The AI power-user gap explodes, using AI to write goes mainstream, and picking the right model becomes a real skill

The gap between businesses that use AI agents well and everyone else has quadrupled since January, a famous investor made AI-assisted writing publicly respectable (with caveats), and companies are learning to mix cheap and premium models instead of paying flagship prices for everything.

• The gap between top AI users and average users grew from 2.6x to 8.3x in six months
  My take: New enterprise data from OpenAI shows the top 10 percent of business users now consume 8.3 times as much AI as the average firm, up from 2.6 times in January. The driver is agents: work delegated to AI that runs on its own, rather than chat back and forth. And the fastest-growing users are not engineers. Legal teams grew their agent usage 108x since February, sales 41x, finance 20x. The compounding part is what should get your attention: firms that figure out agents get faster at figuring out the next thing, so the lead widens every month. If your company's AI use is still mostly summarising meetings and drafting emails, the practical move is to pick one team, give them a real budget and permission to experiment, and have them automate one complete workflow end to end. Waiting for this to settle down is the one strategy the data says does not work.
• A billionaire investor got caught using AI to write his op-ed, and said 'of course I did'
  My take: Stanley Druckenmiller published a Wall Street Journal op-ed that was obviously AI-written, and when called out, refused to apologise: 'I write everything using AI now for the same reason I use a calculator.' The Journal's opinion editor backed him. This is the moment AI-assisted writing became publicly normal at the highest level, and in two years nobody will care that you used it. But the episode also showed what people will still judge you for: laziness. The backlash was not really about AI, it was about him not bothering to remove the obvious tells, which read as not caring about his own argument. The rule for your business writing: use AI freely for emails, summaries and first drafts, but do the thinking yourself first (the argument, the numbers, the structure), and always do a final pass to cut the robotic phrasing. Some companies are now writing this into policy: you must stand behind every sentence you send, and if you generated a document from a two-line prompt, just send the prompt instead.
• Businesses are building model 'stacks' instead of paying flagship prices for everything
  My take: AT&T now serves 40 percent of its internal AI queries with cheap open models and plans to push that to 60 or 70 percent, keeping the expensive frontier models only for hard tasks like coding. Using a router (software that automatically sends each task to the cheapest capable model) cut their AI coding costs 56 percent while quality dropped only 2 percent. Meanwhile OpenAI cut its top model's API prices by roughly a third, and premium agent products dropped from 500 dollars a month into the 60 to 100 dollar range. The takeaway: the era of one model for everything is over. You do not need to become a model connoisseur, but you should ask two questions this quarter: which of your AI tasks actually need premium intelligence (probably fewer than you think), and does your vendor offer routing or cheaper tiers for the rest. The savings are real and the quality cost is now small.
• OpenAI solved the enterprise privacy problem that was blocking its rival's best model
  My take: Anthropic's most powerful model, Fable 5, has seen weak business adoption for a specific reason: it comes with a 30-day data retention requirement (a government safety condition), which many companies simply cannot accept for sensitive data. OpenAI just shipped the counter-move: automated safety scanning that works without keeping your prompts at all, so businesses get frontier AI with zero data retention. Two lessons here. First, if AI vendors' data policies have been blocking projects at your company, re-check them now, because this constraint is actively being engineered away. Second, this is a live example of why data handling terms, not benchmark scores, often decide which AI vendor wins an enterprise deal. When you evaluate AI tools, put retention and privacy terms at the top of the checklist, not the bottom.
• Data centers became the most toxic issue in American politics, and it will touch your AI bill
  My take: Opposition to nearby data centers hit 75 percent in polling, up from 51 percent in February, and governors who championed AI investment a year ago are now competing to restrict it. Pennsylvania signed the strictest rules in the country. At the same time, chip prices are rising up to 17 percent on memory shortages, and those costs will be passed through to cloud customers. For a business owner two things follow. Near term, expect AI infrastructure costs to face upward pressure into next year, which makes the model-stack discipline above more valuable, not less. Longer term, watch whether your region lands on rules or moratoriums, because states that block construction outright will push investment, jobs and eventually cheaper local capacity elsewhere. The good news buried in the polling: most voters prefer clear rules over bans, so this is more likely to slow the build-out than stop it.
• Voice mode and 'watch me work' are the two AI features worth trying this month
  My take: Two newer capabilities are changing how heavy users actually work. First, live voice mode: instead of typing prompts, people talk to their AI continuously while doing other things ('send those 15 client emails and check them against the list, I'll start on the script'). Users describe it as having an assistant in the room rather than a tool on the screen. Second, teaching by demonstration: newer agent products can watch you do a task once on screen and then repeat it, which unlocks automating workflows that were too fiddly to explain in writing. Neither requires technical skill, both take an hour to trial. If you have been meaning to move past chat-window AI, these are the two lowest-effort places to start, and starting matters, because the compounding gap in item one is built from exactly these small adoption steps.