Data & AI · Updated 16 September 2026
AI Engineer (LLM / Agent Applications) salary and career roadmap
Build products on top of foundation models - retrieval, agents, evaluation and cost control - without training a model from scratch.
Salary data checked · outlook from BLS projections released · by Bilal Tahir
- What it is
- A AI Engineer is a software engineer who builds products on top of foundation models somebody else trained - retrieval, tool use and agents, evaluation harnesses, guardrails, and token-cost and latency control - rather than training models from scratch.
- Salary
- A AI Engineer in the United States earns a median of $118,000 entering the field, $155,000 at mid-career and $215,000 at senior level; the top end is $300,000. (Levels.fyi United States AI Engineer reports, reconciled with Glassdoor AI Engineer averages, )
- Outlook
- Employment is projected to grow 10% over the ten years to 2036, with about 106,100 US openings a year. (BLS Occupational Outlook Handbook, 15-1252 Software Developers (nearest proxy; there is no BLS occupation code for AI Engineer), )
- Time to first job
- A career changer starting from zero typically needs 9 to 30 months at 10-15 hours a week to reach a first offer.
- Cost to get in
- The cheapest verified route in (Software engineer adds AI) costs about $400; the most expensive costs about $1,500.
- Roadmap
- Our AI Engineer roadmap is 9 steps and about 1,020 study hours; the fastest route in is Software engineer adds AI at about 9 months.
- Degree
- No degree is legally required and there is no established credential for the role, so a deployed system with an evaluation suite substitutes for one, but most large-employer postings still list a bachelor's degree in computer science or a related field as a preference and the interview loop is a software-engineering loop with LLM topics layered on.
- Automation exposure
- Low. Coding assistants compress the parts of the work that were already mechanical - boilerplate integration code, first drafts of tests, glue between APIs - and model platforms keep absorbing whole task categories, with tool calling, structured output and basic retrieval-augmented generation each moving from application code into vendor features inside two years. What does not automate is deciding what a correct answer looks like in a specific business, building the evaluation set that proves it, and owning the incident when a vendor changes a model's behaviour overnight.
What does a AI Engineer do?
A AI Engineer is a software engineer who builds products on top of foundation models somebody else trained - retrieval, tool use and agents, evaluation harnesses, guardrails, and token-cost and latency control - rather than training models from scratch. The AI engineer is the newest job title in this cluster: it barely existed before 2023 and was a standard req by 2025. The same work is also posted as AI Application Engineer, LLM Engineer, GenAI Engineer, Applied AI Engineer, Agent Engineer.
The AI engineer is the newest job title in this cluster: it barely existed before 2023 and was a standard req by 2025. The premise is that foundation models are now bought rather than built, so the scarce skill is not training a model but making one behave reliably inside a product. The work is retrieval-augmented generation, tool use and agent orchestration, prompt and context engineering, structured output, evaluation harnesses for non-deterministic systems, guardrails, caching, latency and token-cost management, and the observability to know when a change made things worse. You are usually shipping a feature - a support assistant, a document Q&A system, an internal agent that files tickets, a coding or research copilot - not a model.
Who thrives: strong application developers who are comfortable with ambiguity and obsessive about evaluation. The dividing line between good and bad AI engineers is almost entirely about measurement. Anyone can wire an API call and a vector database into a demo in a weekend; the job is turning that demo into something with a test suite, a regression gate, a cost ceiling and a defensible answer to "is the new prompt better?" If you like building things people use and can tolerate a stack that changes every quarter, this fits.
The honest tradeoffs. The pay distribution is bimodal and confusing: Levels.fyi's US AI engineer median total compensation is $154,000 - lower than machine learning engineer because the title is used broadly, including at companies paying ordinary application-developer rates - while Google's AI engineer median is $255,000 and software engineers at frontier labs report medians of $820,000-$880,000. The role is also the least standardised: two jobs with the same title can be a prompt-tuning position and a distributed-systems position. And it is the fastest-decaying skill set on this site; frameworks you learn this year may be irrelevant in two. The durable parts are evaluation, retrieval, systems design and cost reasoning - build on those, not on a particular library.
