Your feed this week is all GPT-5.6. Set it aside for a second.
Saying "we're experimenting with AI" is starting to land like saying "we have a gym membership." Everyone has one. The question is whether you've shown up, and whether you can show what it did for you.
This week, the people signing the checks started asking for receipts.
The board stopped asking whether AI works. Now it asks what it returned.
Until now, the question about AI was whether it worked.
This week the question changed to what it returned.
The sharpest research this week says most teams have been measuring the wrong thing.
The payoff from generative AI isn't showing up as smaller headcount. It's showing up as faster decisions and clearer writing.
Leaders working to justify their spend at the next budget review aren't the ones whose AI didn't work. They're the ones who can't put a true value metric on what they got out of it.
The industry has officially crossed into trillions of dollars of AI spending. This week the coverage turned to one question: what is it all returning?
TechCrunch called it the $3 trillion question. The debate moved out of investor calls and into operating reviews.
The capex story stopped being someone else's problem the moment it hit your budget. Everyone will need an answer to the ROI question sooner than later.
MIT Sloan followed a real organization through its generative-AI rollout. Headcount didn't fall. The work changed shape instead.
Coordination moved out of meetings and into clearer writing. Decisions closed faster. Their conclusion: "hours saved" missed the actual gain.
That's the more defensible scorecard. Not "we cut X hours." Decisions closing faster, fewer meetings to align, clearer writing between people.
Those are measurable and they're actually worth paying for.
Microsoft stood up a new business called Frontier Company: $2.5 billion, about 6,000 people. Their job is getting enterprise AI deployments to actually work. Early clients include the London Stock Exchange Group, Unilever, and Land O'Lakes.
Read what that money admits. The bottleneck isn't the model anymore. It's the gap between buying AI and getting a return from it, and it's wide enough that Microsoft built a company to close it. If the vendor thinks the last mile is the hard part, that tells you where your own gap is.
What to do this week: Before your next budget review, write a one-page AI scorecard: decision speed, communication clarity, coordination meetings removed. Not hours or headcount saved. Walk in with it in hand, instead of assembling it after someone asks "what did we get for it?"
The upgrade you didn't schedule
If your company runs Microsoft 365 Copilot, the engine under your documents changed this week.
Nobody sent a rollout plan. There was nothing to approve. OpenAI shipped GPT-5.6, and it became Copilot's default model in the same stretch of days.
The upgrade arrived on its own. Confirming it didn't quietly change your numbers is the part that's still yours.
OpenAI released its GPT-5.6 family this week. Within the same news cycle, it became the default model in Microsoft 365 Copilot — Word, Excel, PowerPoint, Chat, and Cowork.
The pitch is more capability per dollar and better agent behavior. The delivery was a server-side swap, not a rollout you chose.
Your Copilot got "better" by the vendor's definition, maybe by yours too. But "better on average" and "identical on your specific work" are different claims. What matters is wherever Copilot touches a number or a sentence you'd have to stand behind.
Engineer Armin Ronacher noticed something counterintuitive. Newer, "better" models sometimes handle custom tools worse than the ones they replaced.
His read: they've been tuned so heavily for a vendor's own built-in tools that they get sloppier with yours.
A model upgrade isn't automatically an upgrade for what you built on top of it. A prompt, a workflow, an integration that leaned on the old model's behavior — a smarter model can break it in ways that don't announce themselves. The fix is a short test you re-run every time the vendor bumps the model.
What to do this week: Take your two highest-stakes Copilot outputs, a financial model and a customer-facing document, and re-run them against the new model. Confirm nothing changed that you'd have to explain later. Save those two as the check you run on every model swap.
You can outsource the AI. You can't outsource the risk.
You can buy the AI from someone else. You can't hand them the liability.
If a tool you licensed discriminates, leaks data, or publishes something you were required to disclose, the regulator and the court come to you, not your vendor. That used to be a conference-panel abstraction. This quarter it has a date: the EU's transparency rules start taking effect in August.
Harvard Business Review made a plain argument this week. Delegating an AI system to a vendor doesn't delegate the responsibility.
If the tool discriminates, mishandles data, or harms a customer, the accountability lands on the company that deployed it.
That makes vendor selection a risk decision, not just a procurement one. The useful move is unglamorous: read what your AI-vendor contracts actually say about liability. Most leaders assume the vendor is on the hook. Most contracts say otherwise. You'd find out at the worst possible time.
The EU AI Act's transparency obligations start taking effect in August. That includes disclosing AI-generated content, and every member state standing up at least one AI regulatory sandbox by August 2.
The disclosure wave is already visible. Google said this week it will label ads made with AI.
If you touch the EU market, this is a real deadline with a month attached, not a someday. Someone needs to own it before August: what AI-generated content you publish, where it gets labeled, who signs off. Governance stopped being a topic and became a task with a due date.
What to do this week: Pull your top three AI-vendor contracts and find the clause naming who's liable if the tool discriminates, leaks data, or publishes something you must disclose. If you sell into the EU, put the August transparency deadline on one named person's calendar today.
Of note
Buyers are starting to ask AI who to hire. HBR looked at three small and mid-size businesses adjusting as customers ask an AI assistant "who should I use?" before they ask a person. If the assistant makes the shortlist, your discoverability and structured data just became a sales problem, not only a marketing one.
AI-assisted brainstorms get better and more alike. Across four studies, AI raised the quality of individual creative work but narrowed the range of ideas the group produced. Everyone drifted toward the same good answer. If your team brainstorms with AI, protect the disagreement on purpose.
Alberta put Claude on its security holes. The province used Claude to find and fix cybersecurity vulnerabilities across its systems. Vendor-told, so read the results as directional. But it's the un-flashy kind of deployment worth copying: a real, boring, high-value problem most organizations already have.
The receipts, not the demo
A year ago the impressive thing was watching AI do something.
Today the impressive thing is a leader who can show what it did: the return, the re-check, the line they drew, the risk they own.
The demos got the attention. The receipts are what the room asks for now.