Data & AI watch, week of 28 September to 4 October 2026

The September Power BI feature summary, published on 29 September, confirms the timeline: from October, Power BI Desktop versions from March 2026 or earlier lose the ability to save and share files to OneDrive and SharePoint, as the old file picker is retired. Microsoft is also moving subscription emails to a new delivery platform; every subscription transfers automatically, with no action needed on your side.

The zero-provisioned F0 capacity and on-demand billing across Fabric reach preview "in the coming weeks". We covered why it matters in our FabCon takeaways. Already generally available: workspace-level surge protection, which automatically isolates secondary workloads, paid capacity overage with thresholds and alerts, and the F4096 and F8192 SKUs.

In Copilot Chat you can now search for a Power BI report, attach it to the conversation and ask questions about it, with Power BI permissions and Purview DLP rules applied (generally available). The reverse integration, Microsoft 365 Copilot inside Power BI and Fabric, is in preview, off by default and limited to Microsoft 365 Copilot licences. A useful new setting lets a semantic model editor restrict Copilot access to users with more than read-only rights. It stays open by default.

Power BI "agentic experiences" are generally available: skills published on GitHub (skills-for-fabric), an authoring MCP server that lets an agent edit a semantic model open in Desktop, and a bridge that reloads Desktop after each change. A remote version of the MCP server, usable on Mac with nothing to install, is in preview. The PBIP project format, text-based and versionable in Git, is generally available too.

DirectQuery semantic models can now be created and edited in the Power BI service, collaboratively and from Mac or Linux. In preview, two DAX functions, TEXTCONTAINS() and TEXTSIMILARITY(), add typo-tolerant, meaning-based full-text search (compatibility level 1708, in the service first). Also in preview: approximate distinct count for Import and Direct Lake, and a new Power Query editor that shows, step by step, what is folded to the source. Export to Excel grows from 150,000 to 500,000 rows (generally available).

Microsoft offers a guided Azure Data Factory upgrade: ADF pipelines appear in Fabric, a compatibility assessment comes before any switch, and migration proceeds pipeline by pipeline while existing workloads keep running. Assistants also cover Spark (Synapse) and data warehouses, with incremental replay of changes. The Data Engineering agent (preview) adapts notebooks and pipelines, then validates the result.

At its DevDay on 29 September, OpenAI launched the Agents API in public beta (hosted execution, memory, tools, interface control) and Dots, persistent agents with their own virtual machine, offered on Pro, Business and Enterprise plans outside the European Economic Area, Switzerland and the UK. On 30 September, GPT-6.1 Sol shipped at GPT-6 Sol pricing ($2 input and $10 output per million tokens), with cached input at half the price. The announced performance, close to GPT-6 Astra, comes from the vendor.

On 30 September, the US Federal Trade Commission confirmed an investigation into OpenAI, Anthropic and the evaluator METR, covering deceptive practices and data security. It follows incidents in which agents under test reached real systems, including Hugging Face in July. Formal information requests are expected in the coming weeks. The day before, the White House announced a voluntary agreement with six vendors (OpenAI, Anthropic, Google, Nvidia, Meta and xAI), based on external audits, with no timeline or penalties attached.

The Google Threat Intelligence Group reports that monthly vulnerability disclosures doubled between January and August 2026, and that half of AI-discovered vulnerabilities allow remote code execution, compared with 26% of the others. A BeyondTrust flaw was exploited four days after disclosure. AI components themselves are targeted: Google cites exploited flaws in LiteLLM and Langflow, and attacks on Flowise, vLLM and Ollama.

Both vendors support Apache Ossie, a project incubating at the Apache Software Foundation, which lets you describe business measures and rules once and reuse them across platforms. A converter published on GitHub turns an Ossie document into both a Snowflake Semantic View and a Power BI semantic model. Microsoft aims to have DAX recognised as a query language of the standard.

The Claude Frontier Academy combines a few days of in-person training with a twelve-week residency, during which each engineer leads a real Claude project in their organisation. First participants: Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk and Commonwealth Bank. The programme runs in San Francisco, New York and London, aiming for 10,000 engineers by the end of 2027.