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.

Ramp gives finance teams one dashboard for AI spend (3 minute read)

Ramp launched a tool that pulls AI spending from Anthropic, OpenAI, Gemini and Cursor into a single view, broken down by provider, model, user, team or API key, with spending limits and anomaly alerts. Ramp says the average business using it finds savings worth 12% of monthly AI spend, and it is free to start whether or not you are a Ramp customer. What this means for your stack: if AI costs are hitting your books through engineering credit cards and provider invoices you cannot easily read, this gives you an owner and a number per team before the next budget cycle, and the free tier means you can test the visibility question without a procurement conversation.

Stripe in talks to buy AI model marketplace OpenRouter (3 minute read, Wall Street Journal, paywalled)

Stripe is reportedly in talks to acquire OpenRouter for around $10 billion, up from a $1.3 billion valuation in May, in a deal that would give a payments company control of the layer that routes AI requests across roughly 400 models from about 70 providers. OpenRouter charges a percentage on top of the underlying model cost, which makes it look structurally more like a payments business than an AI one. What this means for your stack: model routing is becoming a billable layer of its own, so if your engineering team is already routing between models, expect that spend to start showing up as a separate vendor line rather than buried inside your Anthropic and OpenAI invoices. Free coverage of the same story at Axios.

BlackLine ships multi-agent AI for reconciliation prep (3 minute read)

BlackLine made Verity Prepare generally available, a system where several AI agents work together to read supporting documents, match transactions, flag reconciling items, write variance explanations and assemble a reconciliation package for a human to sign off. Every recommendation carries its reasoning, a confidence score and an audit trail, and BlackLine says early adopters cut manual preparation time by up to 92%. What this means for your stack: the audit trail and confidence scoring are the part worth evaluating, because your external auditors will ask how an AI-prepared reconciliation got to its answer long before they ask how fast it got there.

Visa and Airwallex target 42-day freight payment cycles (3 minute read)

Visa and Airwallex are partnering to build payments and working capital directly into freight and shipping software, in an industry where the companies say invoices take an average of 42 days to get paid and processing and administrative costs run to nearly a fifth of total transportation costs. The partnership was announced at the intent stage with no named products or launch dates yet.

Survey finds only 7% of finance leaders put governance first (2 minute read)

Avalara surveyed more than 1,500 finance chiefs and senior finance leaders in the US, UK, India and Australia and found 92% under career pressure to prove that agentic AI is paying off, with 70% describing their focus as entirely or mostly on speed and only 7% saying governance is being prioritised over it. Nearly 90% reported some return, but only 38% called it return at scale, and 36% said nobody is specifically responsible for understanding how their AI agents work. What this means for your stack: 23% of respondents could not name who would be accountable for a significant AI error, so if you cannot answer that question for your own finance team in one sentence, that is the cheapest gap to close this quarter.

PCAOB opens comment on its 2026 to 2030 strategic plan (2 minute read)

The PCAOB (the board that oversees audits of public companies) is taking comments through September 4 on a draft five year plan built on six goals, including modernising standard setting, modernising inspections and registration, sharpening enforcement on conduct that harms investors, and modernising oversight through technology and data. What this means for your stack: the technology and data goal is the one to read closely, since how the PCAOB expects to inspect AI-assisted audit work will shape what your external auditors ask you to produce.

Moonshot releases Kimi K3 weights for anyone to download (4 min) Likely the largest open-weight model ever published, and close to frontier quality.

Anthropic releases Claude Opus 5 (2 min) Fourth Claude 5 model in under two months, at roughly half the price.

OpenAI launches GPT-5.6 in three tiers (3 min) Luna, Terra and Sol, with Sol claimed 54% more token efficient on code. Note, actual release from early July.

Worth a second look

EU pushes its high-risk AI deadline to December 2027 The original deadline was this Sunday, August 2, so if you shelved an EU credit scoring or hiring tool over it, you now have about sixteen more months and the same obligations waiting at the end.

Finance and tech payrolls are falling 28,000 a month The clearest data yet that AI is showing up in employment numbers, and finance is exposed because administrative roles are a bigger share of its headcount than any other industry.

AI has made fake vendors much easier to build A useful counterweight to an issue full of AI spending: the same tools are lowering the cost of committing vendor fraud against you.

Terms that turned up in this issue.

Token. The unit AI providers bill you in. Roughly three quarters of a word. You pay separately for tokens going in and tokens coming out, which is why a long document plus a short answer can still be expensive.

Context window. How much text a model can hold in view at once, measured in tokens. Claude Opus 5 and Kimi K3 both advertise 1 million tokens, roughly 750,000 words, or every contract in a mid-size company's repository at the same time.

Model routing. Automatically sending each request to whichever model is cheapest or best for that specific task, instead of sending everything to your most expensive one. This is OpenRouter's entire business, and it is why Stripe is reportedly paying $10 billion for it.

Prompt caching. Storing the unchanging part of a repeated request, a policy document or a chart of accounts, so you do not pay to send it again every time. This is the setting one Ramp customer found was costing them $10,000 a month by not using leveraging prompt caching.

AI agent. A model that carries out a multi step task on its own rather than just answering a question. It decides what to do next, calls other tools, and reports back. The practical difference from a chatbot is that an agent takes actions.

Agentic AI. The umbrella term for using agents in production. When a vendor says agentic, ask two questions: what can it do without a human approving, and what does it write down about what it did.

Multi-agent system. Several specialised agents working on one process, handing off to each other. BlackLine's Verity Prepare is one: separate agents ingest documents, match transactions, and write variance explanations, then assemble the package for a human.

Open-weight model. A model whose parameters are published for anyone to download and run on their own hardware. Kimi K3 is the current example. Practically, it means the option to run AI on finance data that never leaves your environment.

Hallucination. When a model states something false with complete confidence. It is not a bug that gets patched, it is a property of how these systems generate text, which is why every serious finance deployment keeps a human signature at the end. In a finance context this looks like an invented vendor name, a citation to a standard that does not exist, or a number that is plausible and wrong.

Markdown. A plain text formatting shorthand where you type asterisks for bold and hashes for headings instead of clicking buttons. AI tools output it constantly. If you have ever pasted an AI answer and gotten a wall of stray asterisks, that was Markdown that did not render.