AI & Finance Weekly Issue 36 · Tuesday, Sep 08 2026

Get fluent in AI — and in what it means for your numbers.

This week finance stopped being the department that buys the software and started being the department that writes it. Every figure below links to its source.

The Close The Finance Stack Money & Rules Models & Policy Off the Record

Three things to know

01A finance director at OpenAI used a coding agent to rebuild one of the ugliest parts of the monthly close — about five days of work became about five hours.
02Nvidia agreed to buy Hugging Face, the warehouse where most of the world's open AI models live, for $12.93 billion.
03KPMG became the first Big Four firm to get an agentic AI tool independently certified — and said plainly that a certificate is not immunity.

By the numbers

5 hrs

Down from about five days: OpenAI's compute-cost close analysis after a coding agent rebuilt it

CFO Dive, Sep 1

35%

Finance leaders who say they can confidently gauge AI's return, in a Protiviti global survey

CFO Dive, Aug 27

$12.93B

What Nvidia agreed to pay for Hugging Face, per Jensen Huang's own announcement

NVIDIA, Sep 3

8 pts

Revenue-growth gap between firms with the most and least specific AI descriptions in their 10-Ks

CFO Dive, Aug 18

The close

CFO Dive · the close · Sep 1

A finance director wrote code, and five days of close work became five hours

Kyle Kober, director of product finance at OpenAI, went looking for a way to automate the worst workflow in his department and ended up building software. The job was reconciling product-usage data against accounting records for the company's computing costs, then turning that into analysis and reporting — one of the most cumbersome pieces of the monthly close.

He used Codex, OpenAI's coding agent, to rebuild it. CFO Dive reports the compute analysis now takes roughly five hours instead of roughly five days. Kober's framing was blunt: there is now an option to just do it yourself, rather than queue behind IT. The project sits inside a wider push by CFO Sarah Friar toward what she calls an AI-native finance function.

It is not only an OpenAI story. A June report from Boston Consulting Group, cited in the same piece, argues coding agents let finance teams build their own tools for analysis, matching and anomaly detection without waiting in a development queue — while warning that leaders need clear guardrails, auditability among them.

The concept — AI coding agent

A coding agent is a very fast junior developer who works from a plain-English description: you describe the task, it writes and runs the code. What it does not have is your context. It will happily produce a script that executes cleanly, looks professional, and computes the wrong thing — the same way a beautifully formatted spreadsheet can quietly reference the wrong column.

For your desk

A script your team built with an agent that touches the ledger is a spreadsheet with longer reach, and it inherits every spreadsheet control question: who reviewed the logic, who can change it, which version ran, and what evidence exists that it ran against the right data for the right period. If it feeds a number in the close, it belongs in your control documentation — not in someone's personal folder.

Read it at CFO Dive

The Finance Stack

CFO Dive · revenue recognition · Sep 4

Asana started pricing AI by the drink, and the revenue line moved before the customers did

On last week's earnings call, Asana CFO Aziz Megji said the company's shift toward consumption-based AI pricing would contribute a $1.2 million revenue headwind and press gross margins by about 150 basis points in the second half of the year. From mid-September, a base level of AI tools ships inside its Agentic Work Management tiers without a change in tier pricing.

The product news is ordinary. The reporting consequence is not: the same customers, paying similar amounts, land differently on the income statement because the shape of the promise changed.

The concept — consumption pricing

Per-seat pricing charges for access; consumption pricing charges for use — tokens processed, actions taken, documents handled. Cost then tracks activity rather than headcount, which is precisely the variable a headcount-based budget cannot see.

For your desk

If your company sells anything AI-metered, the ASC 606 questions arrive before the cash does. What is the performance obligation once AI is bundled into an existing tier, does bundling reset standalone selling price across the whole arrangement, and does usage-based consideration get recognised as it is consumed rather than ratably? On the buy side, the mirror question: is your AI spend an amount you can accrue, or a meter nobody on your team can currently read?

Read it at CFO Dive

Money & Rules

CFO Dive · agent governance · Sep 1

KPMG had an AI agent audited by an outsider — and still won't call it safe

A KPMG-built agent called aIQ Capture has been independently tested and certified under AIUC-1, a standard for AI agents issued by the Artificial Intelligence Underwriting Company. It is the first such certification for a Big Four firm. The agent went through more than 900 technical tests covering hallucinations, content safety, high-risk domain interactions and prompt-injection attacks, with no critical or major vulnerabilities identified.

The most useful line came from KPMG's own responsible-AI principal, Aisha Tahirkheli, who said the certification is evidence of rigour but does not eliminate risk or grant immunity from it, and that keeping confidence requires continued monitoring, testing and human accountability. Gartner's Alex Levine made a parallel argument in an August Q&A: use your first finance agent to build governance and oversight, not to chase a quick return.

The concept — certification scope

An AI certification works like a SOC 1 report. It attests that a named system, in a named configuration, passed a defined set of procedures on a defined date. It says nothing about how you deploy it, what access you grant it, or what happens after the report period — those are your controls, not theirs.

For your desk

When a vendor waves a certification at you, ask the three questions you would ask of any service-organisation report: what was in scope, over what period, and which complementary controls fall on us. A certificate covering one agent is not a certificate covering the platform.

Read it at CFO Dive

Models & Policy

NVIDIA · M&A · Sep 3

Nvidia is buying the world's open-model warehouse for $12.93 billion

Jensen Huang announced that Nvidia has agreed to acquire Hugging Face for $12,930,300,000. The platform is where more than 18 million developers, researchers and creators share over three million models and half a million datasets, and where more than 200,000 companies go to find, evaluate and deploy AI. Nvidia said it will remain an open platform for the whole ecosystem. Bloomberg reported the price includes an equity-based retention programme of as much as $1 billion for Hugging Face employees who join.

