Get a coffee. Sit down. This one runs longer than usual.
Not because any single story this week is a paradigm shift. No one dropped a product that rewrites the profession overnight. There's no "one thing you need to know." What happened instead is an accumulation of small things that make big waves — and when you stack them up, the direction is unmistakable.
PwC's CEO told partners that anyone who thinks they can opt out of AI is "not going to be here that long." Your bookkeeping vendor launched an AI agent that learns from your corrections and charges you for the privilege. Both Xero and Intuit signed multiyear deals with Anthropic in the same fortnight. Anthropic leaked its most powerful model through a basic security failure. And OpenAI killed Sora because the math stopped working. Five major stories, each one worth sitting with.
The largest firm in your profession just drew a line
PwC US CEO Paul Griggs said it plainly: partners who resist AI "have no place in this firm." Not "should consider adopting AI." Not "might want to explore." No place. No free passes. Anyone who believes they can opt out "not going to be here that long."
This isn't new sentiment. We've heard versions of it from firm leaders for a year. What makes Griggs's statement different is the context around it. PwC is simultaneously restructuring compensation — bonuses now track revenue and margin per professional, not hours billed. Leaders are measured on progress moving services to AI. The firm is hiring fewer traditional accountants proportionally, replacing them with engineers and data specialists.
When Sequoia Capital published their intelligence-versus-judgment framework earlier this year — arguing that most professional services work is intelligence (rules, pattern matching, learnable decision trees) rather than judgment — the accounting profession should have felt the ground shift. PwC just confirmed the thesis. If the largest professional services firm on the planet is telling its partners that resistance means exit, the mid-market conversation changes. This isn't early adopter territory anymore. The Sequoia framework and the PwC mandate point the same direction: the intelligence layer of your work is getting automated, and the only defensible position is the judgment that sits on top of it.
But here's the part that should concern every smaller firm. Griggs didn't just issue a mandate. He dropped a product. PwC launched PwC One — a self-serve AI platform offering tax, compliance, and consulting intelligence on a subscription, without a PwC person in the loop. PwC's partner rates are out of reach for most mid-market companies. PwC's platform isn't. If subscription-priced self-serve advisory scales, it competes directly with the advisory services that small and mid-market firms sell. That could significantly reshape who gets that work and at what price.
Your bookkeeping vendor just started learning from you
Dext launched AI Assist on March 23 — an AI agent built into the platform that watches your team's corrections, categorizations, and tax rule applications, then builds persistent automation rules from those patterns. It doesn't just do the work. It learns how you do the work and replicates your decisions across future transactions. Every suggestion comes with transparent reasoning. Nothing auto-applies without your approval.
This is more significant than another "AI does bookkeeping" headline. This is self-learning at the vendor level. When your bookkeeper corrects a sales tax rate on a receipt, that correction doesn't just fix one transaction — it becomes a rule that applies to every similar transaction going forward. In input tax credit jurisdictions — Canada, New Zealand, Australia, the UK, Europe — that's material. The sales tax rate is the most important data point on a receipt. An incorrect rate means incorrect ITC claims, which means real money at risk and potential audit exposure. Having corrections stick and compound is genuinely useful.
The scale is real. Dext processed 31.4 million receipts and invoices in January 2026 alone, reducing over 2 million hours of manual work to roughly 206,000 hours. AI Assist is designed to push that further by encoding the intelligence your team provides every time they fix something.
Now the pricing. Free trial through April 23. Then £5 per client per month for accountants. Free for Dext Solo — business owners using the platform directly. Read that again. The accountants who teach the AI through their professional corrections pay. The clients who benefit from that collective intelligence don't. Dext is monetizing your expertise back to you while giving it to your clients for free. That's a pricing structure worth questioning, even if the product itself is useful.
The strategic question is who owns the encoded knowledge. Every correction your team makes inside Dext's platform builds Dext's AI, not yours. The context engineering we advocate — building structured, firm-specific knowledge that compounds over time — is happening inside your vendor's walled garden. Your expertise makes their product better for everyone. Whether that trade-off is worth £5 per client depends on whether you're building parallel context in systems you control.
Your accounting platforms are picking sides
Last week we covered Intuit signing a multiyear deal with Anthropic to bring Claude directly into QuickBooks and TurboTax. This week, Xero did the same thing. On March 26, Xero announced a multiyear partnership with Anthropic that works in both directions. Claude's reasoning now powers JAX — Xero's AI assistant — for cash flow analysis, invoice tracking, and revenue performance. And Xero's live financial data flows into Claude.ai, so business owners can ask questions about their books without leaving Anthropic's platform.
