In August 2026 the EU AI Act's transparency rules took effect while its high-risk rules were pushed to December 2027, Claude gained a built-in browser, shared memory and a Salesforce plugin, ChatGPT Business added $125 Premium seats, and Google and Microsoft shipped new agent and data controls. The actions: label AI content for EU audiences, set memory and browser rules, and check which agents can reach your files.
What should an operator do this month?
Three things, in order of how many teams they affect.
- If you serve customers in the EU, make sure any chatbot says it is AI and any AI-generated image, audio or video you publish is labelled.
- Decide whether Claude memory and browser use are on for your team, and which sites the browser may use.
- Review which agents and flows can reach your Google Drive or SharePoint files, and restrict what AI can read before staff find out for you.
The EU AI Act's transparency rules took effect, and the high-risk rules were delayed
What happened (August 2, 2026). The AI Act's Article 50 transparency duties began to apply: people must be told when they are dealing with a chatbot, synthetic audio, images, video and text must be marked, and deepfakes must be disclosed. The enforcement powers of the EU AI Office and national authorities also began, including fines for providers of general-purpose AI models. The high-risk obligations did not arrive. The Digital Omnibus on AI, in force since July 27, moved them to December 2, 2027 for uses such as hiring, credit scoring and education, and to August 2, 2028 for AI built into regulated products. Law firm Jones Walker notes that systems already on the market before August 2 have until December 2, 2026 for machine-readable marking only, not for the other Article 50 duties. Source: European Commission: AI Act enforcement European Commission: AI Omnibus enters into force Jones Walker.
Why it matters. A US business with a website chatbot or AI-made marketing that reaches EU customers is now covered by the disclosure rules. Teams using AI to screen job applicants or score credit have until December 2027, not August 2026, but the work is the same.
What to do. Add a plain "you are chatting with an AI assistant" line to any customer-facing bot, and label AI-generated images and video in EU-facing marketing. If you use AI in hiring, keep a record of what the tool does and who reviews its output. AI governance lite has a one-page policy to start from.
OpenAI added Premium seats to ChatGPT Business
What happened (announced August 10; available August 25, 2026). ChatGPT Business workspaces can now mix Standard and Premium seats. Premium gives five times the usage of a Standard seat and removes the five-hour usage limit. It costs $125 per user per month, or $100 billed annually, against $25 or $20 for Standard. Source: OpenAI.
Why it matters. The heaviest users on a team, usually the ones running agents or long research, were hitting limits. Now the fix is one seat upgrade rather than a whole-team plan change.
What to do. Look at who actually hits limits before buying. Upgrade those people, keep everyone else on Standard, and check again after a month. Claude vs ChatGPT for business compares the two plans.
Green Street made its MCP server generally available
What happened (August 11, 2026). Green Street clients can query its research, company and sector data, market data, forecasts, sales comps and single-property automated valuation model from corporate instances of Claude, ChatGPT and Gemini, covering the US, Canada, Europe and parts of Australia. Source: PR Newswire.
Why it matters. A major commercial real estate data vendor now meets brokers and analysts inside the AI assistant they already use. Expect others to follow, and expect questions about which data your firm's AI can reach.
What to do. If your firm subscribes, ask your Green Street rep to enable it and test it against a deal whose numbers you already know. MCP servers for commercial real estate explains how these connections work.
Microsoft 365 Copilot turned work data on by default and added a credit dashboard
What happened (release notes of August 11 and August 25, 2026). Copilot Chat moved to one interface for web and work with a Work IQ button, with work data on by default. SharePoint admins can mark authoritative sites whose news and policies Copilot Search puts first. A consumption dashboard in Viva Insights shows Copilot credit use for Cowork and the Work IQ API, for tenants on usage-based billing. Word on the web now lets users choose Anthropic models when editing with Copilot. Source: Microsoft 365 Copilot release notes.
Why it matters. With work data on by default, Copilot answers from anything a person can open in SharePoint and OneDrive, so loose sharing settings show up in answers. The credit dashboard is the first easy view of what usage-based agents cost.
What to do. Check sharing on your most sensitive SharePoint sites, mark your policy and HR sites authoritative, and if you pay for Copilot credits, open the dashboard once a month.
Instinct raised $250 million while testers raised privacy concerns
What happened (August 24 and 26, 2026). On August 24, TechCrunch reported testers' problems with Instinct, a personal assistant you text or call: an email sent without approval, inbox data summarized after access was revoked, and terms granting a perpetual license to user materials. Instinct told The Wall Street Journal it was taking the concerns seriously and added a tool to delete external data. On August 26, TechCrunch reported a $250 million Series B co-led by Index Ventures and Benchmark at a $2.5 billion valuation. It was in private beta. Source: TechCrunch on the concerns TechCrunch on the funding.
Why it matters. It is a clear example of what to check before any agent touches business data: the permission list, what happens on disconnect, and the terms.
What to do. Before approving any agent, read its terms for data licenses and for authority to act on the user's behalf. Disconnect it, then check whether it still acts.
