Confirmed
OpenAI launches a Data agent in ChatGPT Work that connects to company databases and builds dashboards
OpenAI launched a Data agent inside ChatGPT Work on September 10, connecting to approved sources including Redshift, BigQuery, Snowflake, Databricks, MongoDB, and ClickHouse to answer business questions in plain language. It reads an organization's semantic layer, business terms, and metric definitions from tools like dbt and Databricks Genie Ontology, then builds interactive dashboards in Tableau, Power BI, Sigma, and other business-intelligence tools. Admins control which connections and roles get access, and queries enforce the account's table, row, and column permissions.
In plain termsInstead of writing a database query, an employee just asks a question in plain English and gets a chart back.
Why it mattersIt's the enterprise version of the same connector pattern already pulling product feeds and analytics into consumer AI answers, worth watching as those data pathways formalize.
Our takeEvery business-intelligence vendor already sells natural-language querying, so that part isn't new. What's different is that permissions and semantic layers travel with the connection, so the agent inherits governance instead of reinventing it, the harder problem most AI-data pitches skip.
Confirmed
OpenAI opens its Agents API in public beta, renting out the harness that runs Codex
OpenAI opened its Agents API in public beta on September 10, giving developers the same harness and infrastructure that runs Codex: a managed session that keeps context across long-running work, searches for relevant tools instead of loading them all, and can delegate pieces of a task to parallel subagents. Developers choose where the agent runs, an OpenAI-hosted sandbox, their own infrastructure, or a partner like Cloudflare or E2B, and pay only for tokens and tool use, no separate API fee.
In plain termsIt's the toolkit OpenAI uses to run its own coding assistant, Codex, now rented out so other companies can build agents without building that plumbing themselves.
Why it mattersFor any team building an agent that reads your site, a shopping assistant, a research bot, this is a rentable alternative to building session and tool-routing logic from scratch.
Our takePublic beta from a company OpenAI's size means enterprise buyers get an agent-hosting decision months before smaller vendors can compete on it, and every 'is this good' question needs an eval score, not a demo, to answer.
Confirmed
AWS ships a plugin connecting Amazon Redshift and its data lake tools to OpenAI's Data agent
AWS announced a Data Analytics plugin for ChatGPT Work's Data agent on September 10, letting employees query Amazon Redshift clusters and Serverless workgroups, plus AWS Glue Data Catalog, S3 Tables, and Athena, using natural language instead of SQL. The plugin translates questions into Redshift-compatible queries, builds shareable interactive dashboards, and supports follow-up questions that refine or drill into results. It runs against a customer's data models and governance rules, and respects each employee's existing access permissions.
In plain termsIt plugs Amazon's own data warehouse into the same ask-a-question agent OpenAI launched the same day.
Why it mattersAlongside BigQuery and Snowflake, this is a second confirmed data source wired directly into a conversational AI product, one more place structured business data now answers questions without a dashboard in between.
Our takeThree cloud data warehouses shipping the same connector on the same day reads as a coordinated launch, not three companies independently reaching the same conclusion, and coordinated launches usually mean the platform (OpenAI) set the timeline.
Observed
A working paper finds Google's AI Overviews cut Wikipedia's search referral traffic by up to 5.45%
A University of Washington working paper by Mehrzad Khosravi and Hema Yoganarasimhan finds that default availability of Google's AI Overviews reduced monthly external-search referrals to English Wikipedia by 5.45% relative to German Wikipedia and 4.82% relative to French, using Wikimedia's own clickstream data and a difference-in-differences design across matched article pairs. The paper, last revised September 2, 2026, has not been peer reviewed. Its abstract frames the finding as evidence that answer-producing intermediaries can materially reallocate attention away from publishers.
In plain termsWikipedia gets fewer visitors from search once Google's AI summary answers the question on the results page itself, and this paper puts a specific number on how much fewer.
Why it mattersCloudflare and Wikimedia's own traffic data already pointed the same direction, and this paper adds the published methodology and baseline that direction lacked: AI Overviews measurably reduces referral traffic to the source it answers from.
Our takeI flagged in 'Google's billions-of-clicks claim' that Wikipedia's own traffic data already showed an 8% year-over-year decline in human visitors. This paper adds a causal estimate for AI Overviews specifically, with a published baseline and denominator, the kind of methodology Google's own click figures still lack.
Deep dive: Google's billions-of-clicks claim: why it can't be checked against your own site data
Confirmed
Google Merchant Center adds a Product videos section for uploading short product videos
Google added a Product videos section to Merchant Center's Video assets page, letting merchants upload short product videos mapped to a single product offer each. For YouTube videos, merchants can accept, modify, or reject AI-suggested product mappings, and AI-generated videos require accepting Google's Product Studio terms before saving. Google's support page gives no launch date but says more video features are planned.
In plain termsMerchants can now upload short product videos straight into the same feed system that already handles their product photos and prices.
Why it mattersPulled into Shopping and organic surfaces, product video is now a Merchant Center asset type, one more feed input that determines whether a listing shows up rich or bare.
Our takeVideo assets sit in the same eligible-not-distributed bucket as every other Merchant Center attribute: uploading a compliant video gets a product past the format check, but whether Google actually surfaces it in Shopping results is a separate, unpublished decision.