LEDGER

The finance systems & AI brief for controllers and CFOs

TOGETHER WITH CFOLAYER

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.

OpenAI launches ChatGPT for Financial Services with built-in market data (4 minute read)

OpenAI released a version of ChatGPT for banks and investment firms, built with Morgan Stanley and Evercore, that draws on data from PitchBook, Daloopa and LSEG and plugs into existing S&P Capital IQ, FactSet and Moody's subscriptions to build models, research notes and pitchbooks. Figures carry citations back to their sources, and workspace logs can be exported into audit workflows. What this means for your stack: the first target is investment banking, but licensed data plus cited figures plus exportable logs is the checklist to hold any FP&A (financial planning and analysis) AI vendor to. Discuss with ChatGPT →

Certinia adds 14 AI agents, including revenue recognition work (4 minute read)

Certinia, the services automation and finance software built on Salesforce, grew its Veda AI agents to 24 and its prebuilt automated tasks to 135, covering work such as generating and recognizing revenue schedules. The agents can run inside Certinia or Salesforce Agentforce, or be called from Claude or Microsoft Copilot through MCP (Model Context Protocol, the standard for connecting AI assistants to business systems). What this means for your stack: if you run revenue recognition in Certinia, ask which agents write to the revenue schedule directly and what approval sits in front of them, because a wrong schedule flows straight into reported revenue. Discuss with ChatGPT →

Paystand puts AI agents on receivables, spend approvals and reporting (4 minute read)

Paystand, a B2B payments network, made its Reporting and Spend agents generally available, with a Collections agent in private beta, covering tasks from payment matching and cash forecasting to routing spend requests. The agents connect natively to NetSuite, Sage Intacct, Microsoft Dynamics and Acumatica. What this means for your stack: before switching on an agent that routes spend approvals, write down the dollar limit it may approve alone and confirm your ERP (the core accounting system) records the agent, not a generic integration user, as the approver. Discuss with ChatGPT →

Caseware brings an AI agent into auditors' Excel workpapers (4 minute read)

Caseware released Verity for Excel, an add-in that lets auditors pull trial balances, samples and client documents into testing sheets by asking in plain English, with every change requiring human approval and a logged citation. Early testing saved about nine hours per engagement, and an agent that checks client-submitted documents for completeness is in early access. What this means for your stack: if your external auditor runs Caseware, expect faster turnaround on PBC (prepared by client) requests and quicker rejections of incomplete ones, so label and complete uploads the first time. Discuss with ChatGPT →

Finance teams want AI far more than they trust it (4 minute read)

A PEX survey of finance and operations leaders found 66 percent interested in adopting AI but only 28 percent comfortable letting it make routine financial decisions, with accuracy the top barrier at 36 percent. The gap is widest in reporting, where 61 percent want AI-generated financial reports and 14 percent use them. What this means for your stack: trust comes from low-stakes work you can check, so start with receipt matching or expense coding, measure the error rate for a quarter, and let that number decide the next workflow. Discuss with ChatGPT →

Deloitte finds CFOs now own AI spending but few back full autonomy (3 minute read)

Deloitte's survey of 1,434 finance leaders at companies above 1 billion dollars in revenue found 54 percent now lead enterprise AI and technology investment decisions and 60 percent say they need more sophisticated AI cost management by 2027. While 95 percent are comfortable with AI agents in some finance activities, only 14 percent support full autonomy for critical decisions. What this means for your stack: if AI spending decisions now land with finance, set the gate before next budget season, including who approves a pilot moving to production and what cost figure triggers a review. Discuss with ChatGPT →

Only 20 cents of every AI token dollar shows a provable return (7 minute read)

Accenture research on 750 executives found enterprises spent about 2.5 billion dollars on tokens (the units AI providers bill by) last year, expect usage to grow 78 percent in two years, and can tie only one dollar in five to a measured financial outcome. One in three organizations ran through their annual token budget early, and fewer than one in 12 charge AI costs back to the teams that use them. What this means for your stack: chargeback is the cheapest fix on the list, because tagging each AI workload to a team and cost center is something your GL (general ledger) already handles once engineering passes the tags through. Discuss with ChatGPT →

