| Typical title | Pay benchmark | Entry route |
|---|---|---|
| AI engineer, applied AI engineer, forward deployed AI engineer | $135,980 median annual wage for software developers, May 2025 (BLS) | Software engineering role first, then LLM application work |
Checked 17 October 2026 at each employer's own page and the BLS Occupational Outlook Handbook.
What an AI engineer does, and how it differs from an ML engineer
An AI engineer builds products on top of existing models. The day to day is writing application code that calls a model API, retrieving the right documents for each request, designing prompts and tool calls, running evaluation suites, and deploying the result so it stays fast, cheap and reliable. Anthropic's posting lists "developing evaluation frameworks," "customized pilots, prototypes, and evaluation suites," and hands-on code reviews with customer engineering teams. Databricks' AI Engineer, Forward Deployed Engineering posting names RAG, multi-agent systems, Text2SQL and fine-tuning, using tools such as Hugging Face, LangChain and DSPy.
A machine learning engineer usually sits closer to the model itself: training, fine-tuning, data pipelines and serving infrastructure. Scale AI's Senior Machine Learning Engineer posting describes retraining production models, hyperparameter tuning and "large-scale fine-tuning pipelines," with PyTorch or TensorFlow required. The two roles overlap, and many teams use the titles loosely. If you want the training-heavy version, read our machine learning engineer career path.
| AI engineer | Machine learning engineer | |
|---|---|---|
| Main output | Shipped features built on foundation models | Trained or fine-tuned models and the pipelines around them |
| Core tools | Python or TypeScript, model APIs, retrieval, agent frameworks, eval harnesses | Python, PyTorch or TensorFlow, data pipelines, GPU training |
| Typical entry | Software engineering experience | Software or data science experience, often a graduate degree |
Checked 17 October 2026 at each employer's own page.
How to become an AI engineer: step-by-step path
These estimates assume about 10 to 15 hours a week of study alongside a job or degree. Someone already working as a software engineer can skip steps 1 and 2 and finish in roughly 6 to 9 months.
- Programming foundations (2 to 4 months). Get fluent in Python, Git, the command line, HTTP APIs and SQL.
- First software job or equivalent experience (6 to 12 months, often overlapping). Most AI engineer postings want production experience. Anthropic asks for 4 or more years in roles such as software engineer or product engineer. A junior developer job, an internship or a sustained open-source contribution record counts here. Our page on AI internships with deadlines and pay lists current student roles.
- Machine learning basics (4 to 6 weeks). Learn what training, loss, overfitting, embeddings and evaluation metrics mean. Google's Machine Learning Crash Course covers these and now includes an introduction to large language models.
- Deep learning in practice (6 to 8 weeks). Train and deploy a small model yourself so the vocabulary becomes concrete. fast.ai's free course has 9 lessons of about 90 minutes each.
- Building with LLMs (6 to 10 weeks). Learn tokenization, transformers, retrieval-augmented generation, tool use, structured output and agents. Hugging Face's LLM course and the model providers' own courses and docs cover this ground.
- Evaluation and production (4 to 6 weeks). Build eval sets, grade outputs automatically, track cost and latency, and add guardrails. Every posting we checked mentions evaluation.
- Portfolio and applications (ongoing, 2 to 3 months of focused search). Ship two or three projects publicly, write up what you measured, and apply to applied AI, solutions and forward deployed roles.
Skills matrix
| Skill | Why it matters | How to show it |
|---|---|---|
| Python or TypeScript | Anthropic requires proficiency in one of them for production applications | Public repo with tests and a working deploy |
| Model APIs and prompt design | Core of daily work; Anthropic lists "advanced prompt engineering" | A project that compares prompt versions with measured results |
| Retrieval (RAG) | Databricks names RAG first in its requirements | A search-backed assistant over a real document set, with citations |
| Agents and tool use | Databricks lists multi-agent systems; Anthropic lists agent frameworks and MCP | An agent that calls at least two tools and logs each step |
| Evaluation | Figma, Anthropic and Databricks all ask for eval work | An eval suite with a pass rate you track across versions |
| Fine-tuning | Figma asks for fine-tuning and reinforcement learning methods such as DPO and PPO | A small fine-tune with a before and after comparison |
| Cloud deployment | Databricks asks for production deployments on AWS, Azure or GCP | A deployed service with a public URL and a cost note |
| Communication | Anthropic and Databricks roles are customer-facing | A written project report or a recorded demo |
Checked 17 October 2026 at each employer's own page.
