What's new:

Gartner is warning CFOs that the business cases used to approve conventional technology investments do not work well for AI. Traditional software follows relatively predictable licensing, maintenance, renewal and depreciation patterns. AI costs fluctuate with usage, compute, inference volume, data requirements and model performance — and the investment may require continuous reinvestment simply to maintain its value.

How it works:

Instead of relying on one deterministic ROI forecast, Gartner recommends modeling several value scenarios based on variables such as adoption, accuracy thresholds, error rates, usage levels and performance improvements. The same variability applies to risk. Because AI crosses data, technology, operations, compliance and workforce decisions, Gartner argues that it cannot be governed solely by IT or a periodic enterprise risk assessment.

Behind the news:

The report connects three issues that CFOs often evaluate separately: investment economics, enterprise risk, and talent. Gartner says only 18.2% of tasks associated with finance leadership are susceptible to automation, compared with 92.9% of tasks related to enabling finance data and technology. Its recommended response is to develop business and finance acumen earlier while adding capabilities in AI-assisted work, coding, business technology, and algorithmic bias detection.

Why it matters:

Finance leaders are being asked to approve AI spending while the cost, output quality, adoption rate, and control environment are still moving. A conventional business case can create false precision: a fixed cost estimate, a stable labor-savings assumption, and a straight path to payback. It can also reward automation of entry-level work without accounting for the future cost of developing reviewers, controllers and finance leaders once the work that trained them has disappeared

Our thinking:

Variable economics should not become an excuse to exempt AI from financial discipline, but rather it should change the discipline. An AI business case needs ranges, thresholds and explicit triggers: what accuracy makes the use case viable, what level of adoption unlocks the benefit, how much review work remains, when usage costs erase the savings, and what evidence justifies another round of investment. The strongest CFO will not be the one who demands a perfectly certain ROI before approving anything, or the one who funds AI on faith. It will be the one who builds a financial model capable of learning as quickly as the technology does.

What's new:

On Aug. 4, the Ninth Circuit vacated Amazon’s injunction against Perplexity’s Comet browser. Judge Milan Smith, writing for the panel, held that Perplexity does not access Amazon’s servers — the user does. It’s the third U.S. legal development since January pointing liability toward whoever operates the agent.

How it works:

California’s AB 316 took effect Jan. 1: A defendant who “developed, modified, or used” AI cannot argue it autonomously caused the harm. Causation, foreseeability and comparative-fault defenses remain. Executive Order 14409, signed June 2, directs the Attorney General to prioritize Computer Fraud and Abuse Act enforcement against anyone who uses AI for unauthorized access.

Behind the news:

Earlier this year Andrew Bird, head of AI at Melbourne-based Affinda, asked a Claude-based agent to book a gym class. The agent found the booking vendor’s API required no authorization to cancel reservations and cancelled a stranger’s. ABC Australia later called it the country’s first known autonomous AI cyberattack. Bird disclosed it himself.

Why it matters:

Agents now touch client portals, bank feeds, tax software and payroll integrations under firm credentials. On the Ninth Circuit’s reading, the firm did the accessing. The exposure sits in E&O and cyber policies written before agents existed, and in engagement letters silent on agent scope. Firms that issue task-scoped credentials and retain agent action logs can answer “who did this” with evidence rather than argument.

Our thinking:

Prosecution is not the likely mechanism. Civil claims, terms-of-service enforcement, insurance underwriting and employer policy all move on clean facts, and “firm directed an agent, agent reached an unauthorized endpoint, third party lost something” is clean. Congressional letters to lab CEOs do not change who the law currently names. The practical step is smaller than a policy rewrite: put agent scope and log retention into engagement letters, and raise agent use with the carrier at this renewal rather than after a claim.

What's new:

In April, Anthropic removed bundled tokens from its enterprise per-seat pricing. The advertised seat price fell from the $40-$200 range to a flat $20, with every use billed on top. The Register reported the change on April 16. Usage cost moved from the vendor’s side of the contract to the buyer’s.

How it works:

Radical CEO Pascal Finette’s argument: A five-person shop gets near-unlimited use from a $100–$200 consumer subscription. A large firm pays metered rates but amortizes a data platform across thousands of seats. The firm in between carries IT, security review and SOC 2 obligations without a platform team, so it buys the same commercial tools its competitors buy.

Behind the news:

Kasa CEO Roman Pedan made a parallel barbell argument in Fortune on Aug.14 — platforms and specialists both gaining while firms in between carry overhead without scale. Finette names the condition that would break his own claim: Consumer-tier AI is subsidized, and if that pricing normalizes toward compute cost, the small-firm advantage disappears.

Why it matters:

The 50 to 1,000 employee band Finette calls worst-positioned is where regional and mid-size accounting firms sit. Metered pricing turns AI from a fixed seat cost into a variable one that moves with busy season, breaking budgets built on per-seat assumptions. Firms that track token cost by engagement can put it into pricing and realization. Firms that do not will absorb it in overhead, undifferentiated from what every competitor bought at the same list price.

Our thinking:

This is a hypothesis, and Finette states its test: revenue per employee at firms under ten people over the next two years. If that figure does not pull away from the middle band, the shape is a ramp favoring size rather than an hourglass. The move that pays under either outcome is structural — give a three-to-six person team its own tooling budget and permission to build workflows that never touch production, hand it one engagement process this quarter, and see what comes back in six weeks.

What's new:

Code changes to Meta’s internal platforms and infrastructure rose 220% year over year, according to an internal post by CTO Andrew Bosworth reviewed by Reuters. Product improvements reaching users rose far less. Meta shelved the second wave of Project OT, its plan to run “AI native” teams.

How it works:

Project OT came out of Zuckerberg’s January retreat: AI agents overseen by smaller “talent- dense” teams, cuts via layoffs, hiring freezes and performance exits in two waves. Reuters reports Zuckerberg pulled back May 19, hours before the first wave; May’s cut landed at roughly 8,000, about 10% of staff, and the November wave was dropped./p>

Behind the news:

Internal posts reviewed by Reuters put major technical and security incidents up 40% from the prior year, with time spent responding up 70%. Meta declined to comment on those figures. In June, attackers exploited Meta’s AI-powered support bot to reach high-profile Instagram accounts, including the dormant Obama White House account.

Why it matters:

Firms measuring AI by preparer-side output (memos drafted, reconciliations prepared, tie-outs completed) will see those counts rise first. Reuters’ reporting shows generation scaling 220% while delivered improvements lagged and incident load climbed. The exposure is a staffing plan built on the first number. Firms that instrument review, rework and time-to-sign-off can price realized capacity before committing to headcount.

Our thinking:

In audit and tax, throughput is constrained by reviewer capacity. AI raises preparer output first, pushing volume into a reviewer pool that did not grow — the shape of Meta’s numbers, code up 220% and incident-response time up 70%. Firms that instrument the review queue now, ahead of the next tool purchase, will learn whether their constraint is drafting or sign-off, and that answer determines whether AI buys capacity or relocates the backlog.