The concept — open-weight models

With a closed model you rent capability through an API and the provider holds the parameters. With an open-weight model you download the parameters and run them on infrastructure you control. That distinction decides where your data physically goes — which is why it shows up in vendor risk reviews long before it shows up in a budget.

For your desk

The retention programme is the line a technical accountant will circle. Equity granted to a target's employees that requires future service generally raises the question of whether it is consideration transferred or post-combination compensation expense — a distinction that moves goodwill and moves next year's operating expense in opposite directions. Nvidia has not said how it will account for it. If you ever sit on the buy side of a deal like this, that analysis starts at signing, not at close.

Read the announcement at NVIDIA

AI, demystified

Prompt injection, in three steps

Prompt injection was one of the attack types in KPMG's 900-test battery. It is worth understanding because it is the failure mode with no software patch — it comes from how these models read.

Step 1

Your agent reads something you didn't write

A vendor invoice, an emailed statement, a supplier's web page, a PDF someone dropped in a shared folder.

Step 2

The model can't tell data from orders

Everything arrives as one stream of text. There is no hard boundary between "content to process" and "instruction to follow."

Step 3

Buried text tells it to act

A line in that document instructs the agent — and it acts with whatever system access you granted it.

The defence is not cleverer prompting. It is the same principle you already apply to a junior with system access: least privilege, and a human approval on anything that writes. An agent that can only read your AP inbox has a bad day. An agent that can also release a payment has an incident.

Hype check

"Our AI agent is independently certified, so it's safe to run in the control environment."

Busted

The certification is real and it is not nothing — but it does not carry the weight this claim puts on it. It records that one named agent, in one configuration, passed a defined battery of tests on a defined date. KPMG's own responsible-AI principal said as much: evidence of rigour, not elimination of or immunity from risk, and confidence that depends on continued monitoring, testing and human accountability.

"Certified" is a statement about a test result. "Safe in our control environment" is a statement about your data, your access grants and your review steps — none of which the certifier saw. Treat the certificate as an input to your assessment, not a replacement for it.

Read the certification announcement at KPMG

Do this week

20 minList every AI-assisted step that touches a number in your close. Steps, not tools — the tool inventory hides where judgment actually got delegated.
30 minPick the riskiest one and write the two sentences an auditor will ask for: what the model did, and who checked it before it posted.
15 minAsk your largest software vendor whether anything on your contract is metered by AI usage, and where that meter can be read mid-month.

Prompt of the week

You wrote the logic below for a month-end
reconciliation. Act as the reviewer, not the author.

[paste the script, formula, or query]

Context: [source system, period, what the output feeds]

Do three things, in order:

1. Restate in plain English what this actually
   computes, step by step. Do not describe what it
   was meant to do.

2. List every input assumption it makes that I have
   not stated: date boundaries, sign conventions,
   currency, duplicates, nulls, period cutoff.

3. Give me three specific cases where it returns a
   clean-looking number that is wrong, and the test
   I would run to catch each one.

Do not fix anything yet. I want the failure list first.

Why it works: it splits author from reviewer, forces a plain-English restatement before any defence of the logic, and demands failure modes before fixes — which is the order a preparer-reviewer control runs in anyway.

Number of the week

Gartner's forecast for the share of enterprise software engineers using AI code assistants by 2028 — up from under 14% in early 2024. Finance is roughly two years behind that curve, and closing.

Gartner, cited by CFO Dive, Sep 1

Off the record

CFO Dive · opinion · Sep 8

Your AI budget is built on a headcount model that no longer describes the spend

Writing in CFO Dive today, Turnstile chief executive Jordan Zamir made an argument most controllers will recognise before they finish the first paragraph. For decades software budgeting followed one formula: estimate headcount, map people to licences, forecast the spend. More salespeople meant more CRM seats. Costs grew with the org chart.

His point is that AI breaks the mapping, and that finance is still reaching for the old playbook because it is the one on the shelf.

Real talk

The uncomfortable version: the number you cannot forecast is the number you end up explaining in a variance meeting. If a line item's real driver is usage and you budget it off headcount, that is not conservatism — it is a miss in a direction nobody has measured yet. Worth finding out which of your AI line items are already metered before someone else finds out for you.

Read the opinion piece at CFO Dive

Personality of the week

Sarah Friar

Chief Financial Officer, OpenAI

In an essay on OpenAI's site last month, Friar set out two ambitions for her function: a zero-day close and a forecast that updates continuously. The description of the target state is the part finance people should read — approved spending plans, general-ledger actuals, purchase orders, accruals and transaction detail held in one continuously reconciled view, with every variance traceable to the underlying activity and AI drafting the first explanation.

What makes it credible is the limit she draws around it. The close, she wrote, does not disappear. What disappears is the scramble to reconstruct the picture after the period ends. Finance still validates the numbers, applies judgment and owns the final sign-off.

Why this week

Her essay is the spine of this week's lead. The Codex work inside her own team is that ambition showing up as ordinary Tuesday-afternoon labour — which is usually where you find out whether a vision survives contact with a reconciliation.

Read her essay at OpenAI

Until next Monday — keep a name next to every number.

AI & Finance Weekly is a free brief for controllers, CFOs and the people who actually close the books. No sponsors, no vendor placements, no affiliate links.

On sources: every story in this issue links to the primary party's own page or a named news outlet — CFO Dive, NVIDIA, KPMG, OpenAI. Figures are stated as the linked source states them. The Carnegie Mellon and Larridin work cited in the stat row is a working paper and has not completed peer review; treat it as one data point. Where a source describes something as reported or forecast rather than settled, so do we. Verify before you cite.

Issue 36 · Tuesday, Sep 08 2026 · sources verified Sep 2026