September 9, 2026
Instacart launched a conversational shopping assistant called Clementine this week. Shipt launched one too, with almost identical pitch language: tell it you're hosting a tailgate for 25 people, or that you need easy school lunches, and it builds the cart for you. Neither launch is a shock. It's the third or fourth version of this exact idea we've seen from a consumer app in recent months, and that repetition is the actual story. When two direct competitors ship the same feature in the same week, it means the technology has stopped being a differentiator and started being a cost of doing business. If you run a business that touches consumers at all -- retail, food, services -- the lesson isn't 'we need a chatbot.' It's that conversational AI is becoming an expected front door, and the companies who treat it as a bolt-on widget instead of a real extension of their CRM and ordering logic will feel it in customer expectations within a year, not five.
At the same time, TechCrunch reported that AI spend per employee actually fell at top firms in August. That's worth sitting with. Falling token costs and cheaper models explain part of it, but 'cheaper' and 'less adopted' are different problems, and this data doesn't cleanly separate them. If enterprises are spending less per employee because efficient models let them do more for less, that's a healthy market maturing. If they're spending less because pilots stalled out after the initial excitement, that's a warning sign hyperscalers should take seriously. My honest read: it's probably both, unevenly distributed by company. Either way, this is exactly why businesses evaluating AI tools should be skeptical of vendors selling seat licenses by headcount rather than tools tied to actual workflows -- see our own thinking on workflow automation for why usage-based value beats subscription bloat.
Jacob Coxon, a researcher at Anthropic, resigned this week and said publicly that self-improving AI amounts to 'gambling with our lives,' calling for pacing agreements between labs. This lands the same week Sequoia doubled down on Cymphony, a startup addressing security risks created specifically by AI agents operating inside companies, now valued north of $100 million after a $25 million Series A. Put those two stories side by side and you get the real tension in this industry right now: the same labs racing toward more autonomous, more capable systems are also generating the enterprise security problems that justify a new category of startup to clean up after them. I don't think Coxon's resignation should be read as fringe alarmism -- when someone inside a leading lab walks away over this, it deserves more weight than the usual outside critic. But it's also fair to note that Anthropic and its peers have safety teams and public commitments of their own, and one departure doesn't settle whether pacing agreements are realistic or just a nice idea nobody enforces. For a business owner, the practical takeaway is narrower and more useful: if you're adopting AI agents that act on your behalf -- placing orders, touching customer data, managing accounts -- you need the same rigor around permissions and oversight that Cymphony is selling as a standalone product. That's not exotic; it's the same principle behind role-based permissions done right, just applied to a new kind of user that happens to be software.
So which is it for your business: is AI adoption cooling off because the easy wins are already banked, or because the harder, riskier deployments -- the ones that actually touch money and customer data -- are the ones companies are hesitating on?
Sources