Since then. On September 8, 2026, Instinct began giving users their own Instinct email addresses so the agent can open and run accounts on their behalf, and Meta launched Muse, a similar consumer agent, the same day. Source: TechCrunch. See September 2026.
Claude's memory started carrying across chat and Cowork
What happened (August 25, 2026). Claude's memory now works across chat and Cowork, on web, desktop and mobile. It is stored as topics a person can read, edit and delete in settings, and by default it leaves out sensitive subjects such as health and politics. Memory is on by default on Free, Pro and Max. On Team and Enterprise, admins decide whether it is available, and it stays off for each person until they turn it on. Source: Claude blog.
Why it matters. Memory saves re-explaining your business every session. It also means client details can persist beyond one chat, so the choice to allow it belongs to the business, not to each user.
What to do. On Team or Enterprise, decide whether memory is available and write one line on what should never be stored in it. On a personal Pro plan used for work, open the memory topics and delete anything about clients.
Claude in Chrome went GA, and Cowork got its own built-in browser
What happened (August 26, 2026). Claude in Chrome became generally available on all paid plans, with Enterprise admins able to limit it to approved domains. The same day Anthropic added a built-in browser to the Cowork desktop app, rolling out to Pro, Max and Team on macOS, Windows and Linux (beta) and available to Enterprise through organization settings. It is separate from your own browser and cannot see your tabs, bookmarks or saved passwords unless you import logins. Source: Claude blog: Chrome GA Claude blog: built-in browser.
Why it matters. Tasks such as pulling invoices from a vendor portal or collecting research can now run in a browser Claude controls, while someone keeps working. Anything that can click and type on a website can also be misled by one.
What to do. Start on sites you trust, keep banking and payment sites out, and on Enterprise set the domain allowlist before rolling it out. What Claude Cowork does covers the basics.
Since then. On September 16, 2026, Anthropic began merging Cowork into the main Claude app, rolling out to Pro and Max first with Team and Free to follow, and at least 30 days' notice for Enterprise admins. Source: Claude blog.
Salesforce and Anthropic announced Claudeforce
What happened (August 26, 2026). Salesforce and Anthropic expanded their partnership as Claudeforce. It starts with Salesforce in Claude, a plugin with 37 prebuilt sales skills for meeting prep, deal reviews, pipeline updates and actions that follow Salesforce's own business rules, running inside Salesforce's trust boundary through Amazon Bedrock. Claude also becomes the default model for Slack AI and Slackbot. The plugin was with select pilot customers, and Salesforce said an open beta was expected in September 2026. Source: Salesforce newsroom.
Why it matters. For a sales team on Salesforce, the CRM can now be worked from Claude with permissions the CRM enforces, rather than through a custom integration.
What to do. If you run Salesforce and Claude, ask your Salesforce account team about beta access, and have your admin review which objects and actions the plugin may touch before anyone connects it.
Google Workspace added admin controls for agents and Gemini data loss prevention
What happened (August 26, 2026). Google's August Workspace feature drop added an agent access management dashboard where admins can view and suspend Workspace Studio flows and revoke specific OAuth scopes, such as Drive access, for individual flows. Gemini data loss prevention can restrict what Drive content Gemini reaches, based on content conditions and labels. Teams can also build reusable Studio skills for brand standards and recurring processes. Source: Google Workspace blog.
Why it matters. Most small businesses on Workspace have never reviewed which agents hold OAuth access to their Drive.
What to do. Open the dashboard, review what has access, suspend anything nobody can explain, and label the Drive folders Gemini should not read.
An item makes the digest only if it changes a decision a small or mid-size team would make that quarter, and only with a dated primary source. Model benchmarks, leaks and funding news without a product consequence are left out.
Common questions
What were the biggest AI changes for business owners in August 2026?
The EU AI Act's transparency rules took effect on August 2 while its high-risk rules were delayed to December 2027, Anthropic added shared memory, a built-in browser and general availability of Claude in Chrome, Salesforce and Anthropic announced Claudeforce, OpenAI added Premium seats to ChatGPT Business, and Google and Microsoft shipped new agent and data controls.
What changed in the EU AI Act on August 2, 2026?
The Article 50 transparency duties began to apply, covering chatbot disclosure, marking of synthetic content and deepfake labelling, and the AI Office and national authorities gained enforcement powers. The high-risk obligations were not applied: the Digital Omnibus on AI moved them to December 2, 2027, and to August 2, 2028 for AI inside regulated products.
Does the EU AI Act apply to a US small business?
It can. The rules follow where an AI system is used or its output reaches, not where the company is based. A US business whose chatbot talks to EU customers, or whose AI-generated marketing reaches EU audiences, should follow the disclosure and labelling duties that began on August 2, 2026.
Who controls Claude memory on a Team plan?
The admin. On Team and Enterprise plans, admins decide whether memory is available to the organization, and it stays off for each person until they turn it on. On Free, Pro and Max plans memory is on by default, and each person can view, edit or delete what it holds.
Is this page current?
It was written in September 2026 as a look back at August 2026. Items describe what happened that month. Where something changed afterward, such as Cowork merging into the main Claude app on September 16, the item carries a dated note saying so.