There are, at a rough count, several thousand AI tools with a landing page and a pricing table. You need about eight of them.
This is the stack I actually build on, organised by the job rather than the category, with honest notes on when each one is the wrong choice.
How to choose (the 3 questions)
Before adding any tool, three questions. They eliminate most of the shortlist.
- What job is this doing that nothing I already have can do? Most new tools overlap something you own.
- Can I get my data out? If the answer is no, you are not adopting a tool, you are adopting a landlord.
- Does it have an API? A tool that cannot talk to your other tools becomes an island, and islands create manual work — the exact thing you were trying to remove.
Orchestration — the backbone
This is the one that matters most, because it is what connects everything else. Pick carefully; migrating later means rebuilding by hand.
n8n — self-hostable, which means cost stops scaling with volume, and it has the strongest AI and agent tooling of the mainstream options. You can drop into code in any node, so you never hit a hard ceiling. The trade is that someone has to own the server.
Make.com — the best balance for most growing businesses. Genuinely good visual debugging, proper branching and loops, nothing to host. Costs more than self-hosted n8n at volume.
Zapier — the widest app catalogue and the shallowest learning curve. Right answer for simple workflows maintained by non-technical people. Gets expensive fast as steps and volume grow.
Pick one and go deep. Shallow familiarity with three orchestration tools is worth less than real depth in one. The concepts transfer in a day when you eventually need another.
The models
You need one primary model provider. Adding a second has real overhead — two sets of keys, two billing accounts, two sets of quirks — so only do it for a specific reason.
What actually differentiates them for business use, in rough order of importance:
- Instruction following. Does it respect your constraints, or drift after a few turns? This matters more than raw capability for production work.
- Structured output. Can you reliably get valid JSON back? This is the difference between an automation that works and one that needs a human checking it.
- Tool calling. Essential if you are building agents. Varies more between providers than the marketing suggests.
- Cost at your volume. Model your real usage, not a single test call. Agents multiply this considerably.
Test with your own prompts and your own data before committing. Public benchmarks tell you very little about how a model handles your messy inputs.
Data & storage
Every automation worth building needs to remember something — what it has already processed, what the current state is, what happened last time.
Airtable — spreadsheet familiarity with a real API. Excellent when non-technical people need to see and edit the data. Gets expensive and slow at scale.
Supabase — a proper Postgres database with an approachable interface, plus built-in vector storage if you are doing retrieval. My default for anything that will grow.
Google Sheets — genuinely fine as a starting point, and people are snobbish about this unfairly. It breaks down around concurrent writes and larger volumes, but for a first automation it removes a whole learning curve.
Building interfaces
Sometimes the automation needs a face — a form, a dashboard, an internal tool.
Streamlit — fastest route from a Python script to something a colleague can use. Ideal for internal tools and data apps.
Lovable / Replit — for spinning up a working web interface quickly. Good for prototypes and internal tools; review what they generate before anything touches production data.
Plain forms — often the correct answer. A Google Form into an automation solves more problems than people expect, and takes an afternoon.
Specialist tools
Add these only when the need is real.
- Vector database — only if you are building retrieval over your own documents. Supabase covers this for most cases without adding a separate service.
- Transcription — for anything involving calls or meetings. Cheap and reliable now.
- Document parsing — for invoices, contracts and forms arriving as PDFs. This is where a lot of manual work still hides.
- Docker — not a tool you use daily, but the thing that makes self-hosting manageable rather than painful.
What to skip
Being direct about this, because the noise is expensive.
- Anything that is a thin wrapper on a model you can call directly. A large share of AI products are a prompt and a nicer interface. If the only value is the prompt, you can write the prompt.
- All-in-one platforms that do everything adequately. They lock you in and are usually worse at each individual job than a focused tool.
- Tools with no export. Non-negotiable. Your data must be able to leave.
- The tool everyone posted about last week. Wait a month. Most of them do not survive contact with real work.
The stack, summarised
| Job | Start with | Add later if |
|---|---|---|
| Connecting everything | n8n or Make.com | — |
| The intelligence | One model provider | A second for a specific gap |
| Remembering things | Airtable or Supabase | Dedicated Postgres at scale |
| Collecting input | A plain form | Custom interface when forms limit you |
| Documents | — | Parsing tool when PDFs pile up |
| Knowledge retrieval | — | Vector store when docs get large |
| Hosting | Managed cloud | Docker + VPS when cost matters |
Three tools will carry you a long way: an orchestrator, a model, a database. Everything else earns its place by solving a problem you have actually hit — not one you read about.