AI & Finance Weekly

The lead

CFO Dive · Deloitte CFO Signals · reported spring 2026

Finance is buying AI faster than it can prove it

The agentic wave has reached the close. The returns, mostly, haven’t shown up yet.

New agentic models keep landing — Anthropic shipped Claude Sonnet 5 on June 30, aimed squarely at professional work at scale, and Intuit says more than three million customers now run its AI agents over their books, payroll, and tax. But the desk-level scorecard is lopsided. In Deloitte’s most recent CFO Signals work, reported by CFO Dive, 87% of finance chiefs said AI would be very or extremely important to their operations this year — yet among those who have fully deployed it, only about 21% report tangible value so far.

That gap isn’t evidence AI doesn’t work. It’s a sign most teams bolted tools onto a 2020 operating model instead of redesigning who decides, how work flows, and where the sign-off sits. The teams reporting real returns tend to be the ones that changed the workflow, not just the software.

The concept · the value gap“Time saved” is the easiest number to produce and the least convincing to a board. Real ROI shows up elsewhere: faster decisions, errors caught before the books close, cash pulled forward. Pick those metrics on day one — or you’ll fund pilots you can’t defend at budget time.

Read the Deloitte read-out on CFO Dive →

The desks

Models

Anthropic · model release · Jun 30 2026

Claude Sonnet 5 aims at “professional work at scale”

Anthropic released Claude Sonnet 5 on June 30, positioning it for coding, agents, and professional work — the mid-tier model most teams actually run in production, now pitched near the top tier’s capability. For finance, the headline isn’t a benchmark; it’s the price-to-capability curve bending down.

The concept · inference costEvery time a model reads or writes, you pay per token. When a capable model gets cheaper to run, work that was too expensive to automate at scale — checking every transaction instead of a sample — can suddenly pencil out.

For your deskRe-price one control you sample today. A full-population check with a cheaper capable model may now cost less than the exception review it would replace.

Anthropic newsroom →

The finance stack

Intuit · finance software · updated spring 2026

Intuit’s agents now do the books, payroll, and tax

Intuit says its Accounting, Payroll, Sales Tax, and Business Tax agents have crossed three million customers since launching in August 2025, and that 74% of surveyed users report a clearer picture of their financial health. The pitch is “done-for-you” categorization and reconciliation, with the human moving to review.

The concept · an “agent”Not a chatbot that answers — software that takes actions toward a goal: reads a bank feed, proposes a categorization, and posts it. The real question is what it’s allowed to do without asking first.

For your deskTreat agent output as a preparer, not a reviewer. If the agent categorizes and posts, someone independent still owns the review control. “Done-for-you” is not “signed-off-for-you.”

Intuit →

Money & Rules

PCAOB · rulemaking · Jun 23 2026

The PCAOB moves to trim its own quality-control rule

The audit regulator opened a comment period on June 23 on narrowing QC 1000, the firm-level quality-control standard adopted in 2024. The proposals would drop the external quality-control function for the largest firms, scrap a “design-only” requirement for firms that don’t perform PCAOB-standard work, and shorten documentation retention from seven years to five. Chairman Demetrios Logothetis has framed the changes as making the rule practicable without moving its December 15, 2026 effective date — and separately wants the board to lean on automation and AI in its own inspections. Comments are due July 9.

The concept · quality managementQC 1000 governs the system a firm uses to catch problems before they reach an audit — the same plumbing that will increasingly run AI tools, and where regulators keep insisting a human stays accountable for the result.

For your deskIf your auditor is a smaller or non-issuer firm, the design-only rollback is the piece to track. The comment window closes Jul 9; the Dec 15 effective date still stands.

PCAOB news release →

AI, demystified

What people mean by an AI “agent”

The word is on every finance-software slide this quarter. Strip the marketing and an agent is a loop with three moves.

STEP 1

Perceive

It reads your data — an invoice, a bank line, the general ledger.

STEP 2

Decide

It plans a step against a goal and your rules: “this looks like a duplicate of last week’s payment.”

STEP 3

Act & log

It does something and records it — holds the payment, flags the invoice, leaves a trail a human can inspect.

The controls question isn’t “is it smart?” It’s which of those actions happens without a human — and can you see every one it took?

Hype check

Exaggerated

“Agentic AI can close the books on its own.”

The tools are real and moving fast, but today’s finance deployments keep a human in the loop by design. Deloitte’s own read-out has only about 14% of firms fully integrating AI agents into finance and roughly 21% of full deployers seeing tangible value; audit regulators on both sides of the Atlantic hold that a person — not the model — remains accountable for the result. “Assisted close,” yes. “Autonomous close,” not yet.

Sources: Deloitte via CFO Dive; PCAOB / FRC audit guidance.

Do this week

15 min
Name one high-friction close task worth piloting — a reconciliation, variance commentary, or spend categorization. Pick by pain, not by hype.
30 min
Inventory the AI already sitting in your ERP and FP&A tools before buying anything new. Most stacks ship features nobody switched on.
20 min
Write the guardrail: one paragraph on what an agent may do unattended, what needs a human sign-off, and where its actions get logged.

Prompt of the week

You are my FP&A analyst. Below is our budget-vs-actual for [cost center], [period]. Walk the top 5 variances by dollar amount. For each, give: – the line item – the $ and % variance – the single most likely driver you can support FROM THE DATA If a driver isn't in the data, say "not determinable from provided data" — do not guess. Flag anything that looks like a mispost. End with 3 questions you'd ask the budget owner. [paste data]

Models good practice: a clear role, your real data, and an explicit instruction not to invent causes the numbers don’t support.

Off the record

Accounting Today · opinion · early June 2026

The person building the AI is the one warning you about it

In an early-June op-ed, the founder of an AI audit platform — someone whose business depends on selling the technology — argued that the fastest way to wreck the profession is to take humans out of the loop. His worry isn’t a dramatic failure; it’s a quiet one. Generative AI produces a polished, authoritative audit memo whether or not the underlying data is right, with none of the uncertainty a junior would have voiced out loud.

Real talkIf young accountants only ever supervise autonomous systems, who develops the judgment to eventually sign the opinion — or catch what the machine missed?

Accounting Today →

Personality of the week

Marie Myers

CFO, Hewlett Packard Enterprise

One of the clearest examples of a finance chief running agents on a core process rather than a pilot. Her team scaled an internal agentic tool — built with Deloitte — and pointed it at the weekly operations review she calls the company’s heartbeat, plus forecasting and receivables, as CFO Dive reported earlier this year.

Why this week: The lead asks whether the agentic wave is paying off. Myers is the version of that bet that’s furthest along — worth watching as the proof-or-bubble question resolves.

CFO Dive →

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

Three things to know

By the numbers