The AI agent will send the email before you realize it should not have
MIT Technology Review — Published: 2026-05-06
Agentic AI tools can now chain together lead follow-up, CRM updates, and invoice generation with no code and no IT department. That is genuinely useful for a lean team carrying too much admin weight. The risk that gets underplayed in the coverage is what operators are calling approval drift: you start reviewing every output, then every other one, then you stop because the last fifty looked fine, and that is exactly when the tool sends a discount offer to your most profitable account without context. (The honest caveat: most bad AI outputs are recoverable. The ones that hit a customer relationship or a financial record are not.) Start small. Stay in the loop longer than feels necessary. Worth Doing: Identify one internal, low-stakes workflow this week, status summaries or lead draft responses. Deploy the agent with mandatory human sign-off for the first 30 days before expanding scope to anything customer-facing or financial.
Your customer survey is telling you what they wish were true about themselves
Behavioral Scientist
Customers describe their ideal behavior. They buy according to their actual one. The gap between the two is not dishonesty. It is the natural distance between intention and action, and most SMB operators are building pricing decisions, service add-ons, and product features on the wrong side of that gap. (This applies with particular force to any survey run during or after a sales conversation, where social desirability bias is highest.) The fix is not more surveys. The fix is treating every stated preference as a hypothesis that needs a behavioral test before it becomes a budget line. Said another way, don’t take your customers words ad verbatim, do verify what they said is what they intended to say. And that is the post-survey hypothesis that is critical. Because any forward looking investment or expansion plans or change in your strategy coming out of a survey carries the risk of you getting trapped in their social desirability bias. Quite simply you might end up building for the persona they wish they are, than they are.
Worth Doing: Pick one decision currently driven by customer self-reports. Design a 30-day live test with a single, observable success metric: purchase rate, usage frequency, or repeat rate. What customers do is the data. What they said is just the hypothesis.
Your annual strategy plan is already wrong. The question is how long before you notice.
MIT Sloan Management Review
MIT Sloan's research makes a pointed argument: in volatile conditions, the cadence of your decision-making matters as much as the quality of your decisions. Annual planning cycles create a dangerous lag. By the time a problem shows up clearly enough to trigger a formal review, the window for a cheap fix has usually closed. The operators who are outperforming are not smarter. They are faster at recognizing when an assumption has expired. (This does not mean abandoning planning. It means treating your plan as a hypothesis you are actively trying to break, not a commitment you are defending.) Build the feedback loop before you need it. Plan to look for signals to know whether the markets are turning materially adverse or materially opportunistic for your business, and recalibrate.
Worth Doing: This month, define three to five metrics your team reviews monthly, one experiment to run, and a written rule for when to kill it versus scale it. Time the first session for within the next 30 days.
Lattice Briefs is written for the people actually running Main Street businesses. If that sounds like someone you know, send a link to this newsletter their way.
New readers can subscribe at latticebriefs.com/subscribe