Nine Weeks In: What Changed in AI This Summer
A recap of This Week in AI, mid-May through mid-July 2026.
Every week since May, we've written down what happened in AI and what it means for the business owners we work with. Read one week at a time, each post is a snapshot. Read together, they tell a different story than any single headline did on its own.
Here's the short version: AI spent the spring being priced like a subscription and treated like a novelty. Over nine weeks, it got billed like infrastructure, regulated like a utility, and audited like a hire. The businesses that are ahead right now aren't the ones with the newest model. They're the ones who stopped waiting for the tools to get easier and started doing the unglamorous work of running them well.
Here's how we got there.
Mid-May: Accountability shows up before the tools are ready for it
The first story we covered wasn't about a new model. It was about a lawsuit. Pennsylvania sued Character. AI for the unauthorized practice of medicine, after a chatbot claimed to be a licensed psychiatrist. The same week, the White House drafted an FDA-style vetting process for frontier models, and Microsoft quietly reconsidered its 2030 clean-energy commitment because AI's power demand had made the math impossible.
None of these were really about the technology. They were about who's accountable for what it does. That question arrived earlier than most businesses had budgeted for.
A week later, Anthropic launched Claude for Small Business, HubSpot shipped a tool to measure whether your business shows up in AI-generated answers, and one of OpenAI's founding researchers joined Anthropic's pre-training team. Different stories, same direction: AI stopped being a side tool and started becoming the layer things run on. Infrastructure, not an app.
Late May into June: The industry starts pricing itself honestly
By the end of May, OpenAI had filed confidential IPO paperwork at a valuation of up to a trillion dollars, while reportedly losing more than a dollar for every dollar it earned. Anthropic was reportedly targeting its own IPO by fall. A Gartner survey found that 80 percent of companies that cut staff after adopting AI saw no corresponding lift in returns, not because AI didn't work, but because they automated the output without redesigning the workflow underneath it.
Going public forces a kind of honesty that venture funding doesn't. So does a bad ROI number. Both showed up the same month.
Then came a story that's stuck with us since: Anthropic published data showing that 80 percent of the code in its own production codebase is now written by Claude. But the detail that mattered wasn't the output. It was what sits on top of it: a mandatory automated reviewer, because raw AI output still isn't trusted to ship, even at the company that builds the model. In the same week, a survey of 740 enterprise executives found that only 10 percent had moved AI agents past the pilot stage, mostly because they didn't trust the output enough to act on it without a human checking. And GitHub Copilot quietly ended flat-rate pricing, moving every plan to a metered, pay-as-you-go model.
The pattern: the companies building AI are using it on themselves productively, because they built serious review processes around it. Everyone else is either stuck in pilot or about to get a bigger bill.
Mid-to-late June: The ground moves on more than one axis at once
One week in mid-June, the U.S. government ordered Anthropic to pull its most capable public model over an export-control concern, and it went dark for every customer within days. No warning, no say. The same week, new data found that automated traffic had passed human traffic on the web for the first time. And a PwC jobs report found that AI is raising the value of human judgment, not erasing it. AI-exposed entry-level roles increasingly require senior-level skills, and they're growing faster than the roles that don't touch AI at all.
Three unrelated facts, one shared lesson: the tools, the traffic, and the skills your business depends on are all less stable than they felt six months ago.
Around the same time, ChatGPT opened a self-serve ad platform: no minimum spend, contextual targeting based on what someone is asking about, not who they are. And by late June, the costs that had been invisible started showing up on paper. FERC ordered the regional grid operators serving two-thirds of the country to justify how AI data centers are affecting electricity rates, and a Reuters study found people are increasingly getting news from AI chatbots and almost never clicking through to check the source. A separate MIT study found that leaning on AI for answers made people measurably worse, a month later, at spotting what's true on their own.
The jobs data that same week undercut the layoff narrative directly. Gallup found only 1 percent of laid-off workers pointed to AI as the reason. Workers who rarely used AI were more likely to be let go than the ones who used it regularly.
July: AI stops feeling free, and the bill for setup comes due
By early July, the free-feeling era was over on three fronts at once. A fintech employee ran up an $81,267 bill in a week just prompting Claude to build a browser game, the visible edge of a much bigger shift toward metered, token-based pricing across the industry. ChatGPT's ad platform scaled to 900 million weekly users, with roughly one in five queries showing direct commercial intent. And a Salesforce survey found AI customer service agent adoption had jumped from 39 to 66 percent in a year, with the best deployments automating the routine and keeping a human reachable for anything emotional or high-stakes.
Then, the story that ties the whole nine weeks together. In the same week, Ford brought back hundreds of engineers into roles automation was supposed to cover. Commonwealth Bank of Australia and Klarna both reversed AI-driven cuts to customer service after results slipped. IBM found its AI could handle 94 percent of routine HR requests. It's now tripling entry-level hiring because the remaining six percent, the judgment calls, is where everything else was breaking. Meanwhile, Microsoft launched a $2.5 billion business unit whose entire job is helping companies deploy AI well, and AWS committed a billion dollars to the same kind of work. Two of the largest companies on earth just put a price on the gap between buying AI and benefiting from it. The price is billions.
What this means
Look at the nine weeks together and the throughline isn't any single technology story. It's a question that kept getting asked in a different costume every week: who's accountable for what this does, and who's doing the work of making it work?
The first half of the summer was about that question arriving: in courtrooms, in IPO filings, in a jobs study that punished companies for skipping the redesign. The second half was about the cost of ignoring it becoming visible: in power bills, in a model going dark overnight, in a week where a company burned $81,000 on a browser game. And the most recent stretch has been about the correction. The companies that cut first and thought later are rehiring, and the companies selling AI are now selling the implementation work around it, because that's where the value was all along.
None of this is an argument against using AI. It's an argument against treating it like a subscription you install and forget. The tool was never the hard part. It never has been.
What business owners should do
Audit before you cut. If you're considering using AI to reduce headcount, map the workflow first. Every company that cut and later regretted it skipped this step.
Name your single points of failure. List the AI tools your business genuinely depends on. If one went dark tomorrow, and one did this summer, do you have a working backup?
Put a number on your AI spend and give it an owner. Token-based pricing is becoming the default, not the exception. Review the bill monthly, the way you would any other line item.
Test how you show up in AI answers. Ask ChatGPT, Gemini, and Perplexity the questions your customers would ask. This is still cheap and organic. It won't stay that way.
Keep a human on anything emotional, high-stakes, or judgment-heavy. Automate the routine. Every company in this recap that got hurt, got hurt by automating something that needed a person.
If any one week's worth of this feels like a lot, that's fair, it was a lot. But the same lesson showed up nine times in nine weeks: the businesses doing well with AI right now aren't the ones with the best tool. They're the ones who did the unglamorous work of setting it up right.
This Week in AI publishes weekly at bampt.co/learn. You can also catch it on Substack at bampt.substack.com and on Instagram @bamptco.