Cowork Turns AI from Chat into Work: What to Tackle This Week

Anthropic shipped a series of announcements over the last 30 days. If you feel you missed something, you did. If you feel you have to read all of it, you don’t.

Some of it matters strategically: the new KPMG alliance, the expanded PwC partnership, the venture with Blackstone, Hellman & Friedman and Goldman Sachs, the Stainless acquisition, the Gates Foundation partnership. All of it points to where the market is heading.

For your Monday morning, only three questions matter:

1. Where can AI stop being a chat and start doing real desktop work?
2. Which prompts will behave differently on the new model?
3. Which single process is worth measuring this week?

1. Cowork Is Out of Preview — and It’s Not Claude Code for Non-Developers

Claude Cowork has been generally available since April in Claude Desktop on macOS and Windows. The real change isn’t that it’s “smarter”. The real change is the interface.

Cowork in one sentence: an agent you give a goal and permission to access selected files or apps, and it finishes the task instead of just telling you how to do it.

That’s the difference from a normal chat. Chat writes you a process. Cowork can walk through a folder, read files, create outputs, and operate the tools you allow. It still needs scope, oversight, and sensible guardrails. But it isn’t just a text box anymore.

Examples where it’s worth trying:

• Monthly reporting. Input: CRM and finance exports. Output: a draft report in the same structure as last month.
• Meeting prep. Input: a folder of materials. Output: a summary, risks, and questions for the decision.
• Shared drive cleanup. Input: unnamed PDFs. Output: renamed and sorted files with a changelog.
• Contract processing. Input: a batch of contracts. Output: a table with key parameters and clear places for human review.

Week one of adoption

Don’t launch a company-wide pilot. Pick one repeating process that today takes at least 30 minutes a week and has a clear output.

Use a simple frame:

Sample brief:

Walk through the files in this folder. Produce a monthly report in the same structure as report-last-month.docx. Don’t change any numbers without citing the source. For every uncertainty, write a question into a “For review” section. Save the output as a new document and add a short changelog.

This is exactly the kind of small test that builds on the AI pilot trap: one task, two weeks, measure the result, only then scale.

2. Opus 4.7: A More Precise Model Means Fewer Excuses for Vague Prompts

Claude Opus 4.7 came out on April 16, 2026. Anthropic describes it as a better model for complex, long-running, tool-heavy tasks. Practically, three things matter:

1. it holds the line in long tasks,
2. it handles tools and errors better,
3. it takes instructions more literally.

The first two are good news. The third is only good news for people who write precise briefs.

A prompt that used to work “in spirit” can suddenly fail because it isn’t clear what relevant, similar, concise, or done actually means. That’s not a bug in the model. It’s a new deal: you get more of what you wrote, and less of what you meant.

What to rewrite

Don’t go through every prompt. Go through the five most-used: slash commands, agent instructions, system prompts, report templates, or briefs you hand to Claude Code.

Of each line ask: “If a lawyer read this, what loophole would they find?”

This isn’t rewriting everything. It’s maintenance on the prompts you actually use. One day of work that gives you back consistency.

What it means for agents

Opus 4.7 cuts friction on agentic tasks, but it doesn’t make agents a cheap default. It’s worth reopening use cases that previously failed on loops, tool errors, or long context. But the frame from Skills vs. Agents still holds: repeatable skills first, then agents with monitoring.

3. Claude for Small Business: Interesting Mainly as Direction

Claude for Small Business launched on May 13, 2026. Anthropic positions it as a bundle of connectors and ready-made workflows for small companies. It runs through Claude Cowork and integrates with tools like Intuit QuickBooks, PayPal, HubSpot, Canva, DocuSign, Slack, Google Workspace, and Microsoft 365.

It ships with 15 pre-built workflows and 15 skills. Typical examples: month-end close, cash-flow planning, invoice chasing, lead triage, campaign execution, contract review, or pulling materials for the accountant.

For a European mid-sized company this doesn’t mean you fire your integrator on Monday and everything works. Some of the connectors fit the US small-business stack more than the European one. But the direction matters: AI is being packaged into specific work patterns, not into an empty chat.

For a sub-50-person company, the practical question isn’t “should we buy Claude for Small Business?”. The better question is:

Which of our repeating processes looks exactly like a ready-made workflow — invoices, leads, reporting, contracts — and can we describe it well enough that AI can safely take over 50% of it?

If you can’t describe the process, no connector will help you. If you can describe it, you may not need a six-month project.

4. KPMG and PwC: A Signal for Approval, Not Universal Proof

Hold two thoughts at once here.

Yes, large firms have started deploying Claude seriously. KPMG announced a global alliance with Anthropic — Claude is to be available to 276,000+ employees and integrated into the Digital Gateway platform. PwC has expanded its Anthropic partnership, with planned Claude Code and Cowork rollouts and certification for 30,000 people.

No, that doesn’t mean a European mid-sized company should buy everything tomorrow. KPMG and PwC have different budgets, different teams, and a different compliance apparatus.

What this means practically:

• For CEO/CFO: the “we’ll wait until enterprise approval is easier” argument is weaker than it was a year ago.
• For IT and security: there’s now precedent that AI tools can be handled with access controls, auditability, governance, and human-in-the-loop.
• For teams: a big rollout isn’t a substitute for the work of nailing down one concrete workflow. It’s only proof that the topic moved from experiment to operational agenda.

What to Do This Week

No big AI program. Three small moves:

1. Run Cowork on one repeating task. Pick a workflow that takes more than 30 minutes a week. Prepare an input folder, describe the output, measure time and the number of corrections.

2. Rewrite your three most-used prompts. Not all of them. Only the ones you actually use every week. Add limits, format, stop rules, and a definition of done.

3. Send your leadership one enterprise example. Not as proof you should buy Claude. As proof that AI is no longer just an experiment for enthusiasts.

If you’re not sure whether to start with Cowork, with rewriting prompts, or with an agent, book an AI audit. In 55 minutes we pick one process, score the risk, and define the first step. If you already know what to roll out but need to walk it through the team, AI mentoring is a better fit.

Summary

• Cowork is a practical shift from chat to desktop work. Test it on one repeating task.
• Opus 4.7 is better at long, tool-heavy tasks — but it punishes vague briefs harder.
• Claude for Small Business shows the direction: AI as ready-made workflows, not an empty box.
• KPMG and PwC are a strong signal for internal approval, but they don’t replace your own measurement.

Pick one process. Measure the result. Only then scale.

— Jirka

You Might Also Like

• Skills vs. agents: When you need a recipe and when you need a chef
• The AI Pilot Trap: Why 77% of Corporate Projects Never Reach Production
• Shift+Tab Is a Trap: Why Your AI Plan Ignores Half Your Codebase
• How to Prepare for an AI Workshop