The finance systems & AI brief for controllers and CFOs
Two stories this week are really one story. A Goldman partner said out loud that automating the analysis junior bankers used to do risks what he calls cognitive atrophy, then conceded that the bank's own AI platform is better at sounding thorough than being thorough. Days later the US Tax Court found three cases cited in a filed brief that do not exist. Goldman is worried about what people will stop catching. The Tax Court just published an example.
Also worth your time: RSM shipped AI control testing built on Andera, the company we resurfaced three weeks ago, and Protiviti found 77 percent of finance teams using AI while only 14 percent can point to a strategy.
Was this newsletter forwarded to you? Sign up here.
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
Thomson Reuters, the company behind Checkpoint and CoCounsel, trained its own large model on its proprietary tax and legal content instead of renting one, spending roughly $40 million over several years with a final training run that cost under $450,000. It is embedded inside its products rather than sold as a separate subscription. What this means for your stack: the cost of a specialized model has fallen far enough that your research and content vendors may start shipping their own, which moves the diligence question from which model they use to whose data it was trained on. Discuss with ChatGPT →
Intuit moves the QuickBooks Online Reports application programming interface to a new backend on September 1 (the day before this newsletter is released), changing how empty values, row ordering and account hierarchies come back, which can leave reports pulled by outside tools misaligned until the code behind them is updated. What this means for your stack: if anything outside QuickBooks pulls your reports, a dashboard, a consolidation workbook or a custom script, ask whoever maintains it whether they tested against the migration flag. Discuss with ChatGPT →
This article argues that reconciliation is moving into the payment itself, because a transaction carrying the customer identifier, invoice reference and virtual account number arrives ready to apply rather than needing to be matched afterward. It also reports that 71% of executives at companies above $1B in revenue say organizational readiness, not the technology, is what limits AI. What this means for your stack: before buying another cash application tool, ask whether your bank or payment provider can attach remittance data to the transaction, because that removes the matching work instead of automating it. Discuss with ChatGPT →
A survey of 1,808 practitioners put UltraTax CS at 22.3% share and Lacerte at 16.7%, with Drake Tax and Lacerte tied highest on overall rating at 4.4 out of 5, and found 65% now use AI for tax research while only 16% have no plans to use it at all. Discuss with ChatGPT →
🤖 AI IN FINANCE
A Goldman Sachs partner said automating the analytical work junior bankers used to do risks what he called cognitive atrophy, because the apprenticeship model depends on people doing that work themselves. He also said the bank's own AI platform is currently better at sounding thorough than being thorough. What this means for your stack: if you are automating the reconciliations and schedules your staff accountants learn on, decide now what replaces that training, because the review judgment you will need from them in five years is built by doing the work. Discuss with ChatGPT →
RSM, a top ten accounting firm, launched Beacon, which reads control descriptions and supporting evidence and produces annotated, review-ready control testing workpapers, flagging anything ambiguous for a person rather than deciding it. It is built with Andera, the audit AI company that raised $37M in June, and RSM is looking at extending it past SOX (the internal control rules for public companies) into SOC, NIST and PCI work. What this means for your stack: when your auditors start reading evidence this way, the consistency of what you hand them, meaning predictable file names, dated screenshots and clean Excel support, is what decides how much rework comes back to you. Discuss with ChatGPT →
Protiviti's global finance survey found 77% of finance organizations now use AI and use in forecasting rose from 58% to 76% in a year, but only 35% say they measure the return effectively and just 14% run AI against a defined strategy. What this means for your stack: the gap between 77% using it and 14% having a strategy is where duplicate tool spend hides, so an inventory of what your team already pays for is usually the fastest saving on the table. Discuss with ChatGPT →
EY reorganized its consulting and advisory work into five cross-functional bundles covering enterprise trust, growth and mergers, productivity and technology, business and operations, and finance, each combining its AI platforms, proprietary data and methodology rather than selling by service line. Discuss with ChatGPT →
⚖️ REGULATION & REPORTING
The United States Tax Court found that three cases cited in a taxpayer's brief do not exist, and the judge wrote that submitting a brief with fictitious caselaw is a recipe for sanctions. The article maps the exposure back to Circular 230 (the rules governing practice before the IRS) and the AICPA tax standards, and includes a checklist for firm AI policies. What this means for your stack: any AI output that cites an authority needs a verification step before it leaves your team, and the cheapest version is a standing rule that whoever signs checks the citation, not the tool that produced it. Discuss with ChatGPT →
A Treasury watchdog report found the IRS took in a record $5.3 trillion in fiscal 2025 while examination staffing fell 27% in a year and revenue from examinations fell 35%. Examination starts for individuals dropped 30%, and for taxpayers earning above $400,000 they dropped 27%. Discuss with ChatGPT →
🛠️ THE PRACTITIONER
Anthropic's coding tool runs several agent sessions at once, and for projects tracked in Git each session works inside its own isolated copy so one session cannot overwrite another until changes are deliberately committed. A companion feature keeps the agent running on your own machine while you drive it from a phone or a browser, so files and execution never leave the local environment. What this means for your stack: this is the clearest working model yet for segregation of duties with AI agents, one isolated workspace per task and a review step before anything merges, and it is worth copying when you scope agent access to your own systems. Discuss with ChatGPT →
⚡ QUICK LINKS
Google Cloud put Gemini Enterprise for Financial Services and a parallel legal version into preview, with a managed research agent, more than 50 role-specific skills and enterprise data connectors. LEDGER covered this announcement last week; the new detail is the legal build and the disclosure that close to 90% of the Fortune 100 already use Gemini Enterprise.