Why AI Engineer pay is high
Companies are under board-level pressure to ship AI features and have discovered that the bottleneck is not model access - everyone can call the same APIs - but the engineering discipline to make outputs reliable enough to put in front of customers. That capability is new enough that there is no established pipeline producing it, so employers bid for people who have actually shipped something that works, and demand extends well beyond technology companies into legal, healthcare, insurance, financial services and enterprise software. The upper tail is driven by a separate force: a small number of frontier labs and AI-native startups competing for the same engineers with equity-heavy packages - Levels.fyi shows software engineer medians of $880,000 at OpenAI and $820,000 at Anthropic, with the bulk of that in stock. That tail is real but it is a few thousand jobs; the broad market for AI engineers pays like senior application engineering plus a premium, which is what the $154,000 Levels.fyi median and $145,471 Glassdoor average reflect.
What's good
- The most accessible high-paying AI role: no PhD, no distributed training experience, and a strong application developer can transition in about nine months.
- Demand spans far beyond tech - legal, healthcare, insurance, financial services and enterprise software are all hiring.
- Short build cycles and a visible product, so you see users touch what you made within weeks.
- No entrenched credential means a working portfolio genuinely substitutes for a degree.
- A real upper tail: Levels.fyi software engineer medians are $880,000 at OpenAI and $820,000 at Anthropic.
- Skills compound toward ML engineering, product engineering and technical leadership.
What's hard
- Fastest-decaying skill set on this site; frameworks and even entire problem categories are absorbed into platforms within a year or two.
- The broad-market median is lower than the headlines suggest - $154,000 on Levels.fyi, $145,471 average on Glassdoor - because the title is applied loosely.
- Wildly inconsistent job definitions: two reqs with the same title can be prompt work or distributed systems work.
- Working with non-deterministic systems is genuinely frustrating; the same input can produce a different answer tomorrow.
- Heavy expectation-management burden - executives see a demo and assume the hard part is done.
- Vendor dependence: a model provider's pricing, deprecation or behaviour change is your incident.
What a AI Engineer does all day
- 09:00 - Check overnight traces: the support assistant escalated 14% of conversations, up from 9%. Something changed.
- 09:30 - Find the cause - a vendor model update changed refusal behaviour on a category of question. Pin the model version, open an incident note.
- 10:30 - Run the eval suite against the new model version; three of the 180 golden cases now fail. Adjust the system prompt and rerun.
- 11:30 - Design discussion with product: they want the agent to issue refunds. Argue for a human approval step above a dollar threshold.
- 13:00 - Deep work: rewrite the chunking strategy for the policy corpus; retrieval recall goes from 0.71 to 0.86 on the eval set.
- 15:00 - Cost review: 60% of spend is one endpoint that could be served by a smaller model. Implement routing with a quality gate.
- 16:00 - Add a prompt-injection test case after a user pasted instructions into an uploaded document.
- 16:45 - Write the weekly update: eval scores, escalation rate, cost per conversation, what shipped.
- Ongoing - Read release notes and one or two papers a week; resist rewriting the stack every time something new ships.
AI Engineer salary in 2026: by level
US, annual, USD. Base plus typical bonus where the source reports it.
A US AI Engineer earns a median of $118,000 entering the field, $155,000 at two to four years and $215,000 at senior level, with the top end at $300,000. These are base-salary figures in US dollars as of September 2026, synthesised from Levels.fyi United States AI Engineer reports, reconciled with Glassdoor AI Engineer averages. Pay varies about 20-40% by metro.
Read this ladder carefully, because the title spans a wider range than any other role on this site. Levels.fyi US 'AI Engineer' total compensation (retrieved September 2026): 10th percentile $83,200, 25th $110,000, median $154,000, 75th $215,000, 90th $300,000, with base at $145,000 median and median equity of $0 - most companies using this title pay a base-heavy package. Glassdoor's US average for 'AI Engineer' is $145,471; its 'AI and Machine Learning Engineer' average is $173,482 and 'ML/AI Engineer' is $206,724, which shows how much the label itself is worth. Big tech pays far above the broad market: Google AI Engineer on Levels.fyi has a $255,000 median (10th $112,500, 75th $389,000, 90th $535,000) and Microsoft's AI Engineer band runs roughly $184,000-$502,000 with a $282,000 median. The top rung is a genuinely different market: Levels.fyi software engineer medians are $880,000 at OpenAI (10th $450,000, 90th $1,092,000) and $820,000 at Anthropic (10th $375,000, 90th $1,325,000), with equity making up most of the package, and Scale AI spans roughly $231,000-$1,230,000. Those are small samples of a small number of employers and should be treated as the ceiling of the profession, not as a level you plan for. Equity at private AI labs is illiquid and its paper value depends on the next round.
