Every week somebody tries to sell a small business an "AI platform." The pitch is always the same: one login, one invoice, one magical dashboard. And it's almost always the wrong first move, because what a growing business actually needs isn't a platform. It's a stack. Five boring layers, stacked in the right order, most of them cheaper than your coffee budget.

Here's the whole thing, bottom to top.

Layer 1: Your systems of record

Your CRM, your calendar, your accounting software, your inbox, your project tool. You already own this layer. The only question that matters is whether the data in it is accurate, because AI built on messy records automates the mess. Before anyone sells you intelligence, clean the filing cabinet it will read from.

Layer 2: The plumbing

Integration tools like Zapier, Make, and n8n move information between your systems without anyone copying and pasting. This layer existed long before the current AI wave, and it's still where most of the ROI lives. A lead form that creates a CRM record, notifies your phone, and schedules a follow-up is not glamorous. It also never forgets.

Layer 3: Rented intelligence

The models themselves: GPT, Claude, Gemini, and a long tail of open-source options. Notice the word rented. You don't buy models, you call them, and the price of calling them keeps falling. Inference costs have dropped hard enough that features which were experiments two years ago are now rounding errors on an invoice. Model choice matters less than everyone thinks; the layer above matters more.

Layer 4: The assembly

This is where the value actually shows up, and where the work is. The assembly layer connects rented intelligence to your plumbing and your records: a knowledge assistant that reads your documents, an intake workflow that drafts quotes, an agent that answers the phone and books the appointment. Nothing here is a product you buy off a shelf. It's a system somebody designs around how your business runs.

Layer 5: The adult supervision

Monitoring, logging, and a written policy for what AI may and may not touch. Small businesses skip this layer, then quietly stop trusting their automations the first time something goes sideways. A system nobody trusts is a system nobody uses. Logs, human checkpoints, and a monthly review keep the whole stack honest.

What you don't need

  • A GPU cluster. You are renting intelligence, not manufacturing it.
  • A data lake. Your data already lives in tools with APIs.
  • An "all-in-one AI platform" subscription. That's usually layers 2 through 4 with a markup and a lock-in clause.
  • A head of AI. You need one accountable owner and a partner who explains things in plain English.

Build the five layers in order, spend the least you can at each one, and expand what proves itself. That's the whole architecture. Everything else is a keynote.