Read time: 4 minutes

Leader, welcome back.

Two of the biggest AI labs shipped new products this week aimed directly at how your team works.

Not how you work individually. How your team works together.

That shift is worth paying attention to.

On Friday, Anthropic launched Claude Design at claude.ai/design.

Think of it as a designer on demand, available to anyone with a Claude Pro, Max, or Team subscription.

Here is what it can do.
You describe what you want, or drop in a screenshot, an existing deck, or a Figma file as a starting point. Claude asks clarifying questions before generating anything, which matters more than it sounds. Then it builds the output as editable components, not a flattened image you cannot touch.

The things you can make: websites and landing pages, slide decks and board presentations, marketing one-pagers, social assets, internal reports.
Anything a modern company produces visually. When it is ready, it exports to PDF, PowerPoint, Canva, or standalone HTML.

The gate to branded, production-quality work just moved for every knowledge worker in your organisation.

The PM who needed mocks, the sales lead who needed a battle card, the finance director who needed a client-ready deck: all of them can now produce that work without a design handoff. The floor for what a non-designer can produce rose sharply this week.

Then on Tuesday, OpenAI launched Workspace Agents.

These are Codex-powered agents that run in the cloud, connect to Slack, Salesforce, Google Drive, Notion, and more than 90 other tools, and keep working even after you close the tab.

The use cases OpenAI highlighted are telling. One of the flagship templates is a weekly metrics reporting agent that automatically pulls data every Friday, generates charts, drafts the narrative commentary, and delivers the finished report to the team. That is not a generic demo. That is a finance workflow described as a headline feature.

Other templates: a product feedback agent that monitors Slack and external channels and turns signals into weekly action items. A chief of staff agent that manages briefs across meetings and tools.

The agents are free until May 6, then move to credit-based pricing.

Here is the point both announcements are making, whether either lab has said it directly or not.

Anthropic and OpenAI both moved this week, and they moved in the same direction.
Claude Design places AI inside every creative output your team produces. Workspace Agents places AI inside the workflows that run your operations.

Neither is asking you to go to a new tool.

Both are embedding AI in the places where the work already happens.

The individual productivity era of AI, where one person opens a chat window and saves themselves an hour, is giving way to something different.
AI is becoming team infrastructure. It runs in the background. It connects to your existing tools. It does the repeatable work automatically and delivers the output to whoever needs it.

For finance teams: the weekly reporting pack, the variance commentary, the board deck. The work that absorbs your team every close cycle.
That is exactly the category both companies built for this week.

Cat Wu is head of product for Claude Code and CoWork at Anthropic. She was on Lenny's Podcast this week and said something that stopped me.

"If an automation doesn't work 100% of the time, it's not really an automation."

Most finance leaders are sitting at 90 to 95% and calling it good enough. Cat's point is that good enough is actually worthless. That last 5 to 10% is exactly where the trust breaks down: a wrong number in a board pack, variance commentary that cites a driver that does not exist in your P&L, an output that costs more time to check than it saved to produce.

Here is the pattern I see constantly. A finance leader runs Claude on a clean example dataset. The output looks impressive. They share it with the team. Then they try it on the actual March close and it breaks on three things specific to their business. They fix those three manually. It breaks on two more. They quietly go back to doing it themselves and tell people AI is not quite there yet.

The problem is not Claude. The problem is they stopped at 95% and expected it to hold.

Here is how you get to 100% on a variance commentary workflow. This is worth fixing first because it happens every single month and the quality directly affects how your CFO reads the numbers.

Step one: show it what good looks like. Open your Claude Project for your finance function. If you do not have one yet, create it this weekend. Paste in two or three examples of variance commentary you have written yourself that you are genuinely proud of. Tell Claude this is the standard. Match this tone, this level of specificity, this structure.

Step two: run it on last month's real close data. Not a demo. Not a cleaned-up version. The actual numbers, with all the noise.

Step three: note every failure specifically. Not "the commentary was off." Write down exactly what it got wrong: wrong cost driver attributed to headcount variance, missing context on a one-off restructuring charge, too much passive voice. Specific failures only.

Step four: add that context directly to your Project instructions. If it keeps misattributing a variance, tell it explicitly: our Q1 headcount variance was driven by the delayed EMEA hire, not attrition. If it does not know about a restructuring charge, paste in the relevant paragraph from your board memo.

Step five: re-run on the same data. Compare. Repeat until you can use the output without reading every line twice.

This process takes longer upfront than just doing it yourself. That is the point.
You are building a workflow, not running a one-off prompt. The finance leaders saving hours every close are the ones who put in that elbow grease once and then rely on it every month after.

Pick one workflow. Get it to 100%. Then move to the next one.

The full prompt chaining workflow I use for this, including the exact project instructions and before/after outputs, is inside my AI for Finance Leaders course. Currently at a discounted rate for a limited time. Check it out here.

Every Claude Design output looks the same out of the box: teal gradients, serif headline, containers stacked on containers. If you use it this week without fixing this, your deck will look like every other AI slide on LinkedIn.

The fix is giving it a brand guide before you ask for anything.
You do not need a Figma file or a design agency.
Give it your hex codes and your font names and ask it to build your brand system first, then generate screens against that system.

Two ways to get started quickly: use ChatGPT Images, which also launched this week and is the best image generation tool available right now, to create reference visuals for the look you want.
Or go to getdesign.md, grab the design files for brands whose aesthetic you want to borrow, and drop those straight into Claude Design as your starting point.
That one step is the difference between generic output and something that actually looks like your company.

Cat Wu, head of product at Claude Code, put the shift happening right now into one sentence on Lenny's Podcast this week: "The 2024 generation of AI products was chat-based. The Claude Code generation is action-based." Both launches this week prove her point. Finance is on the same curve. The teams pulling ahead are not asking better questions. They are letting AI do the work and reviewing the output.