There is no BLS occupation for AI engineer - the title is too new for the classification system, which is itself a useful signal about how young the role is. The nearest official proxy is Software Developers (15-1252): 1.9 million employed in 2025, median $135,980 in May 2025, +10% projected 2025-2035, about 106,100 annual openings, with BLS explicitly naming artificial intelligence as a demand driver. Computer and Information Research Scientists (15-1221) is +22% and Data Scientists (15-2051) is +35%, both also citing AI. The candid assessment: demand for this specific skill set has grown far faster than any of those series, but it is also the least stable role here - it is young, definitionally unsettled, and dependent on a model-provider ecosystem that could consolidate. Automation risk to the job itself is low in the near term (you are the person automating things, and coding assistants make you faster rather than redundant), but the specific tasks are volatile: tool-calling, structured output and basic RAG have each moved from application code into model and platform features within two years. Skill half-life here is short, and the roles that survive are the ones anchored in evaluation, systems and domain knowledge rather than in a framework.
“Demand for software developers, software quality assurance analysts, and testers is projected to be strong due to the continued expansion of software development for artificial intelligence (AI), Internet of Things (IoT), robotics, and other automation applications.”
AI Engineer salary by city
National bands scaled by metro wage differentials from the BLS May 2025 OEWS release.
AI Engineer pay is highest in San Jose / Silicon Valley (mid-career median about $220,100, ×1.42 the national figure) and lowest among large metros in Salt Lake City (about $147,250). The multiplier moves the offer, not what you keep after rent and state tax.
| Metro | Entry | Mid | Senior | vs national |
|---|---|---|---|---|
| San Jose / Silicon Valley | $167,560 | $220,100 | $305,300 | ×1.42 |
| San Francisco Bay Area | $159,300 | $209,250 | $290,250 | ×1.35 |
| New York City | $151,040 | $198,400 | $275,200 | ×1.28 |
| Seattle | $141,600 | $186,000 | $258,000 | ×1.2 |
| Boston | $135,700 | $178,250 | $247,250 | ×1.15 |
| Washington DC metro | $132,160 | $173,600 | $240,800 | ×1.12 |
| Los Angeles | $127,440 | $167,400 | $232,200 | ×1.08 |
| Chicago | $123,900 | $162,750 | $225,750 | ×1.05 |
| Austin | $123,900 | $162,750 | $225,750 | ×1.05 |
| San Diego | $122,720 | $161,200 | $223,600 | ×1.04 |
| Denver | $120,360 | $158,100 | $219,300 | ×1.02 |
| Philadelphia | $120,360 | $158,100 | $219,300 | ×1.02 |
| Dallas | $118,000 | $155,000 | $215,000 | ×1.0 |
| Minneapolis | $118,000 | $155,000 | $215,000 | ×1.0 |
| Raleigh-Durham | $118,000 | $155,000 | $215,000 | ×1.0 |
| Houston | $116,820 | $153,450 | $212,850 | ×0.99 |
| Atlanta | $115,640 | $151,900 | $210,700 | ×0.98 |
| Phoenix | $112,100 | $147,250 | $204,250 | ×0.95 |
| Miami | $112,100 | $147,250 | $204,250 | ×0.95 |
| Salt Lake City | $112,100 | $147,250 | $204,250 | ×0.95 |
A multiplier raises the number on the offer letter, not what you keep. San Francisco pays about 35% more than the national median for these roles, but median Bay Area rent and California state income tax eat most of that for anyone below the senior rung. Texas, Florida, Washington, Tennessee and Nevada levy no state income tax, which is worth roughly 4-10% of take-home versus California or New York City, where city tax stacks on top of state tax. Run the comparison on after-tax income minus housing before you move. Fully remote roles are the edge case worth chasing: a national pay band spent in a 0.88 cost market beats a 1.35 salary spent in a 1.6 cost market for most people.