Month-end roundup of August's five largest fintech deals, including Visa's $2.4 billion all-cash purchase of fraud prevention company BioCatch and confirmation of the Stripe and OpenRouter deal above $7 billion.
Fraud prevention company Apate.AI raised $8.15 million to expand its use of AI agents against scam operations.
🔁 ICYMI
Worth a second look.
Published August 20 and never sent in a LEDGER issue. A working question bank for the hire most finance teams get wrong, with a chart of accounts review as the take home. Discuss with ChatGPT →
Goldman is worried that reviewing AI output all day makes people worse at thinking. This is that worry with a number attached: 1,053 professional testers caught a clearly dangerous command 13.6% of the time, and worse the longer the session ran. Discuss with ChatGPT →
Protiviti found that 14% of finance teams use AI against a defined strategy. If you want to be in that 14%, this is the clearest published account of what the strategy actually contains. Discuss with ChatGPT →
📖 WORTH THE READ
This article argues that the number of months needed to recover customer acquisition cost is a more usable operating metric than the lifetime value to acquisition cost ratio, because that ratio compounds too many assumptions to steer on day to day, and it includes benchmarking tables across nine software sectors. What this means for your stack: if your board pack leads with lifetime value to acquisition cost, put payback months beside it, because payback is the number your sales and marketing leaders can actually move inside a quarter. Discuss with ChatGPT →
📘 DEFINITIONS
Terms that showed up in this issue
Inference. Running a trained model to get an answer, as opposed to training it in the first place. Training is the one time cost of building the model. Inference is what you pay every time somebody asks it something. It matters for finance because inference is the line that scales with usage, which is why AI spend behaves like a utility bill rather than a software subscription. When a vendor quotes you per seat but the underlying cost is per token, somebody is absorbing the difference, and it is worth asking who.
Prompt injection. Hiding instructions inside content an agent reads, a web page, a PDF, an email or a supplier invoice, so the agent follows those instructions instead of yours. LEDGER defined this in Issue 031 and it earns a second entry because this issue puts agents in front of a lot more documents. The accounts payable version is concrete: a vendor sends a PDF invoice carrying white text on a white background that reads approved vendor, remit to the account below. A person never sees it. An extraction agent does. The control is that the agent proposes and a deterministic rule approves, never the other way around.
Hallucination. A model producing something fluent, specific and false. The tax case in this issue is the clean example: three citations with volume numbers, page numbers and years, formatted exactly like real ones, for cases that never existed. Hallucinations are not vague or hedged, which is precisely why they survive a skim and only fail a check.
Frontier model. The largest and most capable class of general purpose model at any given moment, the tier the leading labs compete in. Thomson Reuters using that word about its own model this week is the news. The tier used to be defined by who could spend billions, and a specialized version now costs a content company about forty million dollars to reach.
Continual learning. Training an existing model further on new material instead of starting from scratch. Thomson Reuters built on Qwen, an openly available base model, then kept training it on proprietary tax and legal content. It is why the final training run cost under $450,000 against a $40M program. Most of the general capability was already paid for by somebody else.
Open weights. A model whose trained parameters are published so anyone can download and run it themselves rather than calling it over the internet. Thomson Reuters released a small version of its model this way under an academic license. For finance teams the relevance is deployment, because an open weight model can run inside your own environment, which changes the data residency conversation with your auditors.
Git worktree. A way of checking out several working copies of the same project at once, each isolated from the others. Claude Code now gives every parallel agent session its own worktree, so one session cannot overwrite another until changes are deliberately committed. The finance analogy is a separate scratch ledger per preparer, with nothing touching the books until it is reviewed and posted.