The summer that changed the AI rules, why your AI bill needs a plan not a cap, and how to make an agent finish the job

A look back at the summer shows the ground shifted under every business that buys AI: governments now decide when models ship, costs replaced capability as the boardroom question, and managing agents became a real skill. This week is about what to do with that, from cost architecture to giving an agent a finish line it can check itself against.

• Washington became the release gate for new AI models, and the gap between what labs have and what you can buy has never been wider
  My take: The summer started with the strongest models ever released, Fable 5 and Mythos 5, being pulled within days after a government export letter. OpenAI's next model was announced weeks before anyone could use it. Chinese labs released open models that matched the frontier while the American ones sat behind closed doors, and the big labs are now openly asking the US government to pace releases. The practical consequence for a business owner is that the model you can buy today is further behind the lab's best than at any point in this technology's history, and the terms attached to the best models (retention rules, regional limits, sudden withdrawals) are now set partly by regulators, not vendors. Do not build a workflow that only works with one specific model. Keep the model swappable, test new releases against your own real tasks when they arrive, and put data handling terms above benchmark scores when you pick a vendor. That rule has paid off every month this year.
• The boardroom question moved from 'can AI do it' to 'what does each task cost', and the answer is a stack, not a subscription
  My take: Companies got excited about agents in spring, and by summer the finance teams had noticed the bill. AI is not a 20 or 30 dollar per seat tool anymore. A person running agents can usefully spend hundreds or thousands a month, and the companies that win are the ones who spend that well, not the ones who spend least. Three moves became standard over the summer. Routers that send each task to the cheapest model that can do it (Stripe paid a reported 7 billion dollars for the leading one). Model families, where OpenAI cut its cheaper tiers by up to 80 percent so you keep the flagship for hard work only. And open-weight models you can run yourself, which AT&T and Thomson Reuters now use for large parts of their volume, and which reached about half of enterprise usage on the main open marketplace. If you have one AI plan for everything, you are overpaying. This quarter, list your recurring AI tasks, mark the few that truly need premium intelligence, and ask your vendor what routing or cheaper tiers cover the rest.
• The most useful AI skill right now is giving an agent a finish line it can check itself
  My take: Power users stopped prompting AI step by step this summer and started running loops: give the agent a goal, a way to check whether it is done, and let it work until it gets there. Every agent tool already loops under the hood, but it stops after one polite pass unless you tell it what done means. This came from software, where tests either pass or fail. In normal business work you have to build the referee yourself, and the trick is to be boring. 'Make it insightful' is not checkable. '200 competitor data points, each with a source and date, none duplicated, summary under 150 words' is. Then add a cap, say 30 rounds or two hours, and keep the agent in a sandbox where a mistake costs nothing. Good candidates are research, campaign optimisation where you can measure clicks, compliance checks and content audits. Bad candidates are anything where your judgment is the actual work, like hiring or an executive email. Pick one overnight research task this week, write the finish line before you start, and see what you get in the morning.
• Autonomy has to be earned, and every AI workflow needs a named owner
  My take: A widely shared list of what makes a company 'AI native' had two items that matter more than the rest. First, agents climb a ladder: observe, suggest, act with approval, act alone, and only then run a whole workflow inside a set boundary. You do not hand a new agent your bank login on day one for the same reason you do not hand a new hire the company credit card. Agents should inherit the permissions of the person asking, and those permissions should be enforced in your systems, not in the prompt. Second, and this came from readers rather than the list itself, someone has to own the result. AI can automate the work but not the accountability. Every workflow you automate should have a named person, a measurable goal, and a clear answer to who is responsible when it goes wrong. If you cannot fill in those three blanks for an agent you already run, that is the one to fix first.
• Do not turn your old process into an agent, and plan to rebuild every few months
  My take: The instinct when automating is to map how people do the job today and give that to the agent. The best operators are warning against it. Human processes are shaped by human limits, attention, hours in the day, how many things one person can hold in their head, and agents do not share those limits. Give the agent the goal and the guardrails, not the twelve steps Dave has always followed. Two other habits separated the companies pulling ahead. They treat legal, HR and IT as partners who design the enabling rules up front, rather than a gate at the end, so they are not re-arguing permissions every project. And they expect to throw away and redesign their AI workflows roughly every quarter, because the tools improve that fast. If you built something clever in spring and are proud of it, that is exactly the one to re-examine now. Do not fall in love with a workflow. Fall in love with the outcome it produces and stay ready to get there a cheaper way.