Data Analyst to AI Engineer roadmap
The longest of the common transitions into AI engineering, and an entirely achievable one. What you bring - SQL, data intuition, business framing, stakeholder fluency - is real and useful. What you're missing is the whole of software engineering, and that is a bigger gap than most course syllabi admit.
16 stages286 topics
Plan for 9–15 months of consistent part-time work, not a 12-week bootcamp. The single most common way this transition fails is underestimating section 3: becoming a programmer, not someone who can write scripts. Sections 3, 4 and 5 are roughly half the total effort and there is no shortcut through them. Everything after that is genuinely easier than it looks.
Audit: What You Already Have
Tick this honestly. It's a shorter list than the data scientist's, but the items on it are genuinely valuable and several teams are short of them.
The Shape of the Gap
Read this before planning your time. The most expensive mistake in this transition is treating it as a content problem when it is a skill problem.
Programming for Real
The biggest section, and the one that decides the outcome. Budget four to six months. Fluency means writing working code from a blank file without a tutorial - not recognising code when you see it.
Software Engineering
What separates a script from software. This is the vocabulary and practice you will be assessed on in every interview.
From Report to Service
The bridge exercise, and the most valuable month of this roadmap. Take a report you already produce manually and turn it into software that runs without you.
Statistics You Actually Need
Not a full statistics course. Enough to evaluate systems honestly and avoid drawing confident conclusions from noise.
Just Enough Machine Learning
Two to three weeks, then move on. You need intuition and vocabulary, not the ability to train state-of-the-art models.
LLM Fundamentals
Prompting as Engineering
Reliability
How you make an unpredictable component safe to put in front of users. New territory, and heavily assessed in interviews.
Embeddings and Vector Search
RAG
The most common thing an AI engineer builds, and the best portfolio project available to you.
Agents and Tools
Evaluation
Where your analyst instincts become an advantage. You already refuse to accept a number without asking how it was computed - that is precisely what this requires.
Backend, Data and Deployment
Your SQL background makes half of this quick. The service and deployment half is new.
Making the Move
The honest path
Three stages with a gate at the end of each. Do not move on until the gate is met - arriving at LLMs without programming ability is the standard failure of this transition.
Stage 1 - Become a programmer (4–6 months)
- Sections 3, 4, 5 and 6
- Gate: you can build, test, containerise and deploy a small API from an empty folder, without a tutorial open
Stage 2 - Learn the AI layer (3–5 months)
- Sections 7 through 13
- Gate: you have a deployed RAG system with an evaluation set you built, reporting retrieval and generation metrics separately
Stage 3 - Make it a job (2–4 months, overlapping)
- Sections 14 through 16, plus applications
- Gate: you can talk through your system's cost, latency, failure modes and evaluation without notes
The shortcut that is real
- Automate something at your current job - a recurring report, a manual triage, a data check
- Put an LLM feature into an internal tool nobody else has time to build
- Real production usage beats any personal project, and an internal move is the most likely way this transition actually completes
Analysts who make this move well tend to keep their old strength rather than abandon it. Knowing what a metric actually means, and refusing to accept a number without asking how it was computed, is exactly the instinct that AI evaluation requires - and it is in short supply on most teams.