The number everyone quotes, and the one that matters
The headline from PwC and Anthropic's expanded alliance on May 14, 2026 is a big round number: 30,000 PwC professionals certified on Claude, with Claude Code and Cowork rolling out across U.S. teams first and a stated ambition to reach a global workforce of hundreds of thousands. Alongside it came the proof points — insurance underwriting cycles compressed from ten weeks to ten days, and delivery improvements of up to 70% across deployments — plus a joint Center of Excellence and a new Claude-native finance business group.
The metrics are the part that gets quoted in board decks. The number of people being trained is the part that should change how you think about your own AI strategy. One of the largest professional-services firms on earth looked at frontier AI and concluded that the limiting factor was not access to a model — anyone can buy that — but how quickly thirty thousand humans can learn to work with agents, and how fast the work itself can be re-shaped around them. That conclusion is the real signal.
The model was the easy part
Here is the uncomfortable truth that the size of PwC's program makes impossible to ignore: the model is now the cheapest, fastest-to-acquire input in the whole equation. Claude Code is an API key and a CLI. You can have frontier-class coding and reasoning in your environment this afternoon. What you cannot buy this afternoon is an organization that knows how to use it.
The ten-weeks-to-ten-days underwriting result didn't come from a model existing. It came from someone redesigning the underwriting workflow so an agent could carry the parts it's good at, with humans positioned at the decision points that actually require judgment. That redesign — figuring out where the agent goes, what it's allowed to touch, who reviews its output, and how the existing process bends to accommodate it — is the work. The model just made it possible.
This is why the training number is so large. You don't certify 30,000 people to teach them to type prompts. You certify them because the bottleneck to value is human and organizational, not technical, and closing it at scale is a year-long change-management program, not a software install.
What "certified" actually has to cover
A serious adoption program isn't a prompt-engineering webinar. To get from "we have Claude" to "underwriting is 7x faster," the training has to cover the things that determine whether AI helps or quietly creates risk:
- Where the agent fits in a real workflow — which steps it owns, which it assists, and where a human must stay in the loop. Most failed AI rollouts skip this and just bolt a chatbot onto an unchanged process.
- How to review AI output competently — the reviewer needs enough domain skill to catch the confident-but-wrong answer. Agentic speed is only an asset if the review keeps pace; otherwise you've just industrialized the production of plausible mistakes.
- The guardrails — what data the agent may access, what it must never do unsupervised, and how its actions are logged for audit. In regulated functions like underwriting and finance, this is the difference between a deployment and a liability.
- When not to use it — teaching people to recognize the tasks where the agent's failure mode is worse than its speed is worth.
Notice that almost none of this is about the model. It's about people and process. That's the bottleneck PwC is spending to clear.
The two halves of the real work
What PwC is buying, underneath the press release, is two things that have to happen together. One is building the agentic systems — wiring Claude into underwriting, diligence, and finance workflows with the right tools, permissions, and audit trails so the agent can actually do the work safely. The other is bringing the humans along — certifying thousands of people to direct, review, and trust those systems, and reshaping the processes around them.
Skip the first and your people are trained on a tool that isn't connected to anything. Skip the second and you've built an elegant agentic pipeline that no one trusts or uses. The 70% delivery improvement is what you get only when both land at once.
Where Sonnet Code fits
This is the seam between the two things we do, at the scale most companies actually live at — not 30,000 people, but a team that needs AI to move from demo to production without a year of flailing. AI development is the engineering half: wiring frontier models into your real workflows with the tools, permissions, and audit trails that make the deployment safe, and rebuilding the process around the agent instead of bolting it on. AI training is the human-in-the-loop half: the domain experts and senior engineers who define what "good" looks like for your work, build the evals that tell you whether the agent is actually reaching it, and stand up the review discipline that lets your people trust the output instead of rubber-stamping it.
PwC's lesson, scaled down to your size, is the one worth internalizing before you sign a single AI contract: the model is the easy part. The value lives in the workflow you redesign and the people you teach to run it — and that's the part worth getting right.
