Stop buying AI tools. Buy the workflow that connects them.

There's a particular kind of founder who shows up to our first call with a list.

OpenAI. ElevenLabs. Pinecone. n8n. Make. Zapier. HubSpot. Notion. Airtable. Slack. Plus three more they're "evaluating."

They've spent $400/month on subscriptions. They've done demos. They've watched YouTube tutorials. And — critically — they've automated exactly one thing: their receipt of SaaS receipts.

This is the most common pattern I see. And it's the wrong one.

The problem with "tool collecting"

Tools don't automate your business. Workflows automate your business. A workflow is a sequence: trigger → process → output. Tools are the bricks; the workflow is the building.

When you collect tools without workflows, you end up with:

• App sprawl. Twelve subscriptions, two used weekly, one used daily.
• Manual glue. Copying data from OpenAI into HubSpot into Notion by hand. (Real example from a call last month — the founder had a 14-step manual process to move one lead from chat to CRM to onboarding.)
• Fragmented data. Customer info lives in 4 places. None of them is the "source of truth."
• Brittle processes. When the person who set up the Zap leaves, no one knows how to fix it.

The founder's instinct is to buy another tool to fix this. The actual fix is to delete three tools and connect the remaining ones properly.

The 4-step framework I use

I won't build anything for a client until we walk through these four steps on a whiteboard. The order matters.

Step 1 — Discover: where is the business leaking?

Before talking about AI, we map the actual business process. We list every place where:

• Time is being lost (manual work that should be automated)
• Leads are being missed (calls, emails, DMs that don't get a fast reply)
• Revenue is being left on the table (no-shows, churn, no follow-up)
• Data is being entered more than once (the kiss of death for productivity)

This step usually surfaces 8–12 candidates for automation. Not all of them are worth automating. The next step decides.

Step 2 — Design: minimal viable automation

For each candidate, we ask three questions:

1. How often does this happen? (Daily > weekly > monthly > one-off. Don't automate one-offs.)

2. How much time does it cost? (10 minutes/day = 40 hours/year. Below that, probably not worth the build.)

3. Does it have a clear input and output? (If you can't describe both in one sentence, the workflow isn't ready to automate.)

We pick 1–2 workflows to start. Never more. Building one workflow that actually runs is worth more than 5 half-built ones.

The other key decision: where does the human stay in the loop? Most workflows shouldn't be fully automated. The "80% auto, 20% human approval" pattern is usually the right balance. It's also the safest.

Step 3 — Build: connect, don't create

This is where most projects go wrong. People try to build custom AI for everything. That's slow and expensive.

The right approach is to use existing tools as much as possible:

• For AI: OpenAI for text, ElevenLabs for voice, Vapi for voice agents, Pinecone for retrieval. Don't build your own model.
• For workflow logic: n8n or Make. Don't write custom code if a node can do it.
• For storage: Supabase or Postgres. Don't roll your own.
• For CRM: HubSpot or Pipedrive. Don't build a custom CRM.

The job of a "build" is to wire these together with webhooks, API calls, and a small amount of custom logic. If your build is mostly glue code, you're doing it right. If it's mostly new infrastructure, you're probably over-engineering.

A typical first project takes 5–14 days. If it's taking longer, the scope is wrong.

Step 4 — Optimize: automation is a compounding asset

After launch, the workflow runs. People stop thinking about it. That's a mistake.

Every workflow generates data — logs, latency, error rates, conversion rates. That data tells you what to optimize next:

• If the AI voice agent routes 30% of calls to a human, that's a prompt-tuning signal
• If the content approval queue rejects 40% of drafts, that's a brand-voice-prompt signal
• If no-shows are still 25%, that's a reminder-timing signal

Monthly tuning (1–2 hours) keeps a workflow healthy. Without it, the workflow drifts — the AI prompt becomes outdated, the CRM schema changes, the API breaks. Within 6 months, a workflow that wasn't tuned is usually broken.

The takeaway

If you're a small business and you're trying to "do AI," don't start with tools. Start with a leak.

Pick one process where you can quantify the loss in dollars or hours. Map it. Design the smallest possible workflow that fixes it. Wire together existing tools. Launch in 2 weeks. Tune for 90 days.

Then — and only then — pick the next leak.

A small business with 2 well-tuned workflows will outperform a small business with 12 half-built ones every time.

The math doesn't lie. And it doesn't care how many SaaS logos you can fit in a footer.