PwC and Accenture — two of the biggest consulting firms in the world — spent this week publicly restructuring around AI. Anthropic shipped Opus 4.7 and quietly moved enterprise contracts to per-token pricing, while investors circled at an $800B valuation. OpenAI bought a personal finance startup. Every one of these stories lands the same message in a CAS practice's inbox: the comfortable window for deferring AI decisions just closed.
The advisory firms just stopped waiting
PwC unveiled a global consulting overhaul this week, explicitly restructuring its member-firm network to support the rollout of PwC One — an AI-mediated delivery platform where clients describe a problem, autonomous agents do the work, and PwC professionals review rather than drive the outputs. The same week, PwC published its 2026 AI Performance Study, based on 1,217 senior executives across 25 sectors. The headline finding: 80% of the financial returns from AI are being captured by 20% of companies. Those companies aren't focused on productivity. They're focused on growth. That 80/20 gap is the thesis of Thursday's piece — what the 20% are actually doing, why most CAS practices are stuck on the productivity side of the line, and the one thing that moves you across. Come back for it.
Accenture is the more uncomfortable data point. The firm cut 11,000 people in the last quarter as part of an $865M restructuring charge, trained 70,000 in AI technologies, and is now hiring again — 4,000 added in Q1 FY2026, specifically for AI capability. CEO Julie Sweet was unusually direct: "We are exiting on a compressed timeline people where reskilling is not a viable path."
If you're a CAS owner talking about "moving to advisory," the two biggest advisory firms in the world just wrote your case study. They can't defer this. They're not trying to. If you're still telling yourself you can wait another year, the firms whose pricing you'll eventually benchmark against have already started.
The AI bill just got bigger — and the productivity did too
Anthropic shipped Opus 4.7 on Wednesday. Same $5 / $25 headline pricing, 3x production-task accuracy versus Opus 4.6, and the benchmarks that matter most for CAS jumped the most — Finance Agent to 64.4%, Office QA Pro from 57% to 81%. The catch is a new tokenizer that can emit up to 35% more tokens for the same input text. So "same per-token price" doesn't mean "same monthly bill."
Then the quieter story from the newsletter sweep this week: Anthropic moved every enterprise contract from flat-rate seats to per-token consumption pricing. The flat-fee era is over. Read with the $800B valuation talks, the October IPO speculation, and a revenue run-rate now past OpenAI, the picture is consistent. Anthropic is narrowing on enterprise. OpenAI is broadening into media, personal finance, and consumer ads. Different bets.
For budgeting, the shift is subtler than "AI just got more expensive." The $20 Claude Pro plan still exists. Claude Team's Standard seats are also $20 per user and start at five seats, making Team the realistic firm-scale plan for most CAS practices. Enterprise — with SCIM, audit logs, compliance APIs, and custom data retention — is now $20/seat plus variable token usage, but it requires a 20-seat minimum and annual-only billing, which puts it out of reach for most smaller practices anyway. What changed this week is the Enterprise billing structure itself. Legacy tiers were reportedly $40–$200/seat flat-rate. The new structure drops the base to $20 and puts token consumption on a separate, uncapped meter on top. For firms on legacy Enterprise tiers, the per-seat line item went down. What replaced it is a token meter. Heavy users — Champions running agents, operators with nightly Routines — move to Claude Max at $100/mo or Team Premium seats at $100/seat for 5x usage in a single fixed line item. The planning question shifts from "how many seats do we buy?" to "how much token usage do we let accumulate before we cap or instrument it?"
There's a second risk to plan for that most firms haven't priced yet. Practice-management systems and the accounting platforms themselves are absorbing these same token costs, and those costs have to show up somewhere. Expect subscription price increases or metered AI features from Karbon, Financial Cents, Xero, and every platform baking AI in by default. The productivity gains in Opus 4.7 are real enough to justify the spend. The variability is what will surprise firms that haven't instrumented their usage.
