute read)

Nasdaq, NYSE Arca, Cboe EDGX, MEMX and 24X are approved to run overnight sessions on Sunday through Thursday nights starting December 6, putting US stocks on about 23 hours a day, five days a week, with clearing and the consolidated price feed already extended to match. Overnight trading is still under 1 percent of share volume but up 359 percent from a year ago. What this means for your stack: check that insider trading windows, buyback instructions and your equity plan administrator's settings still work when your shares can trade overnight, and confirm every policy that cites a stock price names the official close. Discuss with ChatGPT →

Also: WilmerHale: what firms should know before 23x5 launches

GAO puts federal tax fraud at $116 billion to $304 billion a year (3 minute read)

The Government Accountability Office estimated tax fraud costs the US $116 billion to $304 billion a year, about 2 to 6 percent of estimated tax liability for 2022, and said the IRS still lacks an agency-wide antifraud strategy; the IRS partially agreed with both recommendations. Discuss with ChatGPT →

PayNearMe launches an AI agent for payment servicing and collections (2 min read) Calls and texts customers, schedules payments, escalates hard cases to people.

Finsider automates quality-of-earnings work for small-deal diligence (3 min read) Beta firm cut engagement time 60 percent. "Nobody's signing generative output."

A founder's map of the 2026 AI finance stack (8 min read) Assistants up front, AI-native ERPs underneath; by loopfour's CEO.

OpenAI holds back GPT-6.1 Astra over scope and authorization (3 min read) It failed on staying within the authority it was given.

Worth a second look.

Grant Thornton: 84 percent of CFOs say AI returns meet or beat expectations (5 minute read)

Grant Thornton's survey of 229 finance leaders found a record 80 percent expect profit growth over the next year, 40 percent say AI returns beat expectations and another 44 percent say they met them, and 39 percent name finance and accounting as a top area for AI transformation.

Why now: Published September 24 and never sent in LEDGER. Read it beside Datarails: in the same quarter, CFOs say AI is paying off while their teams spend a quarter of the week checking it. Discuss with ChatGPT →

BCG's playbook for CFOs funding and scaling AI agents (8 minute read)

BCG's James Tucker estimates about 80 percent of FP&A work is data extraction, cleansing and basic calculation suited to agents, and cites one deployment that automated more than 90 percent of business intelligence work in nine months instead of the usual three years.

Why now: Never sent in LEDGER. It is the planning-side version of Mercor's finding: AI wins on structured work, and BCG says that is most of FP&A. Discuss with ChatGPT →

OpenAI's ChatGPT for Financial Services (Issue 036, September 16)

Why now: This week's Dots launch is the always-on agent layer on top of it. Worth revisiting what OpenAI built in for finance users, including cited outputs and audit log exports, before an agent version shows up in your workspace. Discuss with ChatGPT →

Simon Taylor: fintech is dead, and agents and tokens take its place (11 minute read)

Simon Taylor argues fintech's playbook of apps, cloud and APIs is now table stakes, and the next shift runs on three tracks: AI models making decisions, stablecoins and other tokens keeping the record, and agents with wallets and spending limits becoming the customer. He points to Stripe's roughly $7 billion deal for AI router OpenRouter, Ramp passing $1 billion in annualized revenue, and Allica Bank deciding small business loans end to end in 12 minutes. What this means for your stack: ask your banking and payments partners how they will handle payments started by your own AI agents, including spending limits, approvals and audit logs, because Taylor expects agents carrying a budget and a mandate to become the main way companies buy and pay. Discuss with ChatGPT →

AI terms that came up in this issue, in plain language.

Always-on agent
An AI agent that keeps working toward a goal in the background, on its own computer, instead of waiting for each prompt. OpenAI's Dots work this way. In control terms it is closer to a standing employee than a tool, so it needs access rights, approval limits and a log like one.

Human baseline
A score from real professionals doing the same test as an AI, so the AI's result means something. Mercor's licensed CPAs averaged about 37 percent on its close tasks. Without one, a benchmark score tells you how a model ranks, not whether it can do the job.

Rubric-based scoring
Grading AI work against a checklist of specific requirements, such as each number and step a task must get right. Mercor priced results per criterion met, which shows cost per correct output rather than cost per attempt.

Verification time
The hours people spend checking and fixing AI output before it can be used. Datarails puts it at 26 percent of finance time. It is the hidden cost line in most AI business cases.

Non-deterministic output
Getting a different answer when you ask an AI model the same question twice. 56 percent of CFOs in the Datarails survey have seen it. It is why you test AI on repeat runs, not one demo, before trusting it with a recurring task.

Auditability
Being able to show how a number was produced, from which data, by which step, and who approved it. Three in four CFOs call its absence the biggest problem with finance AI, and it is what your auditors will ask for first.

Agent orchestration
Designing and running several AI agents so they hand work to each other in the right order, with the right data and checks. Bank job postings mentioning it rose 1,721 percent this year. LangGraph and LlamaIndex are popular software frameworks for building it.

Responsible AI
The policies, testing and oversight that keep AI fair, accurate, secure and explainable. Bank postings for these roles rose 657 percent, a sign governance work is growing faster than the build work.

Scope and authorization
Whether an AI agent stays within the task and the permissions it was given. OpenAI held back GPT-6.1 Astra because it fell short here. For finance, it is segregation of duties for software.

Agentic business account
A bank or payments account built so an AI agent can operate it, moving money between entities, investing idle cash or converting currency, within rules and approvals the company sets. Airwallex launched one this week.

Intelligence tokens and value tokens
Simon Taylor's split between units of AI work (intelligence tokens, whose output varies) and units of money on a shared ledger (value tokens, such as stablecoins, which do not). His argument is that finance increasingly uses the second to pay for the first.

Stablecoin
A digital token designed to hold a steady value, usually one dollar, backed by cash and short-term government debt. USDC is the best known. Visa now settles with some US card issuers in USDC seven days a week, including weekends when banks are closed.

23x5 trading
Stock trading that runs about 23 hours a day, five days a week, with a one-hour pause each night. Five US exchanges start it December 6. The official closing price still sets most valuations, but more trading will happen while your team is asleep.