Most "AI stack" posts are lists of forty tools with affiliate links.

This is the opposite. It is the short list of what I use to run several businesses, what each one is for, and the rules that decide whether a tool earns a place. I'm not going to quote what any of it costs. Your usage will differ and prices change. The rules matter more than the receipt.

The stack, by job

I sort tools by the job they do, not the brand.

The thinking and writing layer: Claude

Claude is the main tool. I use it for strategy drafts, reading long documents, writing and editing, and working through problems out loud. I also use Claude Code, the coding agent, to build and maintain the scripts and sites behind my businesses.

Two habits make it work better:

  • It reads from a written source of truth. My notes, decisions and project context live in a notes vault (Obsidian), and I point the AI at the relevant note before it starts. It doesn't rely on whatever it remembers from last week.
  • I match the model to the job. Mechanical work like formatting, bulk edits and boilerplate goes to a cheaper, faster model. Planning, review and anything where a mistake is expensive goes to the stronger one. Context is the real cost, not the output, so I keep sessions short and end them when the task is done.

The models behind client builds: Claude and GPT

In client builds I use Claude or the GPT models through the API, depending on what the workflow needs. Which one is better changes by task and by month, so I'd rather test on the actual job than trust a benchmark. If someone tells you one is simply better for business, they are selling something.

The plumbing: Make.com and GoHighLevel

When a client needs follow-up that runs on its own, the tools I reach for are Make.com for connecting systems and GoHighLevel as the CRM and messaging layer. That is the delivery stack at Di-Hy, my AI implementation firm. For a trades business, the usual targets are missed-call recovery, form capture, follow-up sequences and reporting.

I don't recommend either as a purchase for a business that has no written process yet. They run what you give them. See what to check before buying AI.

The board where work gets assigned: Monday.com

For one of my businesses, work is assigned and tracked on a Monday.com board that people other than me can see. The point isn't the software. It's that each item has one view and one owner, so nobody needs to open a folder or ask me where something stands.

The record: HubSpot and Google Analytics

Anything that needs to be measured gets a CRM record or an analytics tag (HubSpot and Google Analytics are the ones I use), and I check that the tag is actually there. A missing tag is the classic silent failure: most pages carry no tracking code and the numbers you read for weeks mean nothing.

That's the whole stack. A handful of jobs, a handful of tools.

The rules that keep it small

This is the part worth copying.

1. A tool has to do one job I can name

"AI for marketing" isn't a job. "Draft the first version of the weekly client report from the numbers" is. If I can't say what a tool replaces, it doesn't get a trial.

2. The process has to exist on paper first

If I can't explain the process manually, step by step, the tool will only automate the confusion. I write it down before I buy anything. This is also what I tell clients, and it's the single most common reason their projects stall.

3. Nothing that touches money or a customer runs unsupervised

Software finds, drafts and prepares. A person decides, signs and sends. I run this rule strictly: nothing runs on its own, and a named person signs. I trust the output of an AI tool like I'd trust a smart new hire's first month. Useful, fast, and checked.

4. I check its work, especially when it sounds sure

AI is confident whether it's right or not. An AI review of a deal file can spend much of its effort correcting conclusions that had already been replaced, and cite page numbers that don't exist. Give it a corrected second pass and it can still find the most important number in the file. Both things are true. It's a smart opinion, not a verdict.

The cure is boring. Give it the source documents. Keep old conclusions marked as superseded so they don't get cited as current. Ask where each claim comes from.

5. A tool has to connect to something I already use

The demo always works. What matters is where the data comes from, where it goes and who keeps it current. If the answer is "I'll just update it by hand," that part won't happen, and the tool becomes shelfware.

6. I decide what the saved time is for

Ten hours saved a week isn't profit unless those hours produce something. Before I adopt a tool, I decide what I'll do with the time, and I measure that. I make the full case in why hours saved is not profit.

What I'd cut first from most stacks

I can't tell you what I've cancelled, because I'm not going to invent a "tools I quit" story for effect. What I can tell you is what I'd cut first from most small business stacks:

  • Anything bought for a demo and never connected to real data.
  • Two tools doing the same job because each was bought in a different month.
  • Anything nobody owns. If no named person looks at it weekly, it's already broken and you haven't noticed.
  • Content generators used as a first AI project. For most service businesses the money is leaking in follow-up, not in how often you post. I made that point in my rule for what to automate first.

The honest limit

A stack doesn't make a business. I've seen small companies with three well-used tools beat ones with fifteen, and I've seen the reverse go badly. A clear process, an owner and a baseline do more than any subscription. I wrote more about that in AI Isn't Magic. It's Margin.

If you're staring at a list of tools and can't tell which ones your business needs, that's usually a process question, not a software question. Di-Hy's free audit is built to sort that out.

FAQ

What AI tools does a small business owner actually need?

Usually fewer than they think: one general assistant for writing and analysis, a CRM that records every lead, and one automation layer for follow-up. Start with the process, then pick the tool that fits it.

Is Claude or ChatGPT better for business?

It depends on the task, and the answer changes as the models do. Test both on your real work, such as a customer email, a report or a document review, and keep the one whose output you edit less.

How much should I spend on AI tools?

There's no honest universal number. Spend nothing until you can name the job, the owner and the measurement. Then compare the cost to what that job is worth.


Sources: Make.com. Tool features and plans change. Check each vendor's current documentation.