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
Lightspeed's Sam Eisler, whose firm backs DualEntry, argues that next-generation ERP (the core accounting system) startups stalled because they built add-ons around NetSuite and SAP instead of solving the migration itself, citing Gartner's finding that 70 percent of ERP projects miss their original goals. His reverse wedge case is that AI can now move a company's history in about 24 hours instead of a six to nine month implementation. What this means for your stack: test the migration claim before the product demo, by asking any AI-native ERP vendor to convert a sample of your own messiest entity, intercompany and custom-field history. Discuss with ChatGPT →
Mercury, a business banking company with more than 300,000 customers, launched Mercury Books, a double-entry accounting system that categorizes and reconciles account, card, invoicing and bill pay activity as it happens, with an AI agent that can post journal entries and change the chart of accounts on request. It is free through December and 35 dollars a month after that, with unlimited seats for outside accountants. What this means for your stack: if a subsidiary or a portfolio company already banks with Mercury, this may displace QuickBooks there, so decide up front which journal entries and chart of accounts changes the agent may make without a person approving them. Discuss with ChatGPT →
Stripe's head of Connect reports new platform businesses on Stripe up 182 percent year over year and self-serve integrations up 360 percent, with platforms launched in January reaching 1 million dollars in payment volume faster than any earlier group, despite software stocks losing about 1 trillion dollars of value in 30 days earlier this year. More than 55 percent of new Stripe integrations in August were built with AI help. What this means for your stack: Stripe's test for a platform that lasts is whether it runs a real workflow, keeps valuable data and moves money, which works as a three-question screen for every point tool on your renewal list. Discuss with ChatGPT →
Claude Projects now run as a coordinator that splits work across parallel threads, each its own cloud session, with project memory that builds up over time and a library of past files and outputs, starting in beta for some Pro and Max subscribers. The same day, Anthropic folded Cowork into the main Claude app and launched Docs and Slides in beta, with checkpoints that pause automated work for a person to review. What this means for your stack: a coordinator handing work to parallel threads is the preparer and reviewer split you already run at close, so put the human checkpoint wherever someone signs off today, not only at the end. Discuss with ChatGPT →
Emburse moves from expenses into accounts payable and payments (2 minute read)
Emburse, best known for expense management, released Emburse AP for mid-sized finance teams, covering purchase order and non-purchase order invoices with two- and three-way matching, duplicate detection, AI anomaly flags and role-based permissions that support segregation of duties, with payment by ACH, virtual card or check in the US and Canada. Discuss with ChatGPT →
🤖 AI in Finance
Agents pushed Harvey's gross margin to negative 50 percent (5 minute read)
Legal AI company Harvey sells flat per-seat licenses but pays AI providers per token, and after a March agent update drove token use up 20 times while revenue only doubled (still impressive), its gross margin fell from about 50 percent to about negative 50 percent. It moved to its own model, built on the open-weight Kimi K3, and cut cost per query about 90 percent, while Bloomberg reports Ramp, Abridge and Rogo are weighing similar moves. What this means for your stack: at your next AI renewal, get written notice before the vendor changes the model behind your workflows and the right to rerun your acceptance tests when it does, because a locked price can now sit on top of a different model. Discuss with ChatGPT →
Nearly half of CFOs say they follow AI over their own judgment (3 minute read)
A Board survey of 300 CFOs, CIOs and COOs at companies above 100 million dollars in revenue found 61 percent name large language models among their top outside influences on strategy, and 48 percent of CFOs follow an AI recommendation that conflicts with their own judgment, against 11 percent of COOs. Only 39 percent have a formal governance process for AI-driven decisions, and 83 percent say their board acted on a forecast it knew was outdated. What this means for your stack: if AI is shaping board-level calls, keep the prompt and the output with the forecast version it relied on, the same way you keep support for a journal entry. Discuss with ChatGPT →
An Auditoria.AI survey of about 300 finance leaders found 66.5 percent increasing AI investment, but just 21 percent reporting meaningful, measurable success and 58.4 percent still exploring or piloting, while 44.5 percent of teams spend 11 or more hours a week chasing follow-up work in the cash cycle. Discuss with ChatGPT →
The AI safety debate lands on the CFO's desk (2 minute read)
After Anthropic disclosed models that escaped test environments and OpenAI disclosed a cyber incident affecting Hugging Face, CFO Leadership Council founder Jack McCullough told CFO Dive that AI is now a capital allocation and governance decision, and that a business case which leaves out the cost of making the technology safe is incomplete. What this means for your stack: add testing, monitoring and incident response as their own lines in every AI business case, next to the labor savings, so approvals reflect the full cost. Discuss with ChatGPT →
⚖️ Regulation & Reporting
French property services company Emeria, backed by Partners Group and TA Associates, added 20 million euros of projected savings from AI agents to its trailing EBITDA (earnings before interest, taxes, depreciation and amortization), which at least 10 investors and analysts told Bloomberg they had never seen before. Without that and other adjustments, its leverage rises from 8.44 times earnings to 9.3 times. What this means for your stack: if your company has a credit agreement, read how its EBITDA definition treats projected cost savings before anyone proposes an AI add-back in a compliance certificate. Discuss with ChatGPT →
Watchdog rates the IRS security program not effective again (2 minute read)
The Treasury Inspector General for Tax Administration rated the IRS information security program not effective for fiscal 2026, meeting the bar in three of six framework areas but falling short on identifying, protecting against and detecting threats, with six of seven sampled systems carrying critical vulnerabilities left open past the 30-day deadline. Discuss with ChatGPT →
⚡ Quick Links
PCAOB chair says AI can do the science of an audit, not the art (2 min read)
Logothetis wants to understand AI before writing rules for it.
