The New Blendtec PM Problem

Sometimes, even in 2026, using less AI is the right move for PM leaders.

There’s a stage in every PM’s AI journey where you feel like you’ve cracked it. You’ve graduated from casual chatbot prompting to actually deploying agents with real context and automating their tasks: meeting transcripts, product briefs, Slack threads, the whole knowledge graph of your work life. The outputs are impressive, the time savings feel real, and you start to wonder why you ever did anything without it.

And then something quietly goes wrong.

I hit this wall a few weeks ago while working on an H2 planning doc. I’d spent a couple of hours on a first draft, developed a clear outline, and had a real point of view on where we were going and why. Then I handed it to an AI and started doing what felt like responsible, rigorous iteration: reading the output, catching mistakes, flagging hallucinations, pushing for revisions, convinced that staying engaged with my agent in every round meant I was doing this right.

By the fourth or fifth pass, I barely recognized the document. The original structure had quietly dissolved, new framing had crept in from nowhere, plausible-sounding but disconnected from anything I actually believed or had evidence for, and the sections I hadn’t touched had morphed to accommodate the sections I was fixing. What started as a tight, opinionated planning document had become something more like a well-formatted approximation of one, a document that looked like rigorous PM thinking but had the intellectual texture of consensus mush.

The harder lesson wasn’t that the AI made mistakes. It was that I had been so focused on catching the local errors that I missed the global drift, and what I thought was steering had actually been an extended, gradual handoff of the wheel.

---

This is the thing nobody tells you about the drift problem: it’s not just a technical limitation you can prompt your way out of, it’s a structural property of how these systems work. Every revision is probabilistic, every pass through the document is an opportunity for the model to reweight what matters, smooth out edges you actually wanted to keep, and introduce elegant-sounding connective tissue that subtly rewrites your premises. You can try to constrain it by telling it to touch only section three and leave everything else alone, but at that point you’re doing so much scaffolding and oversight that you might as well just edit the section yourself, which, it turns out, is often faster.

The deeper problem, the one I think the PM community isn’t being honest enough about, is what this workflow does to your thinking over time. Writing a tight planning doc, a strategy brief, a product narrative that will get pressure-tested by executives and engineers and skeptical cross-functional partners, isn’t just about the output. It’s about what happens to your thinking in the process of producing it. When you write every line yourself, you discover the gaps in your own logic, find the places where you don’t actually know what you believe yet, and earn the ability to defend it in a room because you built it rather than reviewed it.

Prompting is a genuinely different cognitive act, one that takes real skill and judgment but doesn’t produce the same depth of ownership as drafting something line by line. If you’ve only ever directed an AI to write your H2 planning doc, you’re going to feel that absence the first time someone in the room pushes back hard on the strategy rationale and you realize you can describe the prompt for the document better than you can defend the document itself.

---

I want to be clear that I’m not writing this as an AI skeptic. I use it constantly, and I genuinely can’t imagine working without it for large categories of my job: triaging and drafting communications, synthesizing research, producing executive briefings on topics I need to get up to speed on fast, running competitor analyses, doing the first-pass fact-checking on something before I commit to an assertion. These are all cases where AI has made me meaningfully more effective, and I’m not going back. The leverage is real and the tools are only getting better.

But I’ve had to develop a cleaner mental model for when AI is the right tool and when it isn’t, and for me the dividing line runs roughly here: AI is extraordinarily good at recalling, summarizing, and synthesizing things that already exist, and it’s a liability, or at minimum a very high-maintenance collaborator, when the job is to produce something that needs to be original, logical, and defensible.

For canonical PM documentation, the kind of document that travels up and down the org, anchors a quarter or a roadmap or a strategy pivot, and will be read by people specifically looking for evidence that you understand the problem deeply and have thought it through rigorously, AI is your sidecar rather than your driver. You can use it to pressure-test your logic, fact-check your claims, and tell you when a paragraph is unclear or a section is structurally weak, but the drafting, the actual sequence of choices about what to say and in what order and with what emphasis, has to stay in your hands, or you risk producing a document that nobody truly owns.

---

Being an AI-first product leader is the right goal, and I believe that fully. But AI-first doesn’t mean AI-always, any more than getting a Blendtec for your kitchen means everything in your refrigerator goes in the Blendtec. The appliance is remarkable, handling things you couldn’t have imagined doing in your kitchen before, but if you put in the wrong ingredients, or the right ingredients at the wrong moment, you don’t get a better meal, get a very uniform mess.