What's new:

Anthropic has released an economic scenario explorer and accompanying research modeling showing how AI could affect U.S. growth, employment and wages through 2030. Its three scenarios span modest productivity gains to a profound economic transformation. These are conditional scenarios, not predictions, and the researchers assign no probabilities to them.

How it works:

The model treats occupations as bundles of tasks that AI can augment, automate or leave unchanged, while allowing new tasks to emerge. The explorer lets users vary assumptions about AI capabilities, adoption, autonomy, productivity and how quickly displaced workers find new jobs. It translates those assumptions into possible economic outcomes, making the distinction between what AI can do and what organizations actually adopt as central to the exercise.

Behind the news:

In the substantial scenario, U.S. gross domestic product (GDP) is 8.3% higher in 2030 than it would be without AI, but knowledge-worker wages are essentially flat. In the extreme scenario, GDP is 32.4% higher while knowledge-worker wages fall by more than 10% relative to the no-AI baseline. More of the economic gains flow to capital. The model also leaves out policy responses, business cycles and potential financial-market disruptions, factors that could materially change actual outcomes.

Why it matters:

For CPAs and finance leaders, the planning implication is that stronger economic growth does not automatically mean stronger demand for the same professional work. A client’s business could expand while its need for human preparation, analysis or reporting changes. That creates two questions for firms: how AI changes their own delivery economics and how it changes what clients need from them. Budgeting only for productivity gains addresses the first while leaving the second largely unexplored.

Our thinking:

The useful exercise is to pressure-test the firm’s assumptions. Take one service line and model three conditions: AI assists the team, AI performs most production with human review, and clients handle more of the work themselves. What happens to fees, staffing, training and the reason clients hire the firm? Then identify the evidence that would trigger a change in the plan. Anthropic’s model does not tell accounting leaders which future will arrive. It gives them a reason to stop building every forecast around the same firm, selling the same work, with faster delivery.

What's new:

OpenAI shipped two industry products. ChatGPT for Financial Services (Sept. 10) bundles GPT-6 Astra with hosted PitchBook, Daloopa, LSEG and Crunchbase data. Astra for Law (Sept. 17) adds a legal search index that raised correctness on Vals AI’s Legal Research Bench from 38.7% to 54.0%, in OpenAI’s testing.

How it works:

OpenAI pairs its frontier model with three things vertical software vendors sell: licensed data, domain retrieval and firm templates. The finance product hosts provider data on OpenAI’s servers so answers cite specific tables. The law product searches 230 million URLs, including CourtListener’s case law. Both run on ChatGPT Enterprise controls such as SSO, role-based access and information barriers.

Behind the news:

Tax is a favorite AI proving ground. The U.S. code is long, and nearly every American files. Greg Brockman’s GPT-4 launch demo in 2023 calculated a couple’s taxes. Column Tax’s TaxCalcBench (2025) found the best model got under a third of returns right. On Vals AI’s Tax Agent Bench, GPT-6 Astra ranks 24th (63.3%), and Claude Fable 5.1 leads (77.6%).

Why it matters:

For transaction-advisory teams, the finance product already handles adjusted-EBITDA reconciliation, leverage buyout (LBO) models and earnings analysis. A firm that pays separately for a model, a data feed and a specialist tool could buy all three from one vendor. That vendor also grades its own work. Every performance figure in both launches comes from OpenAI’s testing, and audit-grade reliance needs independent evidence.

Our thinking:

OpenAI calls its approach composable, and Harvey and Legora will build on Astra for Law. But the data deals, retrieval and templates that set those vendors apart now ship inside ChatGPT. We read this as a revenue move. Labs need income beyond API fees, and professional services pay well. OpenAI says it will expand into other financial services categories, and tax is already the benchmark labs reach for. Firms buying AI tools should price each point solution against what a lab is likely to bundle within a year and favor tools whose edge is proprietary data or assumed liability.

What's new:

Only 32% of 532 large U.S. and UK companies report even one quantified AI result, according to Amin Mrini's AI Value Gap Index for Q2 2026. By contrast, 71% report a live AI deployment. The gap between the two widened this quarter.

How it works:

Mrini scores companies on a five-level “Proof Ladder,” from a stated AI plan (L1) to AI reported as a recurring line item (L5). The sources are 10-Ks, 10-Qs, 8-Ks, earnings calls and web research. The index is weighted by revenue and by how much each company talks about AI. It measures disclosed evidence, so a company that uses AI quietly scores low.

Behind the news:

Regulators already police the talk side. In March 2024 the SEC settled “AI washing” charges against advisors Delphia and Global Predictions for $400,000 combined. Mrini flags a similar pattern at Block. Its CFO attributes an expected doubling of gross profit per employee (to about $2 million) to AI. Block also cut 40% of its staff, and its earnings calls don’t repeat the claim.