How the multipliers are derived
Anchor: the BLS May 2025 OEWS national mean wage across all occupations is $33.54/hr ($69,770/yr). Metro all-occupation means from the same release: San Jose-Sunnyvale-Santa Clara $57.32 (1.71x national), San Francisco-Oakland-Fremont $48.19 (1.44x), Washington-Arlington-Alexandria $44.20 (1.32x), Seattle-Tacoma-Bellevue $44.13 (1.32x), Boston-Cambridge-Newton $43.09 (1.28x), New York-Newark-Jersey City $41.50 (1.24x), Denver-Aurora-Centennial $39.28 (1.17x), Atlanta-Sandy Springs-Roswell $34.57 (1.03x), Chicago-Naperville-Elgin $34.42 (1.03x, May 2024), Dallas-Fort Worth-Arlington $33.96 (1.01x), Phoenix-Mesa-Chandler $33.48 (1.00x). We damp the top end of those raw ratios. All-occupation means exaggerate the gap for the careers on this site, because national pay bands, remote hiring and company-wide equity grids compress geographic spread for high-skill professional roles more than they do for service and hourly work. Levels.fyi shows the same damping: its Bay Area software engineer average total compensation of about $291k sits roughly 1.3x-1.4x the US median, not 1.7x. Non-US multipliers convert local market rates to USD and are directional, not survey-grade.
How to become a AI Engineer
Every route we could verify, with honest time, cost and difficulty.
There are 5 routes we could verify into AI Engineer work. The fastest is Software engineer adds AI at about 9 months; the cheapest is Software engineer adds AI at about $400. Most AI Engineer hires are existing software engineers who added the AI layer, very often by volunteering for the AI feature nobody at their current employer had owned yet; the honest route for someone with no programming background runs through 12 to 18 months of becoming an employable application developer first, which is why the range above is so wide.
Software engineer adds AI
The fastest and most common route. If you can already build and ship a backend or full-stack application, the AI-specific layer - APIs, embeddings, retrieval, evaluation, agents - is 200-400 hours of study plus two real projects. Many people make this move inside their current company by volunteering for the AI feature nobody has owned yet.
Data scientist or ML engineer pivots
You already have the modelling intuition and the evaluation instinct; what you need is application engineering (APIs, async, front-end integration, deployment) and the LLM-specific stack. Usually a 6-12 month transition and frequently a pay rise, because AI engineer reqs are often easier to fill from this direction than from application engineering.
Career changer via application development
The honest sequence for someone starting from zero: learn Python and web development properly, get a junior developer or technical support engineer job, then move into AI work from there. Trying to skip straight to 'AI engineer' without being able to ship software is the most common failed plan in this field, because the interview is a software interview with LLM topics layered on.
Domain expert plus AI tooling
Lawyers, clinicians, underwriters, analysts and support leads who learn enough Python and LLM engineering to build the internal tool their function needs. This is an unusually strong position in 2026 because the hardest part of most enterprise AI projects is knowing what 'correct' looks like, and that is domain knowledge. Often becomes an internal AI lead role rather than a conventional engineering one.
Build in public
Ship a real AI product - open source or a small paid tool - with users, an evaluation suite and a public write-up of what failed. In a field with no established credential, a system strangers use is the strongest signal available and routinely converts into interviews that a resume would not. Pairs with any of the paths above rather than replacing them.
AI Engineer roadmap: 9 steps, 1,020 hours
Becoming a AI Engineer from zero takes about 1,020 study hours across 9 steps, roughly 9 to 30 months at 10-15 hours a week plus a job search. Step one is Be able to ship software. The roadmap starts with shipping ordinary software because an AI Engineer is an application engineer whose application calls a model, and almost every failed attempt at this career change produces someone who can write a prompt but cannot build the system around it. Evaluation comes at step five, before agents and before production work, because it is the skill interviewers use to separate professionals from demo-builders.
0 of 9 steps done · saved in this browser ·
-
1
Python to production standard: async/await, HTTP clients, FastAPI, pydantic validation, environment and secrets management, pytest, Docker, and deploying to a cloud host. Git and GitHub Actions. Enough TypeScript and React to put a usable interface on what you build.
Why now: An AI engineer is an application engineer whose application calls a model. Almost every failed attempt at this career change skips this step and ends up with a candidate who can write a prompt but cannot build the system around it. If you can already do this, skip ahead - and if you cannot, nothing later in this roadmap will land.
- course Automate the Boring Stuff with Python, then the FastAPI official tutorialAl Sweigart / FastAPI · 60 h · Free
- course The Missing Semester of Your CS EducationMIT · 15 h · Free
- course Associate Python Developer career trackDataCamp · 45 h · Included in DataCamp Premium, about $28/month billed annually
-
2
Tokenisation and why token counts drive cost and latency; embeddings and vector similarity; context windows and the lost-in-the-middle effect; temperature, top-p and sampling; the transformer and attention at a conceptual level; what pretraining, instruction tuning and RLHF each contribute; what a model can and cannot know.
Why now: Interviewers open here, and it separates people who understand the system from people who have memorised a framework's API. It is also practical: knowing that retrieved chunks in the middle of a long prompt are systematically underused changes how you build retrieval.
