ISSUE 05 · SEPTEMBER 8, 2026
AI is all we do. We keep up so you don't have to.
WHAT WE'RE SEEING IN THE MARKET
The fastest returns on AI agents come from three decisions made in advance
A new survey of 2,025 companies running AI agents found the fastest returns went to teams that did three unglamorous things before they turned anything on: cleaned up the data an agent actually touches, drew a narrow line around what it's allowed to do, and decided ahead of time when a person steps in. Unglamorous work, evidently the kind that pays off. More on why that matters below.
THE WHY
The real training gap
The training gap is real, and it's the mistake I see most often with mid-market companies. Picture a 40-person ops team at a distributor that rolled out Claude to everyone six months ago. Logins were up, usage was up, adoption looked great on paper.
But ask any manager there what "good with AI" means for a warehouse coordinator versus a customer service rep, and you'd get a shrug, the corporate version of "I'll know it when I see it." Nobody had written it down. So training stayed generic, one session on prompting basics that everyone forgot by Friday, instead of something built around what each role actually needs.
The Salesforce survey behind this week's market box backs this up. The fastest returns come from three decisions made in advance: clean data, a narrow job for the agent, a clear point where a person steps in.
Microsoft's research adds the human piece. When a manager actively used AI in front of their team, and expected people to use it too, employees reported far more trust and far more real impact. Login counts and usage stats can look great and still miss the point entirely. A team can hit 100% adoption and still have no idea what they're supposed to be getting better at.
The fix is smaller than it sounds: decide what "good" looks like for the role, then have someone actually show it in front of people.
YOUR MOVE
Ask this before you schedule the next training session
This week, ask whoever leads your most AI-active team, ops, support, sales, doesn't matter which, one question: what would someone need to be able to do with AI for you to call them AI-competent?
Not something vague like "uses it sometimes." Something specific, like drafting a first-pass customer response that only needs light editing, or summarizing a vendor contract's key terms correctly on one read. Write that sentence down.
Then ask yourself the harder version: have you shown your team you can do that yourself in the last 30 days?
Whatever you find, build your next session around that one sentence instead of a generic "AI basics" deck. People actually use training built around something specific.
WHAT TO SKIP
Low usage numbers don't mean you need a bigger AI budget
If proficiency looks thin this quarter, it's tempting to reach for a better model or more licenses. Most of the time the simpler fix works better. It means figuring out what "good" looks like for the role, then getting whoever leads that team to actually show it. That closes more of the gap than a bigger budget ever will.
THE ROUNDUP
Signals from the noise this week.
GPT-6 Astra: A new generation of intelligence (OpenAI): OpenAI's newest model just shipped, and it's the first of theirs to cross their "critical" cybersecurity capability threshold, which is exactly why they built new internal safeguards on top of the Hugging Face lessons before letting it out the door. Worth knowing the ceiling moved again, even though we build on Claude. Read it here
Developing Enterprise Frontier Safeguards with our customers (Anthropic): If your team uses Claude for anything sensitive, this is the feature to ask about. Activity logs used for misuse detection can now stay in your own cloud instead of Anthropic's, which matters if you've got compliance requirements to answer to. Read it here
We still don't know how people are really using AI (MIT Technology Review): A new independent research project found company usage reports miss a lot of how people actually use AI day to day. A good reminder before you make a decision off a vendor's usage stats, or your own login numbers. Read it here
Agentic AI Study: Preparation Beats Speed for ROI (Salesforce): The full data behind this issue's numbers, straight from the source. Worth a look if you want to see exactly how they define "meaningful ROI" and which factors predicted it best. Read it here
Want help figuring out what "good with AI" should look like for your team? Let's talk it through.
Got a question about any of this, or want a second opinion on something you're building? Just hit reply, I read every one myself.
Justin
Why of AI