What is an AI stack?
The five layers, three example stacks, and the rules for keeping a stack small.
An AI stack is the set of AI tools a person or team actually uses together — not a list of every product that launched this month. A good stack has one assistant for thinking, one or two tools for making things, one layer for automation, and a place where the work is remembered.
The mistake most teams make is buying five tools that all do the same thing. The second mistake is buying a specialist tool before the workflow exists. This guide is about the smallest stack that covers the work.
The five layers
The assistant
The place you think, draft, and ask questions. Usually ChatGPT, Claude, or Gemini. Pick one as the default so the team builds shared habits.
The makers
Tools that produce a specific artifact: a design, a video, a deck, a voiceover. Add these only when the output is a recurring deliverable.
The connectors
Automation that moves work between systems. Zapier or Make for most teams; n8n if you can self-host and want control.
The memory
Where decisions and context live: Notion, Google Drive, or your existing wiki. AI over your own docs is more valuable than another chat window.
The private layer
For confidential work, a local model or a vendor with a no-training business plan. This is a policy decision before it is a purchase.
Three stacks that work
These are the smallest useful stacks for three common teams. Prices are list prices for the seat count shown.
Solo marketer
Five-person law firm
Three-developer startup
Four rules for keeping it small
One default assistant
Let the team standardize on one assistant for a quarter. Two defaults means two sets of prompts, two sets of limits, and no muscle memory.
Buy the artifact, not the promise
If a tool does not produce a deliverable you already need, it is a experiment, not a subscription. Put it on a trial, not the company card.
Price per seat, then per month
A $20 tool for five people is $100 a month — $1,200 a year. Team plans are often cheaper annually, but only if everyone actually uses them.
Review the bill every month
Every tool should map to shipped work, a closed deal, or saved hours. If it does not, cancel it. The stack should get smaller as the team gets clearer.
Questions
Two to four is usually enough: one assistant, one maker, one connector, and a place for memory. If you have more than five paid AI subscriptions, you probably have overlap.
Start there, but check the limits. A $20 plan that runs out of messages by Wednesday is more expensive than a $30 plan you can use all month.
Only if you are building a product or automating at volume. For normal team work, the app is enough. The API is a separate bill and a separate skill.
They are the right answer for confidential data and predictable volume. You pay in hardware, setup, and maintenance instead of tokens. Most teams should run one cloud assistant and keep a local option for sensitive work.
Build the smallest stack that works.
Two questions, a monthly cost, and a first move. No signup.