- course Hugging Face LLM CourseHugging Face · 40 h · Free
- course Deep Learning Specialization (Course 5: Sequence Models, for attention and transformers)Coursera (DeepLearning.AI) · 30 h · Free to audit; $49/month or Coursera Plus for the certificate
- book AI Engineering: Building Applications with Foundation ModelsO'Reilly (Chip Huyen) · 35 h · About $70 print (532 pages), or an O'Reilly subscription
-
3
System prompts, few-shot examples, chain-of-thought and when it helps, structured output with JSON schemas, tool and function calling, prompt templating and versioning, and managing the context window deliberately - what goes in, in what order, and what gets dropped.
Why now: Prompts are code: they belong in version control with a change log and a deploy process, and interviewers ask specifically about that discipline. This is also the cheapest lever in a production system - most quality problems are fixed by context design before anyone needs to fine-tune.
- course ChatGPT Prompt Engineering for Developers and LangChain for LLM Application DevelopmentDeepLearning.AI · 6 h · Free during the platform beta
- practice Anthropic prompt engineering documentation and interactive tutorialAnthropic · 15 h · Free
- practice OpenAI structured outputs and function calling guidesOpenAI · 10 h · Free
-
4
Document parsing and chunking strategies, embedding model choice, vector stores (start with pgvector), hybrid search combining BM25 with dense retrieval, reranking, metadata filtering, citation and grounding, and handling the case where nothing relevant exists. Then measure retrieval quality separately from generation quality.
Why now: RAG is the single most common thing an AI engineer is asked to build, and the most common thing built badly. Almost every 'the model hallucinates' complaint is a retrieval failure. Track the no-retrieval rate and the hallucination rate separately - interviewers are impressed by candidates who instrument the retrieval layer rather than blaming the model.
- course Building and Evaluating Advanced RAG ApplicationsDeepLearning.AI · 4 h · Free during the platform beta
- practice LlamaIndex documentation and end-to-end RAG examplesLlamaIndex · 30 h · Free
- practice pgvector - build a retrieval layer on Postgres before reaching for a managed vector DBpgvector · 20 h · Free
-
5
Build a golden dataset of 100-300 real examples with expected behaviour. Write deterministic checks where possible (schema validity, citation presence, refusal correctness) and LLM-as-judge rubrics where not. Learn RAGAS-style faithfulness and answer-relevance scoring. Wire evaluation into CI so a prompt change that regresses quality fails the build. Add tracing so you can see every step of a failed run.
Why now: This is the highest-leverage step in the entire roadmap and the most reliable interview differentiator. Non-deterministic systems fail in ways a normal test suite never catches - prompt regressions, retrieval drift, runaway tool calls - and the engineer who can say 'here is my eval set and here is the regression gate' is immediately more credible than one who says 'it seemed better'.
- practice RAGAS - automated RAG evaluation frameworkExploding Gradients · 20 h · Free, open source
- course Automated Testing for LLMOps and Evaluating and Debugging Generative AIDeepLearning.AI · 5 h · Free during the platform beta
- book AI Engineering (evaluation chapters)O'Reilly (Chip Huyen) · 15 h · About $70 print, or an O'Reilly subscription
-
6
Tool definitions and schemas, the plan-act-observe loop, state and memory, multi-step control flow with LangGraph or an equivalent, human-in-the-loop checkpoints, retry and timeout policy, sandboxing side effects, and the Model Context Protocol for connecting tools and data sources. Build one agent that does something genuinely useful and then break it deliberately.
Why now: Agentic systems are where most 2025-26 AI engineering budgets went and where interviews concentrate. The questions are about failure modes and observability, not about frameworks: runaway tool calls, retrieval drift, cost blowouts, and how you stop an agent that has decided to delete something. Building one and watching it misbehave teaches this faster than any course.
- course Hugging Face AI Agents Course (smolagents, LlamaIndex and LangGraph)Hugging Face · 28 h · Free, including the certification
- course AI Agents in LangGraphDeepLearning.AI (with Harrison Chase and Rotem Weiss) · 4 h · Free during the platform beta
- practice Model Context Protocol specification and reference serversAnthropic / MCP · 20 h · Free
-
7
Token accounting and per-request cost budgets, semantic and exact caching, batching, streaming responses, routing cheap requests to smaller models, self-hosting with vLLM or Ollama when volume justifies it, rate limits and backpressure. Then safety: prompt injection defence, PII redaction, output validation, refusal handling, audit logs and human review paths.