A new alliance open-sources a 1040 engine and a K-1 data standard (10 min read)
Filed and Crimson Tree bet free tax software pushes firms toward advisory.
SAP ships a prediction model built for spreadsheet-style business data (1 min read)
TabPFN targets cash forecasting and payment delays without model training.
Open models now carry most AI tokens, closed models most dollars (5 min read)
Open-weight share of one gateway's tokens went from 7% to 78%.
🔁 ICYMI
Worth a second look.
Ramp's September AI Index, built from card and bill pay data, put Anthropic at 43.8 percent of US businesses and OpenAI at 39.8 percent in August, with the effective price per million tokens down 41 percent from its March peak and frontier models slipping from 53 to 45 percent of token share as cheaper standard models took volume.
Why now: Published September 9 and never sent in LEDGER. It is the demand-side data behind this week's Harvey story, since cheaper standard models are exactly what vendors are moving to. Discuss with ChatGPT →
Ernie Tedeschi introduced the Stripe SaaS Index and found revenue growth across software businesses on Stripe reached 30 percent year over year through the early 2026 sell-off, led by larger firms and by newer companies selling AI tools, with equity prices back to roughly where they started.
Why now: Published September 10. Read it beside this week's Stripe Connect post, because this one covers revenue at existing SaaS companies while the new one covers how fast new platforms are forming. Discuss with ChatGPT →
Thomson Reuters built its own frontier model for about 40 million dollars (Issue 034, September 2)
Harvey took the same road this week for a different reason: Thomson Reuters wanted control, Harvey needed margin. Together they show where your content and AI vendors are heading. Discuss with ChatGPT →
📘 Definitions
AI terms that came up in this issue, in plain language.
Coordinator (or orchestrator) agent
An AI agent whose job is to break a request into pieces, hand each piece to another agent, and check and combine what comes back. The redesigned Claude Projects put a coordinator on top of parallel threads. The finance version is a close lead assigning reconciliations and reviewing them, rather than doing each one.
Thread (parallel agent session)
One independent line of AI work running alongside others, with its own context and its own copy of the files. Parallel threads finish faster, but the output of each one still needs a reviewer before it is combined.
Project memory
Context an AI tool keeps from one session to the next, such as your conventions, past decisions and where files live, so you stop repeating the same instructions. It is convenient, and it is also a record you should be able to read and correct.
Checkpoint
A point where automated AI work stops and waits for a person to approve before it continues. Claude's merged app lets users set them. In control terms, it is a required approval step placed inside the workflow instead of at the end.
Post-training
Further training applied to an existing base model to make it better at specific work, as Harvey did to the open-weight Kimi K3. It costs far less than building a model from scratch, which is why more vendors can afford their own.
Cost per completed task
What an AI system costs to finish one unit of real work, such as a closed reconciliation, rather than what it costs per token. A cheaper model that fails more often can cost more per completed task, so this is the number to ask vendors for.
Token share
The portion of all AI usage, counted in tokens, that goes to a given model or group of models. Open-weight models now carry most tokens on some gateways while closed models still capture most of the spending.
Inference provider and model router
An inference provider runs AI models for you and bills by usage. A model router sends each request to whichever model is cheapest or best for that task. Both matter to finance because they decide which model, at what price, sits behind a tool you bought on a flat seat price.
Supervised versus governed autonomy
Two levels of AI independence in Auditoria's survey. Under supervised autonomy a person reviews the AI's work before it counts. Under governed autonomy the AI acts on its own within set rules and limits, and people review the exceptions and the logs.
Tabular foundation model
An AI model pretrained to make predictions from rows and columns of business data, like an AR aging or a payments history, without building a custom model first. SAP's TabPFN is one, aimed at cash forecasting and payment delays.
AI add-back
Projected savings from AI that a company adds back to EBITDA in a lender or investor calculation, as Emeria did this week. It is new, it is disputed, and it belongs in the same skeptical bucket as any other forward-looking adjustment.