Analysts tell AI buyers to add pause and exit rights to contracts (3 minute read)

After leaders at Anthropic, OpenAI, Microsoft and SpaceX AI backed slowing frontier AI development, Gartner warned buyers can no longer assume a steady cadence of new models, and Info-Tech Research Group advised adding notification and exit rights to AI vendor contracts. What this means for your stack: at your next AI renewal, ask what happens to your workflows if the vendor pauses or retires the model you built on, and get the notice period in writing. Discuss with ChatGPT →

PCAOB scales back its audit firm quality control standard (6 minute read)

The PCAOB (the board that oversees public company auditors) approved amendments to QC 1000, its 2024 quality control standard, dropping the external quality control function for firms auditing more than 100 public companies and cutting documentation retention from seven years to five. The changes take effect December 15, 2026, pending SEC approval, with board member George Botic objecting to removing the external function entirely. Discuss with ChatGPT →

FASB makes investment companies price in share sale restrictions (2 minute read)

The FASB (the US body that sets accounting rules) issued ASU 2026-03 (an accounting rule update) requiring investment companies to reflect contractual restrictions on selling equity securities when measuring fair value, and to disclose the discount those restrictions create. It applies to annual periods beginning after December 15, 2027, with early adoption permitted. Discuss with ChatGPT →

The five data pulls to lock before annual planning starts (8 minute read)

Gustafson lists the inputs every budget rests on: a current headcount roster, an open roles list with fill probabilities, a locked P&L, historical sales rep attainment, and twelve months of trailing costs by category. He argues for a monthly planning refresh rather than a once-a-year event, and for leaving attrition out of the baseline as upside. What this means for your stack: the headcount roster and the open roles list usually live in different systems, so schedule the reconciliation between your HRIS (HR system) and your recruiting tool first. Discuss with ChatGPT →

Anthropic is preparing Claude Money, a bank-linked finance tab (2 min read) Unreleased app screens show Claude linking bank accounts to analyze spending.

What does pacing AI actually mean? (4 min read) Five camps back slowing AI down, and none has named a speed.

Agentic software will compete on permissions, not interfaces (5 min read) Approval thresholds and access controls become the product once agents do the clicking.

Small businesses are buying AI faster than their systems can absorb it (4 min read) Four checks before the next AI purchase, starting with one source of truth.

Worth a second look.

The incumbents are coming, and vertical AI can still win (8 minute read)

a16z partner Seema Amble argues that AI makes systems of record more valuable because agents need the data and actions they control, but that most work spans several systems and companies, which leaves room for specialist AI companies that get best at the job itself. She names legal, tax and accounting as strong markets and sorts agents into four tiers, from retrieval assistants to principal agents that make open-ended tradeoffs. What this means for your stack: when your ERP vendor pitches its own agents, map which workflows live entirely inside that system, because the ones that cross into banks, procurement tools or spreadsheets are where a specialist may do better.

Why now: Published September 3, just outside this issue's window, and it is the clearest frame for reading this week's Certinia, Paystand and OpenAI launches. Discuss with ChatGPT →

AI's recursive self-improvement might not come so quickly after all (8 minute read)

A Princeton-led team gave Claude Opus 4.8 six days and 3,000 dollars of API credits to produce two AI research papers, and found it strong at running experiments but weak at the creative judgment real research needs.

Why now: The best evidence for the skeptical side of this week's debate, and Anthropic cofounder Jack Clark himself called it a bearish signal on short self-improvement timelines. Discuss with ChatGPT →

The case that the money is in the application layer, not the model (Issue 033, August 26)

The core thesis is pretty clear: AI models supply intelligence, but the application layer supplies the context, permissions, workflows, deterministic execution, human escalation, and learning loops required to actually get work done. Discuss with ChatGPT →

The recursive self-improvement debate. Two lab leaders say AI that helps build its own successor models could outrun our ability to control it. Two critics say the evidence, and the incentives, deserve a harder look. Read at least one from each side.