Free official courses and docs
Use one course per step. These are the providers' own learning pages.
| Resource | Provider | Best for | Format |
|---|---|---|---|
| Machine Learning Crash Course | ML basics, embeddings, intro to LLMs, production ML systems | Self-paced modules with videos and exercises | |
| Practical Deep Learning for Coders | fast.ai | Training and deploying your first models | Free, 9 lessons of about 90 minutes |
| LLM Course | Hugging Face | Transformers, fine-tuning, datasets, reasoning models | Free; about 6 to 8 hours a week per chapter |
| Generative AI for Beginners | Microsoft | Building generative AI apps end to end | 21 lessons on GitHub |
| Short courses | DeepLearning.AI | Focused topics such as prompt engineering and evaluation | Catalog of short courses |
| Anthropic courses | Anthropic | Building with the Claude API, Model Context Protocol, agent skills | Self-paced courses on Anthropic Academy |
| OpenAI Cookbook | OpenAI | Worked examples for agents, evals, retrieval and fine-tuning | Code examples and notebooks |
Checked 17 October 2026 at each provider's own page.
Hugging Face says its course "requires a good knowledge of Python" and is better taken after fast.ai or a DeepLearning.AI program, which is why it sits at step 5 here. Certifications matter less than shipped work: Microsoft's own page shows its Azure AI Engineer Associate certification as retired.
What employers list in real postings
These four postings were open on 17 October 2026 and publish pay on the employer's own page.
| Employer | Role | Published pay | Experience asked | Degree wording | Location |
|---|---|---|---|---|---|
| Anthropic | Applied AI Engineer, Enterprise Tech | $200,000 to $320,000 annual salary | 4+ years in technical roles; production LLM experience | Bachelor's or equivalent combination of education, training and experience | San Francisco or New York, 3 days a week in office |
| Databricks | AI Engineer, Forward Deployed Engineering | $182,000 to $250,208 local pay range, annual | Production GenAI and LLM applications; not for new graduates | Graduate degree in a quantitative field or equivalent practical experience | United States, open to remote |
| Figma | AI Applied Scientist | $153,000 to $376,000 annual base salary | 4+ years in generative AI, 6+ in ML, NLP or computer vision | No degree requirement listed | San Francisco, New York or remote in the US |
| Scale AI | Senior Machine Learning Engineer, Public Sector | $225,750 to $282,450 annual base | Production GenAI, agentic AI or computer vision; active security clearance | Graduate degree listed as a nice to have | Washington, DC |
Checked 17 October 2026 at each employer's own page.
Every posting asks for production deployment, every one mentions evaluation or model quality, and none of the four makes a specific degree a hard requirement. The Figma AI Applied Scientist posting also lists "Build evaluation systems to measure and improve quality of AI features" among its duties, which is the single most useful skill to put in your portfolio. Databricks states plainly that its role "is not intended for internship, new graduate, or entry-level applicants," so treat these as targets for year two or three, not your first application. For more pay detail by title and city, see our AI engineer salary breakdown, and for live openings use our list of machine learning and AI engineer jobs.
Pay and job outlook from BLS
BLS has no occupation called AI engineer. The closest match for engineers who ship AI features is software developers. The BLS software developers page reports a median annual wage of $135,980 in May 2025, with the lowest 10 percent under $82,460 and the highest 10 percent above $214,670. It projects 10 percent employment growth from 2025 to 2035 for software developers, quality assurance analysts and testers combined, with about 106,100 openings a year, and lists a bachelor's degree as the typical entry-level education.
