I started as a Mathematics teacher. Not a developer, not a computer science graduate — a teacher. Four years later I build AI automation systems for businesses and train other people to do it. So when I say the path is open, I mean it specifically.
This is the roadmap I would give myself if I were starting today, with the detours removed.
What the job actually is
An AI automation engineer takes a manual business process and turns it into a system that runs itself.
That is the whole job. Someone is spending four hours a week copying data between two systems, or answering the same customer question sixty times, or building the same report every Monday. You look at that, design something that does it without them, build it, test it, and hand it over.
What surprises people is the ratio. On a typical project, maybe 40% of the effort is building. The other 60% is understanding the process, deciding what should be automated, and making it reliable enough to trust unattended.
The uncomfortable truth: the technical part is the easy part. Anyone can learn the tools in eight weeks. What takes longer — and what actually gets you paid — is business judgement.
Do you need a degree?
No. I do not have one in this field, and neither do most of the people I know doing it well.
What clients ask for, in order: can you show me something you built, can you explain it in language I understand, and will you still be there in three months if it breaks. Nobody has ever asked me about a certificate.
What does help enormously is any background where you had to break a complex thing into steps and explain it to someone who did not get it. Teaching, accounting, operations, admin, support — all of these transfer better than people expect.
The skill stack that matters
Tier 1 — non-negotiable
- One automation platform, deeply. Pick n8n or Make.com and go far past the tutorials. Depth in one beats shallow familiarity with five.
- APIs and webhooks. What a REST call is, what headers and auth do, how to read documentation, how to debug a 401. This unlocks everything.
- JSON and data shapes. You will spend more time reshaping data between systems than anything else.
- Prompting for production. Not clever prompts — reliable ones. Structured output, guard rails, handling the case where the model returns something unexpected.
Tier 2 — the multiplier
- JavaScript or Python basics. Enough to write a transform step when the visual tool cannot express what you need. This is the single biggest jump in what you can charge.
- Databases. Supabase or Airtable. Every serious automation needs to remember something.
- Error handling and retries. The difference between a demo and a system someone pays for.
Tier 3 — the senior layer
- Agentic patterns. Tool use, memory, when an agent is and is not appropriate.
- RAG and vector databases. For document and knowledge assistants.
- Deployment. Self-hosting, Docker, environment variables, keeping secrets out of the workflow.
The 6-month roadmap
| Month | Focus | What you should have at the end |
|---|---|---|
| 1 | Tool fundamentals | 10+ working workflows. Triggers, actions, filters, error branches. |
| 2 | APIs & real data | Connected 3 services with no pre-built integration, using raw HTTP. |
| 3 | AI in the loop | Two workflows where an AI step produces reliable structured output. |
| 4 | Code & databases | A system with persistent state and a custom transform step. |
| 5 | Portfolio build | 3 complete case studies, each with a before/after and a number. |
| 6 | First clients | One paid project delivered. One testimonial in hand. |
Two things about this table. First, it assumes roughly 8–10 focused hours a week — not full time. Second, the months are sequential for a reason: skipping to month five with a shaky month two is the most common way people stall.
Building a portfolio that converts
Most beginner portfolios are a list of workflows. That does not sell, because a client cannot tell whether a workflow is good.
What sells is a before and after with a number attached.
Weak: "Built an invoice automation using n8n and Google Sheets."
Strong: "A consultancy was spending 5 hours a week generating and chasing invoices by hand. I built a system that generates the invoice on project completion, emails it, tracks payment status, and sends reminders on day 7 and 14. Manual time now: zero. Payment delay dropped from 21 days to 9."
You do not need paying clients to write these. Automate something for a friend's business, a family shop, a local church, a small NGO. Do it free, measure the before and after honestly, and write it up. Three of those and you have a portfolio that beats most people charging money.
Getting your first paid client
Ranked by how well they actually work:
- People who already know you. Almost every first client comes from here. Not a pitch — a specific observation: "I noticed you do X manually every week. I can make that automatic. Let me do it free and if it works, we talk."
- Publish what you build. Post the case studies. Explain the problem and the fix in plain language. This compounds quietly and then suddenly.
- Free audits. Offer to spend 30 minutes mapping someone's process and telling them what could be automated. Roughly a third turn into work, because you have already demonstrated the thinking.
- Freelance marketplaces. Slow and price-competitive, but real. Useful for the first two reviews, then leave.
What you can realistically charge
Ranges vary a lot by market and client size, but as a shape:
- Starting out — small, single-purpose automations. Enough to be worth your time, not enough to live on. The goal here is testimonials, not revenue.
- Once you have 3 case studies — multi-step systems with AI in the loop. This is where it becomes real income.
- Once you can code and deploy — full systems, integrations with a client's existing stack, agentic components. Several multiples of the previous tier.
- Retainers — the actual business model. Monitoring, tweaks, new workflows as the client grows. Predictable and far less exhausting than constant new sales.
Price on outcome, not hours. If a system saves someone 20 hours a month forever, the value is not related to how long it took you to build.
Mistakes that slow people down
- Tool hopping. Two weeks on n8n, then Make, then Zapier, then back. You end up shallow in all three. Pick one, go deep, learn the others later in days.
- Learning without building. Courses feel like progress. Only shipped systems actually are.
- Waiting to feel ready. Nobody feels ready. Take the first project slightly before you are comfortable — that is where the real learning is.
- Ignoring failure paths. Beginners build for the happy path. Professionals ask what happens when the API is down, when the field is empty, when the same request arrives twice.
- Selling tools instead of outcomes. Clients do not care that it runs on n8n. They care that nobody has to do it any more.
The field is genuinely open right now. The demand is real, the barrier to entry is low, and the people already in it are mostly self-taught. What separates the ones who make it is not talent — it is finishing things and writing them up.