OpenAI’s ‘Now everyone can put data to work’ (10 September 2026), Anthropic’s ‘Developing Enterprise Frontier Safeguards with our customers’ (1 September 2026), and RuntimeWire’s 10 September report on Claude for Windows identity controls all point at the same SME decision: who may see company data, where logs live, and which Windows client is approved.
This week’s ChatGPT Work Data agent, Anthropic’s customer-held safeguards, and the shadow-AI / operating-model pieces in the 11 September newsletter give UK firms a concrete checklist: approved plugins, named data owners, and a human review point before an agent acts.
EM-01: OpenAI’s Data agent turns warehouses into a ChatGPT Work question
On 10 September OpenAI introduced the Data agent in ChatGPT Work. It connects to approved company data, investigates what changed, and builds shareable dashboards. Administrators control plugin availability and data-source plugins. AI Build can configure a ChatGPT Business workspace so the Data plugin only sees named sources.
Source: OpenAI, 10 September 2026 — https://openai.com/index/put-data-to-work/
EM-02: ChatGPT Work is an agent that drafts files and needs approval gates
The same day’s Business release notes introduce ChatGPT Work for longer tasks across apps and files, including documents and presentations, and replace the App Directory with Plugins. Compare Business vs personal on admin controls and approval of consequential actions, not just model names. OfficeMaker remains the path for branded, reviewable decks when the output must match a house style.
Source: OpenAI, 10 September 2026 — https://help.openai.com/en/articles/11391654-chatgpt-business-release-notes
EM-03: Shadow AI is now a legal as well as security problem
TechRadar’s piece, carried in the 11 September round-up, treats unapproved tools and sensitive uploads as both a security and EU AI Act issue. An approved ChatGPT Business workspace, simple usage rules and an owner for exceptions is safer than a ban staff will ignore. AI Build can write the short policy and exception path.
Source: TechRadar — https://www.techradar.com/pro/shadow-ai-is-a-security-problem-but-the-eu-ai-act-makes-it-a-legal-one
EM-04: Governance needs a RACI, not another principles slide
Ethosure’s operating-model article argues that ownership, escalation and review make governance usable. A 20–200 person firm can assign data, prompts, approvals and incidents without a full three-lines-of-defence programme. AI Build’s governance review starts there.
Source: Ethosure — https://ethosure.ca/the-ai-governance-operating-model-from-principles-to-structure/
EM-05: Windows AI clients are an identity and procurement decision
RuntimeWire’s 10 September report on a Claude for Windows build describes Entra Continuous Access Evaluation and a Windows authentication broker. Treat it as a distribution signal: verify against Microsoft and Anthropic primary docs before changing a standard. Ask how identity and sovereign-cloud paths work for each approved client.
Source: RuntimeWire, 10 September 2026 — https://runtimewire.com/article/microsoft-let-anthropic-move-in-while-openai-was-sleeping-on-the-couch
EM-06: Anthropic will keep misuse logs in the customer’s cloud
On 1 September Anthropic announced Enterprise Frontier Safeguards: activity data for misuse detection can live in customer-controlled storage, with a phased autumn rollout and zero data retention on Fable 5 / 5.1 for eligible customers until then. Ask where logs live and who reviews flags before putting Claude on client or finance data.
Source: Anthropic, 1 September 2026 — https://www.anthropic.com/news/enterprise-frontier-safeguards
EM-07: Agent oversight and sovereign AI remain an open discussion
The New Stack’s Bluesky post, included in the 11 September round-up, is a discussion signal not a measured trend. The useful questions for an SME remain: where does processing happen, which integrations are on, and what happens when the agent is wrong.
Source: The New Stack via Bluesky — https://bsky.app/profile/thenewstack.io/post/3mufbnsiu7b
EM-08: Trustworthy data still decides whether a Data agent is useful
MIT Technology Review’s 12 August piece on scaling agents with trustworthy data, still in this week’s newsletter, is the pairing for EM-01. Dashboards only help if the sources are trusted. Start with one bounded workflow and a named data owner before connecting more systems.
Source: MIT Technology Review, 12 August 2026 — https://www.technologyreview.com/2026/08/12/1141032/scaling-ai-agents-with-trustworthy-data/
What UK SMEs should do next
Pick one repeatable workflow. Record the approved ChatGPT Business workspace, the data sources the Data plugin may use, the owner, the spend or action limit, and the human review point. Measure time saved and error rate for two weeks before expanding.
Make AI delivery measurable and controlled
AI Build helps organisations identify high-value AI use cases and put the right governance, security, training and implementation controls around them. Start with an AI maturity assessment, an AI governance review or a conversation about a managed ChatGPT Business rollout.