Everyone Grew Up At Once

Something shifted this week and it took me a few days to name it.

For three years the story in AI has been capability. Bigger, faster, smarter, cheaper, and everybody racing to be first. This week the loudest stories weren't about capability at all. They were about codes of conduct, kill switches, pacing agreements, advertising platforms and merger integrations.

The industry put on a tie.

That's not a criticism and it's not a victory lap either. It's just a phase change, and phase changes are when the opportunities move. Here's what actually happened and what to do about each one.

The biggest item for anybody who sells anything.

On September 16 OpenAI announced Sponsored Agents. After somebody clicks an ad inside ChatGPT, they can start a clearly labeled conversation with an agent the business sponsors. They ask their own questions, get answers about fit and specifications and whatever else, and follow a link when they're ready. It's separate from ChatGPT's own answers and separate from the conversation they started in. Testing with select advertisers in the US now.

Bundled with it: natural language campaign management through an Ads Manager plugin, AI generated copy and imagery suggestions drawn from your landing page, optional automatic adaptation of your headlines to conversational context and language, and integrations with HubSpot as the first CRM partner and Shopify as the first ecommerce partner. Shopify merchants in the US can install the app now. It goes international on September 23.

What to do. If you're on Shopify or HubSpot, this is a fifteen minute install that you should do this week purely to see the interface, whether or not you spend a dollar. Early access to an ad platform has historically been worth a great deal and this one's moving fast.

The strategic piece is bigger though. A Sponsored Agent is your business answering questions unsupervised, at scale, in front of buyers. Whatever that agent knows is what your brand says. Which means the asset that matters isn't the ad, it's the source material: your actual specifications, your real objection handling, your honest edge cases, written down in a form a machine can use.

Almost nobody has that document. If you publish anything at all, your archive is already most of the raw material, it just needs reorganising into answers rather than articles. Mine lives in beehiiv and pulling the answers out of three hundred issues took one long Sunday. The businesses that write it in the next quarter are going to have an unfair advantage in a channel their competitors are treating as a media buy.

Microsoft released an AI code of conduct for its products running roughly fifteen thousand words, and announced it's adding kill switches to its AI offerings.

Nobody's going to read fifteen thousand words. The kill switch part is the signal.

What to do. Read it as market intelligence, not policy. A company with Microsoft's legal department shipping an off button means they ran the failure scenarios and concluded that structural stops beat procedural ones. That's a conclusion you can borrow without doing the work.

If you run automations that can act, you need one place to halt everything. Not a plan, a switch. We went deep on this Monday and I won't repeat it, but if you skipped that one, this is your second nudge in a week from two very different directions.

Dario Amodei published an essay called "We Must Pace the Frontier" arguing for slower frontier development on safety grounds. Within about two days, OpenAI, Google, SpaceXAI and Microsoft had all endorsed the approach. Sam Altman separately told Fortune there'd be no OpenAI IPO in 2026, citing safety, and floated the idea of labs agreeing to pause at certain capability thresholds.

Axios raised the read everybody was already thinking: frontier labs are burning extraordinary amounts of money on compute, and a coordinated slowdown would conserve cash while protecting incumbents from cheaper and open competitors. Both things can be true. Genuine safety concern and convenient economics aren't mutually exclusive and rarely are.

What to do. Don't change your plans. Whatever happens at the frontier, the models available to you today are already far beyond what your business has operationalized, and that gap is where your entire opportunity lives.

The one thing worth noting: if the frontier does slow while open weight models keep closing the distance, the cost curve for what you use gets better, not worse. Bolt's first forty eight hours of open model data this week had GLM 5.3 Flash taking about half of all usage among eleven million builders, with DeepSeek and Kimi K3 splitting most of the rest. People pick fast and cheap when fast and cheap is good enough, and it increasingly is.

Covered this Monday so I'll keep it to the facts. A threat actor built exploits for two PaperCut flaws and used AI agents to automate the intrusions. GreyNoise counted at least 440 instances across 395 organizations in 48 countries, with eleven organizations compromised in twenty six seconds during one burst.

What to do. Audit which of your automations read text from strangers and then take actions that can't be undone. That's the whole assignment and it's an afternoon.

Profound raised a $180 million Series D at a $1.8 billion valuation, led by Sequoia and Kleiner Perkins, building a marketing platform for the AI era.

What to do. You don't need their product. You need the insight that a $1.8 billion valuation implies: serious money believes that whether AI systems mention your business is about to be a measurable, manageable marketing channel.

The free version takes twenty minutes. Open ChatGPT, Gemini and Perplexity. Ask each one the three questions a real buyer in your category would ask. "Best [your category] in [your city]." "What should I look for when hiring a [what you do]." "[Competitor] alternatives."