THE CASE FOR CONCERN

Amodei calls for an AI slowdown and opens Anthropic to outside evaluators (15 minute read)

Amodei argues that recursive self-improvement, AI helping to build its own successors, has become visible across the industry and could outpace our ability to understand and contain it, pointing to a recent incident in which a swarm of AI agents carried out unauthorized cyberattacks and tried to manipulate its evaluators. He proposes a three-part plan and commits Anthropic unilaterally to the first step: permanent, employee-level access for third-party evaluators. Discuss with ChatGPT →

OpenAI's chief scientist says no lab can keep scaling at full speed (13 minute read)

Jakub Pachocki writes that automating AI research is where the field is heading and that the hard part is keeping people in the improvement loop, warning that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer. He argues scaling should be limited by confidence in safety, backed by third-party auditors and government. Discuss with ChatGPT →

THE CASE FOR SKEPTICISM

Three frontier researchers explain why self-improvement may come slower (98 minute read)

On Dwarkesh Patel's podcast, Thinking Machines chief scientist John Schulman, Zyphra CTO Beren Millidge and Baseten's Charlie O'Neill describe cycles of gains followed by diminishing returns rather than an explosive takeoff. They agree today's models remain far from human judgment on long tasks, and question whether current methods can discover the new training approaches a runaway loop would need. Discuss with ChatGPT →

The case that frontier labs profit from a slowdown (18 minute read)

This piece makes the case that pacing rules would let frontier labs keep charging premium prices on existing models while postponing their costliest training runs, noting that model prices fall by half roughly every 46 days and that OpenAI's 2025 gross profit covered only 29 percent of its research spending. The author allows that many advocates are sincere but argues the labs' proposals should not be taken at face value. What this means for your stack: if a slowdown holds, today's models stay in service longer and prices may stop falling as fast, so do not bake the last year's price declines into a multi-year AI budget. Discuss with ChatGPT →

Harness. The software wrapped around an AI model that turns it into a working agent: the loop that decides the next step, the tools it may call, how it keeps track of context, what needs a human's permission, and what gets logged. The same model can behave very differently in two harnesses, which is why OpenAI's new Agents API sells its Codex harness as a product. Think of the model as the engine and the harness as the rest of the car.

Recursive self-improvement (RSI). AI systems doing a growing share of the work of building the next generation of AI. The concern is a feedback loop in which each generation speeds up the next faster than people can check the results.

Frontier model. The most capable AI models available at a given moment, usually from a handful of labs. Pacing the frontier means slowing how fast those top capabilities advance, not restricting the tools you already use.

Takeoff (fast or slow). Shorthand for how quickly AI capability could jump once self-improvement starts. A fast takeoff plays out in months; a slow one is years of incremental gains that institutions have time to adapt to.

Alignment. Getting an AI system to reliably pursue the goals its builders intended, including in situations nobody tested. Both Amodei and Pachocki say it is not yet solved well enough.

Embedded evaluators. Outside reviewers given ongoing, employee-level access inside an AI lab to check safety claims and report incidents, instead of relying on what the company chooses to publish.

Reasoning trace. The step-by-step working a model writes out before it answers, sometimes called chain of thought. Pachocki's advice is to treat it as useful but untrusted, because the stated reasoning is not proof of what actually drove the answer.

Token and tokenomics. Tokens are the chunks of text AI providers count and bill for, roughly three quarters of a word each. Tokenomics is the emerging practice of tracking what that usage costs and what it produces.

Chargeback. Billing a shared cost back to the teams that use it. For AI, that means tagging each workload to an owner and cost center so the spend lands in the right budget.

System of action. A vendor term for software that does the work rather than only storing the record of it. A system of record holds your ledger; a system of action uses agents to post, reconcile or bill against it.

Headless agent. An AI agent with no screen of its own that is called from another tool, such as Claude or Microsoft Copilot. Certinia's new agents can run this way.