Your client just got the tool you've been selling
OpenAI acquired Hiro Finance on Monday. Hiro was a fintech specifically trained to nail financial math, backed by Ribbit, General Catalyst, and Restive. The product sunsets on April 20, the team folds into ChatGPT's financial planning capability, and the deal is OpenAI's second vertical acquisition in two weeks. It's also the second time in ten days a major AI company has planted its flag on the "personal CFO" positioning — Perplexity shipped its Plaid integration last week.
Here's the part that matters more than the headline. I was playing with the Xero integration this week. One prompt — not twenty, one — and I had a live-data dashboard pulling from my ledger. It worked. A year ago that would have been a week-long engagement. Your clients are about to figure out the same thing. The AI your client holds in their personal space now reaches into their accounting platform, their bank feeds, and their tax data.
The question this forces back on the CAS practice is uncomfortable, and it's the right one. What does your firm sell that's more than a prompt plus the platform? Relationship, judgment, accountability, context — yes, all of those. But you have to name them. If you can't, the client will run the comparison for you, and they'll run it from a phone. That gap — between the client's one-prompt dashboard and your firm's actual deliverable — is where your next three years of practice value gets built. Saturday I'll show you where to place the first agent that lives in that gap, and why picking the most painful task in your practice is the wrong place to start.
Quick hits
Juno raised $12M. CPA-founded tax prep AI, automating 90% of data entry across 92+ document types, explicitly targeted at the SMB accounting firm rather than enterprise software. 500 firms onboarded in year one. The founder saw an OpenAI demo in 2023 and built the tool he wanted for his own practice. Bonfire Ventures led.
OpenAI's Codex went beyond coding. Computer-use on Mac means Codex can now operate any app with its own cursor — no API required. The mono-thread pattern is the more interesting shift: threads persist context across a schedule, with heartbeats checking sources every 15 minutes. One OpenAI engineer's most useful Codex thread has been running for three weeks. The context-engineering thesis just got productized.
Anthropic shipped Claude Code Routines. Scheduled agents running on Claude's web infrastructure — your laptop doesn't need to be on. Narrow relevance today, but the pattern is right. A firm-specific agent can now run nightly against client notes and account data and hand you a queue at 9am. The constraint is whether your firm has the SOPs to encode.
Claude Design launched. A Figma and Canva competitor built on Opus 4.7, exporting to PDF, PowerPoint, Canva, and HTML. Figma stock dropped 6.8% on the day. For firms producing their own client-facing visuals, the production barrier dropped another notch — and it's sitting on a plan you probably already pay for.
GPT-5.4-Cyber shipped with access gating. OpenAI's cybersecurity variant, available only to KYC-verified defenders. Same pattern as Anthropic's Mythos restraint two weeks ago: frontier capability is increasingly two-tier, general-access and verified-access. For regulated-industry work, the gap between the two will grow.
Xero's "Xero OS" announcement was marketing, not product. On April 14, Xero published a blog post branding its existing platform as "Xero OS" — an "AI-native operating system" with JAX as the coordination layer. Read the post against what Xero has already shipped and the math doesn't hold. JAX has been in beta since August 2024. The agentic repositioning happened at Xerocon Brisbane in September 2025. Automatic bank reconciliation has been in global beta since then. No new feature shipped April 14. No capability moved from beta to GA. If you're budgeting around what your Tier 1 platform actually delivers, ignore the rebrand and read the release notes.
The value test just got harder
Every story this week asks the same question from a different angle. Consulting firms restructuring around AI ask: what's worth charging for when AI does the analytical work? Anthropic's tokenizer change and pricing shift ask: what's your AI bill worth per dollar of output? OpenAI buying Hiro and the one-prompt Xero dashboard ask: what do you sell that a client with a free AI agent and a ledger cannot produce themselves?
The value test isn't hypothetical anymore. Your client will run it this year. Your pricing will reflect the answer by next year. And the advisory firms whose fees you eventually benchmark against have already started.
What is your practice's answer?
Take the free AI Readiness Scorecard at theaiaccountant.ai/scorecard — 25 questions, 5 minutes, and you'll know exactly where your practice stands.