May 24, 2026 · ← Notes
A "healthy AI stack" sounds like marketing language, but it has a real definition. Here's the version I work with when I audit stacks. It's not exhaustive, but it's the closest thing to a rubric I've found.
This isn't a hard ceiling — some businesses legitimately need more. But every tool past 12 adds non-linear coordination cost. Twelve tools is six interfaces to switch between mentally, twelve billing dates to track, twelve places customer data could leak, twelve vendors who could change their pricing model.
If a tool hasn't been opened in 30 days, it's not in your stack — it's in your billing statement. The two are different. Healthy stacks are made of working tools, not paid tools.
Writing, support, CRM, automation, analytics — pick one primary per category. The "I use ChatGPT for some things and Claude for others" pattern is fine as a personal preference, but if you're paying for both at the team tier, you're paying twice for the same job.
When one tool dominates the budget, you've lost negotiating leverage and you have a single point of pricing risk. Vendor changes their tier structure and your stack economics break.
Most operators have never put their stack on paper. Until you do, you can't see where the data lives, where it duplicates, or where it leaks. The exercise of drawing it is itself diagnostic — you'll discover gaps just by trying to draw the diagram.
For solo operators this is trivially true. For small teams, it's often not — knowledge gets distributed and nobody can see the whole thing. The fix isn't documentation; it's appointing one person whose job includes knowing the whole stack.
Most stacks I see fail four or five of these criteria. That's not unusual; it's the default state of an AI stack that grew organically over 18 months. The point isn't to score 6/6. The point is to know which ones you're failing and decide if the failure matters.
The most common failure: too many tools, no primary in each category, and no map on paper. Those three usually go together. Fixing one tends to fix the others, because the act of mapping forces you to name the primaries, which forces you to delete the duplicates.