This issue covers news from the week of September 7, 2026.
Listen to the “podcast” style version of this issue below!
Generated with Google’s NotebookLM.
Most of the AI decisions in front of you have already been made.
Not by you.
Your platform vendor picked a model. Your staff picked their own tools. Gartner priced the layoffs your competitors are making. Agents started working by the week instead of by the question.
All four are still reviewable.
Table of Contents
Your platform vendor is quietly making your model decision for you
This used to be an evaluation your IT team ran. Now it arrives bundled with the platform.
Salesforce named a model partner at Dreamforce. Microsoft shipped a tool days later to pull your CRM onto its own stack.
Both are deciding something on your behalf. One of them already has your data.
Salesforce announced the arrangement under the name Claudeforce.
It arrived bundled with Agentforce 360 and a long run of customer-success news, which is where most of the coverage went.
If you run Agentforce, part of your model strategy was set in a room you weren’t in. That isn’t sinister and it isn’t unusual. It’s invisible, and invisible decisions don’t get reviewed.
Dynamics 365 Activate went into public preview.
It reads your Salesforce data, customizations and dependencies, then plans the migration off them.
Switching costs are the reason most CRM decisions stay settled. A competitor is now using AI to lower them on purpose. Whether or not you wanted it to be, your platform is a live decision again.
What to do this week: Send your CRM or suite account team two questions in writing. Which model powers the AI features we’re using, and what would it take to run a different one. The second question is the real one.
The AI your company bought is not the AI your company is using
You know roughly what you pay for AI. The number you don’t have is what your people actually reach for.
That gap has been a policy nuisance for a couple of years. It just turned into a permissions problem, because the newest consumer AI doesn’t want your questions.
It wants your inbox.
MIT’s GenAI Divide study found that about 40% of organizations had bought official large language model subscriptions.
At more than 90% of those same organizations, employees were already using personal AI tools for work.
Those are two numbers about one company. One is what you approved. The other is what’s happening. The second has never been measured at most organizations, and it’s sitting in logs you already own.
Meta launched Muse, a consumer AI agent that sends email and books travel on your behalf.
The old version of shadow AI was an employee pasting text into a chatbot, and the risk was what they disclosed. This one asks for the mailbox and the calendar. The risk is what they delegate.
What to do this week: Pull the list of AI domains your staff reached in the last 30 days from your identity provider or network logs, and set it next to what you’re paying for. You already have the data. Nobody has asked for it.
The AI layoffs are going to get reversed, and the reversal is already priced
Every couple of decades a company discovers it outsourced something it actually needed. The call center. The maintenance crew. The people who knew why the system was built that way.
It comes back eventually, and it costs more than it saved.
Gartner has now put a date on the AI version. One in three roles, back by 2029, at a higher price.
Gartner forecast that 30% of the positions eliminated in AI-related layoffs will be refilled by 2029, at a higher cost.
Its analyst Tori Paulman said executives will look back and find the mistake was believing automation was the point, when amplification was the opportunity.
A dated forecast with a premium attached is a rare thing to put in front of a board. It doesn’t argue that AI can’t do the work. It argues that a lot of companies are about to find out what the work actually was.
Salesforce published its own account of rolling agentic coding out to 15,000 engineers, on production systems rather than a greenfield side project.
It’s the company’s story about the company, so read it at that discount.
The number survives the discount. Fifteen thousand people were given a new practice instead of being counted as a saving.
What to do this week: Before you sign off on any AI-justified headcount reduction this quarter, ask for the list of tasks the AI actually absorbed. In writing, specific, set next to what a person still does. If nobody can produce that list, what you have is a forecast, not a finding.
Your first long-running agent will still be working three weeks from now. Who’s checking on it?
Salesforce is selling agents that pursue goals over days and weeks, learn new skills and hand work to each other. OpenAI will now run the infrastructure for you.
So the question stops being whether you can launch an agent.
It becomes who reads its output, who owns its mistakes, and what date somebody looks. Most organizations can answer the first two.
Salesforce announced a portfolio of agents across sales, service, commerce and the workforce.
What it’s selling is duration. These agents are meant to pursue a goal over days and weeks, pick up new skills, and pass work to each other.
An agent that answers a question is a feature. An agent that works for three weeks is closer to a role. Roles come with an owner and a review date. Neither of those ships in the box.
OpenAI released the Agents API, a managed service that runs cloud agents on its own infrastructure.
Orchestration, long-running sessions and tool use are handled on their end.
The practical reason not to try an agent was that someone had to build the plumbing and then babysit it. That reason is gone. What’s left is governance, and governance is the part you can write down without buying anything.
What to do this week: Before your next agent runs, name the person who owns its output, and put a date on that person’s calendar to read a full week of the agent’s work end to end. If you can’t name the person, the agent isn’t ready.
Of note
OpenAI stopped selling its $200 tier. OpenAI paused new Pro subscriptions on September 11. GPT-6 Astra had launched two days earlier, and demand outran capacity. Availability, not price, turned out to be the contract term worth asking about.
Free AI licenses for government. OpenAI and the GSA will give eligible federal, state, local and tribal governments $0 license fees and 50% off usage, with cyber defense support attached. If you’re public-sector or adjacent to it, check eligibility. Everyone else just got a number to raise at renewal.
A $5,000 fine for unchecked output. A New Mexico lawyer was fined $5,000 for filing documents in a murder case that cited witnesses the AI had invented. There isn’t much to add to that.
What didn’t make the issue
An Anthropic researcher resigned with a public warning that AI development has gotten out of control. It led the Journal, the Times, the FT, Bloomberg and the AP. You have almost certainly seen it.
It isn’t in here as a theme because it doesn’t change anything you do Monday morning.
That’s not a dismissal. The question is real and it’s unresolved. It also isn’t yours to settle.
The four above are, and not one of them costs anything.