The AI Transformation Brief—September 25, 2026
The AI Transformation Brief
// Today’s Signal
Enterprise AI is moving from a collection of impressive features into a managed business system with clear rules. Microsoft is extending security controls to AI software on devices and in cloud workflows. Databricks is putting a live spreadsheet inside its business data platform with built-in rules. BNP Paribas is scaling multiple cloud providers and deploying multiple AI models across a 180,000-person bank. Google is making real-time visual interaction a business interface. OpenAI is turning workspace administration into detailed access settings. The strategic shift is clear: model access is becoming interchangeable, while authority, identity, data context, and proof of work become the scarce assets. The winners will own the system that lets AI act safely across the company.
// Top Stories
Microsoft’s September security update says AI software now runs on employee devices, cloud platforms, and across developer workflows.
The company is adding capabilities to discover and control AI software running on employee devices, extend strict identity and access checks to traffic from AI software acting for people, and strengthen security operations for work involving AI. Microsoft Security Blog The update also describes Microsoft Purview and Microsoft Entra Global Secure Access bringing data security to network traffic involving people and AI software acting for them. Microsoft positions its security operations foundation inside Defender as a base for security around AI software and brings tools that collect and connect security events together with threat protection. Microsoft Security Blog
The company’s security boundary is moving from the employee and the application to the software that can act for a person. That changes the buying decision from “Which AI tool do we approve?” to “Which identities, data paths, and actions can we observe and stop?” A management layer that sees AI software across devices, cloud services, and developer workflows becomes more valuable than another isolated detector. This quarter, inventory every AI system with permission to read, write, send, purchase, or deploy, then test revocation and recovery against one live workflow. Microsoft Security Blog
Databricks announced its acquisition of Row Zero, a cloud spreadsheet that connects to live, business data with permissions and usage rules and supports collaboration between people and AI software.
Databricks says Row Zero will give Genie, its AI assistant, a native spreadsheet experience for finance, operations, sales, and marketing teams. Databricks The planned experience combines familiar formulas, pivots, and keyboard shortcuts with the Databricks Data + AI Platform, Genie Ontology, Databricks’ business-context system, Unity Catalog, Databricks’ permission and data-management system, and Unity Gateway, Databricks’ AI traffic and spending-control system. Databricks says Row Zero can work across billions of rows at interactive speeds, honor user permissions, keep data connected to authoritative sources, restrict exports, and make actions taken by AI software interpretable and auditable. Databricks
The spreadsheet is becoming the meeting point between business judgment and AI work with built-in rules. Databricks is not adding a prettier front end to a data lake; it is trying to make the place where decisions are modeled the same place where permissions, context, and evidence are enforced. That pulls finance and operations closer to the company’s AI management layer and weakens the old separation between analytics, workflow, and action. Put one recurring planning process into a live spreadsheet with built-in rules this quarter, and measure decision latency, rework, cases where someone needs special access, and the evidence kept after each change. Databricks
BNP Paribas and Google Cloud announced a five-year partnership covering Google Cloud infrastructure, Gemini models, and Gemini Enterprise.
BNP Paribas says its internal LLM@CIB assistant is already available to more than 65,000 employees, while its Nickel business uses Gemini as the base model for Nickel Assist, which supports 200 customer advisers. BNP Paribas Initial Corporate & Institutional Banking use cases include AI support for corporate credit memos, with further applications planned across sales, trading, research, and structuring. The bank says it operates in 64 countries with more than 180,000 employees, and that its security model verifies each AI system, limits access to the resources needed for its assigned task, and monitors its connections and interactions. BNP Paribas
The strategic signal is not that a bank selected another model vendor. It is that a regulated enterprise is treating model choice as a portfolio decision inside a common security and data framework. The bank can add a new model or cloud path without surrendering the rules that define which work may run where. That is the difference between buying AI features and building an enterprise capability. This year, separate model selection from business authority: require every material use case to show where the data is stored, which model handles the work, who is responsible for the result, the human approval point, and the tested way to stop the system. BNP Paribas
Google says Gemini 3.8 Live with Live Avatar is available in Gemini Enterprise.
The feature combines live dialogue, video generation, speech, visual input, background tool calls, and a responsive visual persona. Google lists customer service and interactive walkthroughs as enterprise use cases. Google Google says Live Avatar can move across 97 languages, supports custom avatars through enterprise allowlisting, and embeds SynthID watermarks in generated audio and video. The company presents these controls as safeguards for identity, transparency, and responsible deployment. Google
The interface is becoming a strategic choice, not a cosmetic layer. A live visual AI system can change the economics of customer support, training, and guided service, but it also makes identity, consent, brand control, and escalation visible to the customer. The scarce asset will be the trusted operating design around the avatar, not the avatar itself. Before any public launch, define the approved persona, disclosures, source data, escalation route, and record of what the system said or did, then test the experience in one high-volume workflow. Google
OpenAI’s September 24 enterprise release notes describe two new global-admin permissions for ChatGPT Sites, including whether Sites can use members’ connected software tools and whether applications can access ChatGPT Ads.