Why it matters:

Financial services moved the fastest. Of the sample's 94 companies, 53% now report a quantified AI result, up from 40% a quarter earlier. Only 12 companies in the whole sample (2%) track an AI result through a recurring KPI. That gap is an advisory opening, because clients need baselines that isolate AI’s effect. It is also an exposure in valuation and diligence work, where projected AI savings often rest on no tracked baseline.

Our thinking:

The 39-point gap between deployment (71%) and proof (32%) more likely reflects lag than failure. Companies deploy first and measure later, and the index already rose 17% in one quarter. Hemingway’s description of going bankrupt, “gradually, then suddenly,” fits the likely path. Proof clusters where work is data-heavy and transactional. Ten of eleven information-services companies report results, and accounting has the same profile. Firms that set baselines and track AI metrics now will have numbers to show when clients ask, and firms that dismissed AI on this quarter’s data will be measuring from behind.

What's new:

AI context windows grew from 512 tokens in 2017 to 2 million in 2026, doubling about every 14 months, according to a March 2026 preprint by Netanel Eliav of the Machine Human Intelligence Lab. Between 2004–2026, he estimates that human effective reading span fell from about 16,000 tokens to 1,800.

How it works:

The AI figure comes from model specifications, and the human figure is an estimate. Eliav multiplies session reading time by average reading speed (238 words per minute) and a comprehension factor he modeled himself. The session times come from Gloria Mark’s screen studies, which found that attention on one screen fell from 150 seconds in 2004 to 47 by 2016–2020.

Behind the news:

The attention decline predates ChatGPT. Mark’s data ends in 2020 and had been flat since 2016, which she reads as a new equilibrium. Eliav’s “delegation feedback loop” holds that people who delegate more practice less and then delegate more. It rests on smaller studies. Gerlich (2025, 666 participants) linked heavy AI use to weaker critical thinking. An MIT team (54 participants) recorded lower brain engagement during AI-assisted writing.

Why it matters:

Audit, review and advisory work depend on reading the source in full: the footnote, the contract clause, the reconciliation. Auditing standards require professional skepticism, which means critically assessing the evidence itself. That same judgment is central to a CPA’s role as a trusted advisor. If AI summaries become the default first read, the skills that help CPAs question assumptions, spot inconsistencies and give clients confidence in their advice get less practice. that skill gets the least practice. Firms can respond through workflow design. They can decide which reviews stay manual on purpose, so delegation doesn’t creep down to every task.

Our thinking:

Eliav concedes the full loop is untested. No study has followed the same people across years of LLM use. The component evidence all points the same way, though, and accountants have reason not to wait for the longitudinal data. If judgment is what separates a trusted advisor from the model, judgment needs practice, and delegation removes practice. Firms that use AI to skip junior review work will raise seniors who never learned to spot the error. The firms that do better will keep a deliberate share of the reading human, and test whether their people can still do it.

What's new:

Five frontier labs released new model versions within four days of one another, according to Dewain Robinson of Cloud Wars. He argues that release speed now exceeds what enterprise buyers can evaluate and adopt, and he expects the labs to slow down.

How it works:

Each new model makes the buyer repeat the same work. Teams retest outputs, rerun compliance and security reviews, and rework integrations tuned to the old model’s behavior. Robinson adds a supply problem. He says OpenAI withdrew GPT-6 Astra’s Pro offering within seven days, because compute capacity can’t keep pace with launches.

Behind the news:

Enterprise software buyers are used to long support windows. Microsoft supports a Windows Server release for ten years. Model APIs work differently. Providers publish dated model snapshots but retire them on their own schedules, so pinning a version delays the migration without removing it. Robinson also links release speed to Microsoft’s slower-than-expected Copilot uptake.

Why it matters:

PCAOB AS 1215 requires audit documentation that lets an experienced auditor understand the work performed. If the model behind a workpaper changes, the same prompt can return a different answer, and a retired model can’t be rerun at all. Firms should record the model version with every AI-assisted procedure. Firms that keep their own known-answer test cases can approve a new model in days instead of months.

Our thinking:

Buyers face a real tradeoff. Chasing every release means paying for evaluation and rework every cycle. Pinning a version saves that cost until the provider retires the model or a competitor’s tools pull ahead. Evaluation cost decides the choice. A firm that turns its own past engagements into a standing test suite can rerun it on each new model and switch when the results justify it. A firm without one will move late or move blind. Robinson expects releases to slow, but even a quarterly cadence beats most firms’ procurement cycles, so the test suite is worth building either way.