Issue #1 — The PM AI Stack

Your AI feature is going to get killed in the next reorg. Not because it doesn't work. Because nobody can answer one question about it.

The question is: what workflow does this own?

If you can't answer that in one sentence, your feature is a line item in someone else's roadmap. And in 2026, line items are the first thing cut.

Here is what changed while you were shipping last quarter. Buyers stopped grading software on features and started grading it on workflows completed end-to-end. The Gartner number making the rounds — fewer than 5% of enterprise applications today have embedded task-specific AI agents, projected to hit 40% by end of 2026 — is real, but the more useful framing comes from Bain. Bain maps every AI-enabled product onto two axes: how much of the workflow does it automate, and how much of the workflow does it control? Products with high automation but low workflow control get compressed into features of someone else's product. That is the death zone, and most of what shipped in 2024–2025 is sitting in it.

Three product launches this month make the shift concrete:

The pattern: every meaningful AI launch in April 2026 is a workflow product. None of them are features. If you are still pitching your AI work as "we added X to our product," you are pitching the version of the market that ended last year.

Here is the test I want you to run against every AI initiative on your roadmap. Three questions. Answer them honestly.

1. Who owns the outcome when this works?

Not who built it. Who is accountable when an agent mis-routes a payment, mis-classifies a ticket, or sends the wrong email. If the answer is "the user" or "we'd flag it for review," your feature is assistive, not autonomous. That is fine — assistive features have a market — but you need to know which one you are building, because the pricing, GTM, and moat are different for each.

2. What does the user stop doing because of this?

The strongest test of a workflow product is subtractive. A support tool that surfaces tickets faster is a feature. A system that resolves tickets without human input replaces the tool entirely. If your feature doesn't let the user stop doing something, it is augmentation. Augmentation is harder to defend than replacement, because the moment a foundation model gets 10% better, your differentiation evaporates.

3. If we removed this from the product tomorrow, would anyone cancel?

This is the cancellation test, and it is the one PMs avoid because the answer is usually no. Most AI features added in the last 18 months would not trigger a single cancellation if removed. That is the definition of a feature that gets compressed. If your honest answer is "no," your work is not protected. The reorg is coming for it.

If your AI work fails the test above, you have two options. You can keep building it as a feature and accept that it competes on UX polish — a fight you will lose to whichever competitor's eng team is faster. Or you can reposition it as a workflow.

Repositioning isn't rebuilding. It's reframing what you own. Three moves work in practice:

Expand the boundary. A "smart summary" feature is a commodity. The same model output, framed as "the meeting-to-CRM-update workflow," owns territory. The model is the same. The product is different.

Take the next step. If your feature stops at a recommendation, push it to action. The gap between "we suggest you do X" and "we did X, here's the diff for your approval" is the gap between feature and workflow. It is also the gap between $20/seat and $200/seat pricing.

Own the failure mode. Workflow products own their errors. Feature products surface them to the user. If your AI feature has a confidence score that the user has to interpret, you don't own the workflow. If your feature has a fallback path, an SLA, and a rollback story, you do.

If you are doing one thing as a result of reading this, make it this. Open your roadmap. Pick the AI feature you are most proud of shipping in the last six months. Run the three questions against it. Be honest.

If it passes, write the workflow positioning into your next exec update before someone else writes a worse version of it.

If it fails, you have a choice to make in Q2. Reposition or sunset. The third option — keep shipping it as-is — is the one that gets you written out of the next planning cycle.

When I ran this test against my own roadmap last quarter, I had to sunset two features I was proud of.

The PMs who get through the next 18 months are not the ones with the best AI features. They are the ones who can answer the workflow question in one sentence, every time, for everything they own.

The Pragmatic Engineer's March 2026 developer survey is the most-cited data set in PM circles right now. Around 95% of developers use AI tools weekly, and 70% use two to four AI tools simultaneously. If you are making roadmap decisions about developer tooling, this is the baseline.

The PM AI Stack is a weekly newsletter for product managers shipping AI features in B2B SaaS. If a colleague forwarded this, you can subscribe at pmaistack.

Next week: SAP just blew up the per-seat model. What it means for your AI feature pricing.