Your AI surface area grew.
Choose where it belongs.

What actually changed

• Anthropic launched Claude for Small Business on May 13 as a toggle inside Claude Cowork.
• The package included fifteen workflows for jobs such as marketing campaigns, payroll planning, month-end close, and performance monitoring.
• It also included fifteen reusable skills for cash-flow forecasting, invoice chasing, lead triage, content strategy, and contract review.
• Connectors covered QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365, and Slack.
• Anthropic and PayPal also released a free AI fluency course built around a 4D Framework.

Try this: Give one workflow a real job

• Connect either Canva or Google Workspace, whichever already contains the material needed for this week's work.
• Choose one job, such as auditing sixty days of content or preparing three graphics from existing brand assets.
• Give Claude the source files and the definition of a usable result before it starts.
• Review the finished output once and write down what context was missing. Add that context before repeating the workflow.

My take

This was the first time Anthropic packaged Claude for the stage my members actually operate at. Canva and Google Workspace were the two connectors most likely to earn their keep first.

If your business runs on visuals and shared documents, choose one of those connections and one real job. The number of workflows was less important than whether one of them could produce something you trusted.

What actually changed

• Workday's 2026 research found that 37 to 40% of the time supposedly saved by AI was spent reviewing, correcting, and verifying its output.
• Employees who closely monitored AI output reported 12% more mental fatigue than people who let it run and audited at the end.
• Advisor Perspectives applied the finding to professional services, where solo and small-team operators absorb the review burden themselves.
• The pattern matched earlier research on the cognitive fatigue created by managing multiple AI systems.

Think about this

• What percentage of your last AI session was spent prompting compared with checking the output?
• Where are you reviewing the same kind of AI work twice because the first result still does not feel trustworthy?
• What context, example, boundary, or source could move that correction upstream?

My take

Putting my clinical hat on, this research was measuring hypervigilance. When you do not trust the output, you watch it more closely, and the watching itself is exhausting.

The extra fatigue was not simply from using AI. It came from not trusting it. The fix started upstream with tighter context, clearer examples, and a defined review point.

What actually changed

• Zoom announced its first Solopreneur 50 on May 4.
• Fifty AI-powered businesses of one were selected from nearly 3,000 US applicants.
• Five winners received $30,000 each, for $150,000 in grants.
• Services and consulting accounted for 20% of selected businesses, while technology and SaaS accounted for 5%.
• Solo-founded startups had grown from 23.7% in 2019 to 36.3% by mid-2025.

Think about this

• What concrete AI workflow and measurable output would you have named in an application?
• Which grant programs or business categories might follow this one over the next year?

My take

This was structural recognition that AI-powered businesses are a real category, not a side project anyone needs to apologize for.

That is why the Lab was built around helping people use these tools for the business they are actually building, not around selling a generic course about AI.

What actually changed

• Gemini in Chrome let eligible Google Workspace users save useful prompts and run them with one click across the web.
• Claude, ChatGPT, and Notion already offered their own versions of reusable skills or saved instructions.
• The repeated pattern was save once, run repeatedly, and edit when the work changed.
• Google had not yet specified which Workspace tiers received the feature.

Try this: Save the three instructions you repeat

• Choose the three prompts you typed most often in the prior month.
• Save each one in the reusable-instruction feature available in your primary AI tool.
• Run each saved instruction on real work from that week.
• Delete or combine any saved instruction you have not used in thirty days.

My take

This was a useful sign that context architecture was becoming a product pattern across tools.

Write the best instruction once. Keep the context current. Let the tool name change without rebuilding how you think.

What actually changed

• Memory improvements rolled out to ChatGPT Plus and Pro users in May 2026.
• Memory Sources showed whether past chats, saved memories, files, or connected Gmail had personalized a response.
• Users could delete or correct outdated memories from the source view.
• The feature worked across the models available on a user's plan.

Try this: Audit what ChatGPT thinks it knows

• Open ChatGPT's Memory settings and remove anything stale.
• Run one prompt you use often and inspect which memory sources shaped the answer.
• Add a boundary for anything that should remain inside one conversation rather than become ongoing context.
• Repeat the audit quarterly, not weekly.

My take

If ChatGPT was where you lived, this audit was worth ten minutes. I had found three pieces of context from a project I closed in February still influencing new outputs.

Correcting the memory once was easier than fighting the same stale assumption every time.

What actually changed

• A California class action was filed against OpenAI in May 2026.
• The complaint alleged that embedded tracking technology shared ChatGPT queries with Meta and Google for advertising purposes.
• OpenAI had not responded when the brief was published, and no hearing date had been set.
• The filing was a complaint, not a finding of fact.

Think about this

• What sensitive information had you placed in ChatGPT during the prior ninety days?
• Would a client, employee, or partner expect that information to remain outside a consumer chatbot?
• Have you reviewed the setting that controls whether chats improve the model?

My take

I was watching this one without changing everything based on an unresolved complaint.

If client information had entered ChatGPT, the time to decide whether that workflow was appropriate was before a ruling, not after one.

What actually changed

• OpenAI launched a Personal Finance preview in ChatGPT on May 15.
• It was available to Pro subscribers in the United States at publication.
• Plaid connected bank, brokerage, and credit-card accounts from more than 12,000 institutions.
• The dashboard summarized spending, subscriptions, portfolio performance, and upcoming payments.

Think about this

• Would this solve a financial job you already had, or would it create another dashboard to monitor?
• Were you comfortable with financial information and prompt history sharing the same work surface?

My take

I wanted to like this, but I was not ready to recommend it. Plaid was familiar infrastructure. The concern was the prompt history and the data flow around it.

If you were not already paying for Pro and did not have a clear job for the connection, there was nothing to do.

What actually changed

• Claude for Small Business put AI inside bookkeeping, invoicing, payroll, marketing, and design tools.
• ChatGPT Personal Finance put AI inside bank, brokerage, and credit-card accounts.
• The OpenAI complaint raised questions about where chatbot queries might travel.
• The practical pattern was an expanding surface area, each with its own access and review burden.

Try this: Run a surface-area check

• List the places where AI currently touches your business, including email, calendar, design, bookkeeping, banking, and chat tools.
• Mark each one as chosen, inherited, or unclear.
• Choose one unclear surface to investigate before adding anything new.
• For the next tool, ask which part of the business it will enter and whether that is where help is actually needed.

My take

I tried this audit and quickly found more AI surfaces than I could count accurately. That was the lesson.

The audit did not need to be perfect. It needed to interrupt the moment when another tool entered the business because the marketing found me when I was tired.

The useful question was which surfaces I chose, which ones accumulated, and which one I wanted to close.