Why now: The difference between a prototype and a product is usually a 10x cost reduction and a defensible answer to 'what happens when a user pastes a malicious instruction into a document you retrieve?'. Enterprise buyers ask about the second one in procurement; your CFO asks about the first after the first month's bill.
- practice OWASP Top 10 for Large Language Model ApplicationsOWASP · 15 h · Free
- cert AWS Certified Machine Learning Engineer - Associate (Bedrock and SageMaker deployment content)Amazon Web Services · 80 h · $150 ($75 during the MLA-C02 beta)
- practice vLLM documentation - self-hosted inference, batching and quantisationvLLM · 20 h · Free
-
8
System one: a RAG application over a corpus you know well - your industry's regulations, a company handbook, a large open dataset - with citations, an evaluation suite of at least 100 cases, cost per query measured, and a public write-up. System two: an agent that performs a multi-step task with tools, with guardrails, tracing and a documented failure analysis. Deploy both. Get at least a few real users.
Why now: There is no established credential for this role, so a working system with an evaluation suite is the credential. The write-up matters as much as the code: hiring managers want to see that you measured something, found it wanting, and fixed it. 'Here is my eval harness and here is what it caught' is the most persuasive sentence available to you.
- practice Open-source AI Cookbook - production recipes and patternsHugging Face · 30 h · Free
- practice Langfuse or LangSmith for tracing and evaluation on your own projectLangfuse · 20 h · Free self-hosted / free tier
- book AI Engineering: Building Applications with Foundation ModelsO'Reilly (Chip Huyen) · 35 h · About $70 print, or an O'Reilly subscription
-
9
Five areas: LLM fundamentals, prompt and context engineering, RAG and vector search, agents and tool use, and production operations. Practise an LLM system design question out loud - 'design a customer support assistant over 200,000 documents' - covering ingestion, retrieval, evaluation, guardrails, latency and cost. Keep a coding round in your preparation: 80-150 LeetCode mediums, because most loops still include one.
Why now: Interviewers are looking for someone who understands both the models and the engineering around them: how to ground a model with retrieval, how to evaluate a non-deterministic system, and how to control latency and cost. Candidates who can only discuss frameworks fail; candidates who can reason about tradeoffs pass even when they have not used the company's stack.
- practice LeetCode Top Interview 150LeetCode · 80 h · Free; Premium about $35/month
- book AI Engineering (Chip Huyen) - the most-read title on the O'Reilly platform in 2025O'Reilly · 25 h · About $70 print, or an O'Reilly subscription
- practice Hugging Face Agents Course final challenge (graded, free certificate)Hugging Face · 10 h · Free
Best certifications for a AI Engineer
Which ones matter, what they cost, and how often people pass.
No certification is required for an AI Engineer job and none will get you hired on its own. The best value for time is the Hugging Face AI Agents Course, which is free including its certification and covers smolagents, LlamaIndex and LangGraph in about 25 to 30 hours. If your employer is an Amazon Web Services or Microsoft Azure shop, the AWS Certified Machine Learning Engineer - Associate at $150 (or $75 during the MLA-C02 beta) or Microsoft Exam AI-102 at $165 in the United States signals platform competence to that specific buyer; the DeepLearning.AI short courses are a curriculum rather than a credential and carry no screening weight at all.
Hugging Face AI Agents Course (Fundamentals and Completion certificates)
Hugging Face
- Cost
- Free - the course and the certification process are both free
- Study
- About 25-30 (4 core units plus 3 bonus units, roughly 3-4 hours per unit)
- Pass rate
- Not published. Fundamentals certificate requires completing Unit 1; the Completion certificate requires Unit 1, a use-case assignment and the final challenge.
DeepLearning.AI short courses (LangChain for LLM Application Development, Functions Tools and Agents with LangChain, AI Agents in LangGraph and others)
DeepLearning.AI
- Cost
- Free during the DeepLearning.AI learning-platform beta; each course is 1-2 hours
- Study
- 1-2 hours each; 20-40 hours to work through the agent and RAG-relevant set
- Pass rate
- No exam; completion-based
AWS Certified Machine Learning Engineer - Associate (MLA-C01 / MLA-C02)
Amazon Web Services
- Cost
- $150 USD (MLA-C01); $75 USD during the MLA-C02 beta
- Study
- 60-100 hours of prep
- Pass rate
- Not published. Recommended background is about a year with Amazon SageMaker AI, Amazon Bedrock and related services. MLA-C01's last English testing day is 28 September 2026.