Two of the three major accounting platforms have now signed exclusive AI partnerships with the same provider — in the same fortnight. That's not a feature update. That's a platform convergence event. The infrastructure layer of your profession is consolidating around a specific AI provider, and the integrations are bidirectional. Your clients' data doesn't just stay inside Xero or QBO anymore. It flows into Claude, where it can be analyzed, modeled, and acted on — by anyone with access.
Two platforms choosing the same AI partner makes sense economically. It's cheaper for everyone. It means Anthropic (not Intuit, not Xero) is the strategic arbiter of what accounting intelligence looks like. If you're building practice-specific knowledge in your own systems, you're potentially building parallel to what the platforms are doing. If you're relying on the platforms' bundled AI, you're relying on Claude — which is fine, but it means your firm's knowledge is flowing through a system you don't control, where it's subject to Anthropic's API updates, pricing, and strategic priorities. Neither is wrong. Both are worth knowing.
A quick hit: The week of Anthropic, continued
Speaking of Anthropic — March 18 through March 25 was a chaotic week for the company. On March 18, Anthropic leaked what appeared to be Claude 4, an unreleased model, through a basic authentication failure on its website. The prompt and the outputs went viral. The model showed reasoning about the leaked credentials themselves, which spooked the internet for exactly the right reasons.
Then on March 20, OpenAI reported that Sora — their video generation model that made headlines last year — is being shut down. Publicly, they said the model isn't meeting production quality standards and the compute cost is too high for commercial viability. Less publicly, the math stopped working. You can generate great videos. You can't charge enough to cover the inference cost. That's a clean economic signal: the feature is real but not defensible.
What matters for accountants: Large model inference is expensive. Most of what we're building right now is viable because we're using smaller, faster models (Claude Haiku, Claude 3.5 Sonnet) that cost cents per call. If the economics of large models don't work for OpenAI's video generation — a fully solved problem where the demand is obvious — they won't work for bespoke accounting intelligence either. That constrains what's economically feasible for generalist platforms and leaves the defensible ground to firms that build specific, repeatable, high-margin use cases.
AI economics meet reality
The Sora shutdown is the most important story of the week, and it's getting almost no coverage. Here's why it matters: AI platforms make money on volume. Sora had no volume. OpenAI bet that video generation would be high-margin (people would pay for it) and low-cost (inference would get cheaper). It turned out to be low-margin and high-cost. The business model broke, so they ended the product.
That same calculation is running constantly across the industry. Anthropic is under pressure to prove that large models are profitable at scale. Xero, Intuit, and the other platforms are funding AI features and betting the margin math works out. Every firm building AI skills is quietly doing their own margin math — can we afford to run this skill on every client every month, or only on high-value clients, or only when specifically requested.
The firms that win through this wave aren't the ones with the fanciest AI. They're the ones who understand the economics. High-frequency, low-cost tasks (monthly bank categorization, invoice routing, exception flagging) are defensible because the compute is cheap enough that even at low prices you make margin. One-off, complex, high-touch tasks (strategic financial planning, tax optimization) are defensible because clients will pay for the outcome, regardless of the AI cost. The middle — "nice-to-have" analysis that's expensive to run and clients won't specifically pay for — will get squeezed out.
Three quick hits
Anthropic's tier pricing is live. If you've been thinking about building with Claude, the new pricing model cuts the cost of Haiku by 80% and Sonnet by 40% on input tokens. This is the moment to rebuild anything you've been deferring because the unit economics didn't work.
Stripe launched a native payments copilot inside Stripe itself. For a long time, only API-first companies could delegate transaction logic to AI. Now you can get AI assistance with payment flows directly in the dashboard. Still early, but the pattern is clear: every platform is embedding AI into the product. The gap is where the opportunity is.
Perplexity (the AI search company) is raising another funding round at a $9 billion valuation. Google just launched an internal initiative to make search compatible with AI agents. Neither of these events is surprising on its own. Together they signal something real: search, as we know it, is being rebuilt for agent access. If you're still building workflows that require humans to search for information, you're building workflows that will look quaint in 18 months.
The direction
You don't need to be smart about AI to see where this is pointing. The largest firm in the profession just told partners that resistance is exit. The vendors are consolidating around a single AI provider. The platform companies are embedding AI into core workflows. And the math — the actual unit economics — is getting real. Sora proved that "better AI" doesn't automatically mean "viable business." Anthropic's pricing proves that "cheaper inference" is the real lever for scaling.
The next accounting firm innovation won't be "better AI." It'll be "AI at a price we can afford to run every day, on every client, on the work we're actually doing."
The firms doing that are already building. Are you?