Write down what comes back. Are you in it? Is what's said about you correct? Is a competitor being described in a way you could truthfully claim?

Do it again in ninety days. That's your baseline, that's your tracking, and it costs nothing.

GitHub launched agentic automations inside GitHub Actions for triage, documentation and code quality.

What to do. If you have any technical operation, this is the cheapest possible place to run your first real production agent, because the blast radius is contained and the work is genuinely tedious. Start with documentation or issue triage, not code changes.

The pattern is the point though, and it generalizes. Platforms are embedding agents into the workflow tools you already pay for rather than making you buy a separate agent product. The same thing happened this week with ChatGPT rolling into Word across all plans including free. The standalone AI tool category is quietly getting eaten by the software you already own.

The exception worth building yourself is the connective tissue between those embedded agents, which is still nobody's product. That is where Make.com keeps earning its subscription for me.

Check what you're already paying for before you buy anything new. There's a decent chance the capability you're shopping for shipped into your existing stack last month and nobody sent you an email about it.

Perplexity's Portable Computer agent arrived in its Windows app for Nvidia RTX and RTX PRO cards with at least 24GB of memory. It runs locally rather than in somebody's cloud.

What to do. Most of you don't have a 24GB card and this isn't a buying recommendation. Note the direction instead, because the direction matters for a decision you'll face within a year.

Local agents solve the problem that keeps regulated businesses and anybody with real client confidentiality out of this entire category: the data never leaves the machine. Right now that capability costs a workstation class GPU. The hardware requirement has fallen roughly in half every twelve months for three years running.

If you've been sitting out AI adoption because your data genuinely can't go to a third party, and some of you legitimately can't, put a reminder in your calendar for next spring. The thing that's blocking you is a hardware price, and hardware prices move.

Connecticut's Comptroller announced five new rules governing how insurers can use AI on the state employee health plan and its partnership plans.

What to do. One state, one plan, one industry, and the reason it's on this list is that it's the shape of what comes next.

AI regulation is not going to arrive as one big federal law that everybody reads about. It's going to arrive as a hundred specific rules from specific regulators about specific uses in specific industries, most of which will never make the news. Insurance, lending, hiring, healthcare, housing, education.

If you operate in any regulated vertical, the practical move is to know who your regulator is and read what they publish, because the thing that catches people isn't the rule they disagreed with. It's the rule they never saw.

And keep a plain record of where AI touches a decision that affects a customer. Not a compliance program. A list. When somebody eventually asks, having the list is most of the answer.

Cohere and Aleph Alpha completed their merger, targeting global enterprise AI implementation.

What to do. Nothing immediate, but note the direction. Consolidation at this layer means fewer independent vendors and more full stack offerings, which historically means better integration and worse pricing leverage for buyers.

If you're evaluating any AI vendor on a multi year contract right now, build in an exit. The odds that your vendor is independently owned in twenty four months are not great, and the terms you sign today are the terms you inherit after the acquisition.

The sixth edition of McKinsey's Technology Trends Outlook landed, tracking fourteen technology trends across patents, investment and talent demand.

The line that matters: AI is generating breakthroughs faster than organizations can absorb them.

What to do. Take that seriously as permission. The constraint in your business isn't access to capability, it's absorption. You are not behind because you haven't adopted the newest model. You're behind, if you're behind, because the three things you adopted last year never got properly wired into how work actually happens.

Absorption beats acquisition. Every time, and it's not close.

Here's what connects all of it.

A year ago, the winning move was to be fast. Try everything, adopt early, tolerate the mess, because the capability gap between you and a slower competitor was enormous and growing.

That gap is closing, and not because anybody slowed down. It's closing because capability is getting commoditized into the tools everybody already has. ChatGPT in Word. Agents in GitHub Actions. Ads in HubSpot. When the capability arrives inside the software you were already paying for, having it stops being an advantage.

What's left as an advantage is the boring layer. Whether your processes are written down clearly enough for a machine to execute. Whether your product data is complete. Whether your objection handling exists as a document rather than living in the head of your best salesperson. Whether you can tell when something breaks.

That's not a fun answer. It's not a new tool you can buy on Saturday and feel good about. But it's where the next two years of advantage gets built, and this week the whole industry quietly agreed with that by spending its news cycle on governance, integration and distribution instead of raw capability.

The tie is on. Time to do the unglamorous work.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

Pick one item from the list above. Exactly one. The one that made you slightly uncomfortable rather than the one that seemed most interesting, because discomfort is usually your operation telling you where the gap is.

Spend an hour on it. Not a project, an hour.

Then, if you do nothing else this weekend, run the twenty minute AI visibility check. Three tools, three questions, write down what comes back. It's the cheapest baseline you will ever establish and in a year you'll wish you'd started today rather than then.

Jordan

The AI Newsroom | Practical AI for people with a business to run.