Microsoft Certified: Azure AI Engineer Associate (Exam AI-102)
Microsoft
- Cost
- $165 USD in the US - Microsoft's standard associate-level exam fee; price varies by country
- Study
- 60-90 hours of prep
- Pass rate
- Not published; Microsoft passing score is 700/1000. Renewal is free annually via an online assessment.
Deep Learning Specialization
DeepLearning.AI (via Coursera)
- Cost
- Free to audit; $49/month or Coursera Plus ($59/month, $399/year) for graded work and the certificate
- Study
- 129 across 5 courses
- Pass rate
- No exam; 4.8/5 from 147,248 reviews
Skills employers screen for
Soft skills that decide offers: Setting expectations with stakeholders who think the demo is the product, Deciding when an LLM is the wrong tool, Designing for graceful failure when the system is wrong 5% of the time, Writing evaluation criteria with domain experts, Keeping up with the field without chasing every release.
Best courses for a AI Engineer
Checked on the provider's page on 16 September 2026. Some links are affiliate links.
We list 6 courses for AI Engineer work, checked on the provider's page. Take one structured course to fix the gaps - the Hugging Face LLM Course for model fundamentals or the Hugging Face AI Agents Course for agents, both free - and then stop taking courses and spend the remaining hours and your API budget building the two systems in step eight, because the portfolio is what the interview is actually about.
Browse more AI Engineer courses on Coursera Coursera Plus covers most of the courses above for one monthly fee.
AI Engineer interview questions and format
An AI Engineer loop is typically four to five stages over three to six weeks: a recruiter screen, a coding round, an LLM depth interview, an LLM system design interview, and a behavioural and product-judgment round; startups often swap the middle for a four-to-eight-hour take-home you then defend. The stage that filters most candidates is the LLM system design round, because it is where people who only know a framework's API run out of things to say about evaluation, retrieval quality, guardrails and cost.
Usually 4-5 stages over 3-6 weeks, and the shape varies more than for any other role here. (1) Recruiter screen. (2) A coding round - most loops still include a standard data-structures-and-algorithms screen or a practical Python exercise; some have replaced it with an AI-assisted coding round where you use an assistant and are judged on direction and verification. (3) LLM depth: fundamentals, prompting, RAG, fine-tuning tradeoffs, evaluation. (4) LLM system design: design a production AI feature end to end - ingestion, retrieval, prompt/context strategy, evaluation, guardrails, observability, latency and cost. (5) Behavioural and product judgment, often including 'when would you not use an LLM here?'. Startups frequently substitute a take-home of 4-8 hours - build a small RAG or agent system - followed by a review where you defend every decision. The five areas tested across 2026 loops are LLM fundamentals, prompt engineering, RAG and vector databases, agents, and production operations.
Questions that come up
- Design a customer support assistant over 200,000 internal documents. Cover ingestion, retrieval, evaluation, guardrails, latency and cost.
- How do you evaluate a RAG system? Name the metrics and say how you would build the dataset.
- Users say the system hallucinates. Walk me through how you diagnose whether it is a retrieval problem or a generation problem.
- When would you fine-tune instead of improving retrieval or the prompt? What does fine-tuning not fix?
- Explain chunking strategies and their tradeoffs. How does chunk size interact with the lost-in-the-middle effect?
- What is prompt injection, and how do you defend a system that retrieves untrusted documents?
- Your agent occasionally makes 40 tool calls for a simple request. How do you find out why and how do you contain it?
- Your inference bill tripled this month. How do you bring it down without a visible quality drop?
- How do you version and deploy prompts, and how do you stop a prompt change from regressing quality?
- Compare dense, sparse and hybrid retrieval, and say when reranking is worth its latency.
Prep
- AI Engineering: Building Applications with Foundation Models (Chip Huyen)
- Hugging Face AI Agents Course (free, with certificate)
- DeepLearning.AI short courses on RAG, agents and LLMOps
- RAGAS documentation - RAG evaluation metrics
- OWASP Top 10 for LLM Applications
- LeetCode Top Interview 150 (for the coding round)
AI Engineer FAQ
What is the difference between an AI engineer and an ML engineer?
