The Problem With "Best AI Tool" Lists
Every week there's a new list of "the best AI tools for your business." The problem: they're all generic. They don't know your tech stack, your team's technical level, your compliance requirements, or the specific workflows you're trying to automate.
This article isn't a list. It's a framework — the same one we use with every client at Treesera to figure out exactly which tools belong in their stack, and why.
Start With the Problem, Not the Tool
Before you look at a single tool, answer these three questions:
1. What specific task are you trying to eliminate or speed up?
Be precise. "Improve our marketing" is not a task. "Automatically draft a personalised follow-up email within 10 minutes of a lead filling our contact form" is a task.
2. What does the ideal output look like?
If you automate the follow-up email, what should it say? Should it reference the service the lead mentioned? Should it come from the founder's email address or a generic info@ address? The more precisely you can describe the output, the easier it is to build toward it.
3. Where does the data live today?
Every automation needs inputs. Where is the lead data stored? Is it in a CRM, a Google Sheet, an email inbox? The answer determines which tools can actually talk to each other.
The Four Layers of an AI Automation Stack
A useful mental model: think of your stack as four layers.
Layer 1: Trigger & Data Capture
Something has to start the workflow. Common triggers:
- →A form submission (Typeform, Tally, your own website form)
- →An email arriving in a specific inbox
- →A row added to a Google Sheet
- →A message in Slack
- →A webhook from another system
Layer 2: Orchestration
This is the workflow engine — the tool that connects your trigger to your actions. The three most common options:
n8n — Open source, self-hostable or cloud, the most powerful for complex workflows. Steeper learning curve but no usage-based pricing once self-hosted. Best for: technical teams, complex multi-step workflows, businesses with data privacy requirements.
Make (formerly Integromat) — Visual builder, easier to use than n8n, cloud-only, usage-based pricing. Best for: mid-complexity workflows, teams who want a visual builder without writing code.
Zapier — Easiest to use, most integrations, but expensive at scale and limited for complex logic. Best for: simple one-trigger-one-action automations, non-technical users.
My honest recommendation: if you're building more than three automations, start with Make. If you're serious about automation as a business function, learn n8n.
Layer 3: Intelligence
This is where LLMs come in. Not every automation needs AI — some workflows are just if-this-then-that logic and don't need language understanding. But when you need to:
- →Draft personalised text
- →Classify or score inputs
- →Extract information from unstructured documents
- →Make judgment calls based on context
…you need an LLM in the workflow.
Claude (Anthropic) — Best for nuanced writing, long document analysis, and workflows where tone matters. Excellent context window. Our default recommendation.
GPT-4o (OpenAI) — Strong all-rounder, best multimodal capabilities (image + text), huge ecosystem of integrations.
Open-source (Llama 3, Mistral) — For workflows where data can't leave your infrastructure, or where cost at scale is a constraint.
How to choose: If your workflow involves customer-facing communication, use Claude. If it involves document parsing or multimodal inputs, consider GPT-4o. If it involves sensitive data, evaluate open-source.
Layer 4: Output & Storage
Where does the result go? Common destinations:
- →Your CRM (HubSpot, Salesforce, Pipedrive)
- →A spreadsheet or database (Airtable, Google Sheets, Notion)
- →An email or Slack message
- →Your own application database (Postgres, Supabase)
- →A document (Google Docs, Notion page)
This layer is where most automations break down — not because the AI got it wrong, but because the output format doesn't match what the destination system expects. Plan this layer carefully.
The Build-vs-Buy Decision
Some tools do all four layers in one (like a full AI agent platform). Others are best-in-class for one layer. The trade-off:
All-in-one tools (Relevance AI, Lindy, etc.) — Faster to start, less flexible, harder to customise as needs grow.
Best-in-class per layer — More setup, more maintenance, but far more powerful and adaptable long-term.
For most growing businesses, the best-in-class approach with n8n or Make as the orchestration layer is the right call. All-in-one tools tend to hit their ceiling fast.
Red Flags to Avoid
Building automations without error handling. Every workflow will eventually receive unexpected input. If you haven't handled failures, one bad data point will break everything silently.
Using Zapier for complex logic. Zapier is excellent for simple automations but awkward for anything with branching conditions, loops, or complex data transformation.
LLM outputs going directly to customers. Always have a human review step for any LLM output that's customer-facing, at least until you've validated the output quality across 100+ real examples.
Not documenting your workflows. Six months after building an automation, you won't remember how it works. Document as you build.
How to Audit Your Current Stack
If you already have some automations running, here's how to evaluate whether they're the right tools:
1. Map every manual step in your top five business processes
2. Identify which tools you currently use to handle each
3. For each tool, ask: is this the cheapest option that does this job at our volume?
4. Look for places where two different tools are doing things that one tool could handle
Not sure where to start? Try our free AI Audit tool — it maps your highest-impact automation opportunities in two minutes and gives you a prioritised starting point.