The permissions are in admin preview and off by default. OpenAI Help Center The same release notes describe Voice support for plugins and connected software tools across web, iOS, and Android. OpenAI also says Voice in Work can create documents, presentations, and spreadsheets, use connected software tools, work in a browser, and continue an unfinished task in text after a call. OpenAI Help Center
Enterprise AI is becoming a permissions problem before it becomes a model problem. The important product surface is now the policy that determines which people, applications, connected business systems, and external data can participate in a task. This moves responsibility from the AI team to identity, legal, security, finance, and business owners. Treat every new connected software tool as a decision to let software act for a person: name the owner, define the allowed action, record the evidence, set an expiry or review date, and test removal before enabling it broadly. OpenAI Help Center
// Shelly Palmer Pulse
Shelly Palmer describes YouTube Studio as an AI packaging department that can critique a draft, generate channel-matched thumbnails, test up to three video cuts, and monitor performance.
YouTube says creators have run more than 40 million title and thumbnail experiments since 2024, but it did not provide evidence that the new tools improve creator metrics. Palmer’s angle aligns with today’s enterprise read: once AI can generate, test, and change customer-facing work, approval rights and contractual boundaries become operating controls rather than after-the-fact review. Shelly Palmer
// What It Means For Your Business
The durable asset is shifting from access to models toward a managed system that connects people, data, rules, work, and evidence.
This quarter, select two material workflows and name the accountable executive, permitted actions, approval points, evidence kept, and recovery path before expanding authority. Microsoft Security Blog BNP Paribas
Model access is becoming easier to substitute while distribution with clear rules, identity, trusted context, and customer-facing interfaces become the scarce layer.
Security vendors, data platforms, cloud providers, and model companies are competing to own the path between a business intention and a verified action. Databricks OpenAI Help Center
Trust will be sold through visible control and verifiable outcomes, not through claims about model intelligence.
Make the customer promise specific: what the AI can do, what it cannot do, how a person can stop it, and what record the customer receives afterward. Google Shelly Palmer
Redefine productive work as a completed, reviewed result linked to the person or software that produced it.
Build a standing cadence across operations, security, legal, HR, and technology to maintain test cases, permission policies, exception logs, and stop procedures as business processes that use AI change. Microsoft Security Blog OpenAI Help Center
Fund AI programs against cost per verified completion, rework, failure recovery, and capacity or revenue unlocked.
Do not treat AI activity, connected-app count, or amount of content generated as value. Databricks BNP Paribas
Separate the model from the systems that authenticate data, route requests, enforce policy, record events, evaluate behavior, and manage recovery.
Standardize identity, records of where data or content came from, permission, export, and requirements that make it possible to change AI providers across AI companies. Microsoft Security Blog Databricks
Require management to report which AI tools and software can read confidential data, alter records, contact customers, move money, or deploy code, plus who can revoke those permissions and how quickly.
Treat permission for AI to act as a governance decision, not a software purchase. BNP Paribas OpenAI Help Center
The most consequential shift is that AI is becoming a managed business system with clear rules across identity, data, interfaces, and workflows. The broken assumption is that the model or application vendor owns the whole enterprise path from capability to action. The decision now is whether your company builds a portable control system that can route work across models, enforce authority, and prove the result.
Are you still buying AI applications one department at a time, or are you building the company-wide system that decides what software is allowed to do?
The Transformation Brief is written daily by Les Ottolenghi. Delivered every morning at 6:00 AM MT, a 7-minute read on the AI shifts that matter to operators and boards.
**Build the system that lets AI tools work together. Price the outcomes. Redesign the organization.**
Read On
The AI Transformation Brief—October 14, 2026
Banks, card networks, and regulators are deciding who answers for AI that acts for customers, and companies that wait will inherit their ...
Read the brief →The AI Transformation Brief—October 11, 2026
Gatekeeping decides who benefits from AI, and this weekend shows four different gates. Microsoft's CEO argues the limits on what a model ...
Read the brief →The AI Transformation Brief—October 10, 2026
Permission is becoming the operating constraint on AI. Anthropic is cutting live-internet access for internal AI tests until its monitoring ...
Read the brief →