For research-heavy roles, the BLS computer and information research scientists page reports a median annual wage of $140,300 in May 2025 and projects 22 percent growth from 2025 to 2035, with about 2,900 openings a year. That occupation typically needs a master's degree. Both pages were updated 27 August 2026.
Portfolio project ideas
- Document assistant with citations. Retrieval over a public document set (court opinions, city council minutes, product manuals), answers with linked sources, and an eval set of 50 questions with a tracked accuracy score.
- Text-to-SQL agent. Natural-language questions over a public dataset, with a table of queries the agent got wrong and how you fixed them. Databricks lists Text2SQL by name.
- Prompt and model comparison harness. Run the same 100 test cases across two models or two prompt versions and report pass rate, cost per request and latency.
- Tool-using agent with logs. An agent that books, files or summarizes using two or more external tools, with every step logged and a failure analysis.
- Small fine-tune. Fine-tune an open model on a narrow task using the Hugging Face course methods and publish a before and after comparison.
- Evaluation write-up. A short public report on how you graded outputs, where the grader disagreed with humans, and what you changed. Our AI evaluation career path covers this skill in depth.
Routes without a degree
A degree helps, but the postings above show it is not a hard gate for applied roles. Anthropic accepts "an equivalent combination of education, training, and/or experience," Databricks accepts "equivalent practical experience," and Figma lists no degree requirement. Routes that work:
- Software job first. Get hired as a developer through a portfolio, then move onto the team that ships AI features. This is the most common path for applied roles.
- Contract AI work as a bridge. Model evaluation and coding-task work on training platforms builds the judgment employers ask for; our guide to entry level AI jobs without experience lists where to start.
- Open-source contributions. Merged pull requests to libraries named in postings, such as Hugging Face, LangChain or DSPy, are visible proof of skill.
Research scientist roles are the exception: BLS lists a master's degree as typical, and many lab research postings ask for a PhD. For every role type side by side, see our overview of AI career paths, or start from the AI jobs hub if you are still choosing between salaried and contract work.
How we researched this
On 17 October 2026 we opened each employer posting on its own careers page or Greenhouse board and recorded the published pay range, its unit, the location, the experience asked for and the degree wording. We took pay and outlook figures from the BLS Occupational Outlook Handbook pages updated 27 August 2026, which report May 2025 wages and 2025 to 2035 projections. We opened every course page linked above the same day. Time estimates are editorial judgment based on the course lengths the providers publish and the experience levels the postings ask for. Our general method is on how we rank.
Frequently asked questions
How long does it take to become an AI engineer? About 12 to 24 months from scratch, or 6 to 9 months if you already work as a software engineer. Most postings checked on 17 October 2026 ask for several years of production engineering experience.
Do I need a degree to become an AI engineer? No, not for most applied roles. Anthropic accepts an equivalent combination of education, training and experience, and Databricks accepts equivalent practical experience.
How do I get into AI with no experience? Start with Python and a software role, then add LLM projects. Contract model evaluation work and open-source contributions are practical bridges while you build a portfolio.
What is the difference between an AI engineer and a machine learning engineer? An AI engineer builds products on existing models; a machine learning engineer trains and serves models. The titles often overlap.
What does an AI engineer earn? Posted base ranges run from $182,000 to $320,000 a year at Databricks and Anthropic as of 17 October 2026. BLS reports a $135,980 median for software developers in May 2025.
Which programming language should I learn first? Python. Anthropic's posting accepts Python or TypeScript, and the Google, fast.ai and Hugging Face courses all teach in Python.
Are AI certifications worth it? They matter less than shipped projects. Microsoft lists its Azure AI Engineer Associate certification as retired, while every posting checked asks for production work.
Which free course should I start with? Google's Machine Learning Crash Course for concepts, then fast.ai for hands-on deep learning. Hugging Face recommends taking an introductory course like fast.ai before its LLM course.
What should an AI engineer portfolio include? Two or three deployed projects with measured results: a retrieval assistant with citations, an agent with logged tool calls, and an evaluation suite with tracked pass rates.