An ML engineer builds and trains models and the infrastructure to run them - training pipelines, distributed compute, serving, drift monitoring. An AI engineer builds products on top of models somebody else trained - retrieval, prompting, agents, evaluation, guardrails, cost. The ML role demands more mathematics and systems depth and pays higher at the median ($280,000 versus $154,000 on Levels.fyi); the AI role demands more application engineering and product judgment and is far easier to enter. The boundary blurs at senior level and many people do both.
Do I need to know how to train a model to work as an AI Engineer?
Not to do the job, and not to pass most interviews. You do need to understand training well enough to reason about tradeoffs - what pretraining and instruction tuning give you, what fine-tuning with LoRA or QLoRA can and cannot fix, and why more retrieval usually beats more fine-tuning for factual accuracy. Interviewers ask conceptual questions about transformers and attention; they rarely ask you to implement one.
Which certification should I get to become an AI Engineer in 2026?
None is required and none will get you hired alone. The best value for time is the Hugging Face AI Agents Course, which is free including certification and covers smolagents, LlamaIndex and LangGraph in roughly 25-30 hours. If you are in an AWS or Azure enterprise, the AWS Certified Machine Learning Engineer - Associate ($150, or $75 during the MLA-C02 beta) or Microsoft's AI-102 ($165 in the US) signals platform competence to that employer. The DeepLearning.AI short courses are free during the platform beta and are the fastest way to cover RAG, agents and LLMOps. Spend the rest of your budget on API credits for building, not on credentials.
Is the AI Engineer role a bubble, or is the demand behind it durable?
The compensation at the top is bubble-shaped; the work is not. A handful of frontier labs paying $800,000-plus medians for software engineers is a bidding war among a small number of extremely well-funded employers, and those numbers can compress quickly. But the underlying demand - companies needing engineers who can make model outputs reliable enough to put in front of customers - is broad and growing across ordinary industries. Plan your career on the broad market ($120,000-$220,000 for most roles) and treat the tail as an option, not a plan. Anchor your skills in evaluation, retrieval and systems design, which survive framework churn.
Can I become an AI Engineer without any programming background at all?
Not directly, and anyone selling you a three-month path to an AI engineer job is misleading you. The interview loops include a coding round at a real software-engineering bar. The honest route from zero is 12-18 months to become an employable application developer, a year or two in that job, then the AI transition - about 30 months total. The faster route, if it applies to you, is to use a domain you already have: a paralegal, underwriter or clinician who learns Python and LLM engineering is more valuable on an enterprise AI team than a generalist, because they know what a correct answer looks like.
What actually gets an AI engineer hired in 2026?
A deployed system with an evaluation suite, in that order of importance. Not a certificate, not a list of frameworks, not a tutorial repo. Hiring managers have seen a thousand RAG demos; what they have not seen is a candidate who says 'I built a golden set of 180 cases, my retrieval recall was 0.71, here is what I changed and here is the number afterwards.' Add a public write-up of a failure you diagnosed and you are in the top decile of applicants.
Which framework should I learn - LangChain, LlamaIndex, or neither?
Learn the concepts through one of them and be prepared to work without it. LangChain and LangGraph are the most commonly named in job postings and the DeepLearning.AI and Hugging Face courses cover them, so they are a reasonable default. LlamaIndex is stronger for document-centric retrieval. Plenty of production teams use the model APIs directly with their own thin abstraction, and being able to explain why is a good interview answer. Never build your identity around a framework in a field this young.
What does an AI Engineer actually do all day?
Mostly measurement and debugging, not prompt writing. A normal day opens with overnight traces - the support assistant escalated 14 percent of conversations instead of 9 percent, so you find the cause, which turns out to be a vendor model update that changed refusal behaviour, and you pin the version. Then you run the evaluation suite, argue with product about whether an agent should be allowed to issue refunds without human approval, rewrite a chunking strategy to lift retrieval recall, and cut cost by routing cheap requests to a smaller model. Roughly one day in five is genuinely new feature building.
Is AI Engineer a good career for someone switching at 40 from a non-technical job?
Yes if you already write code, and only with a longer plan if you do not. Age is not the filter here; the coding round is. If you can ship software, the AI layer is 200 to 400 hours plus two real projects and the switch takes about nine months. From a non-technical start, budget roughly 30 months. The strongest position for a 40-year-old is the one younger candidates cannot occupy: pair your existing domain - underwriting, law, clinical work, tax - with Python and LLM engineering, because knowing what a correct answer looks like is the hardest part of most enterprise AI projects.
How much does an AI Engineer make in the San Francisco Bay Area versus the national median?
Expect roughly 20 to 40 percent above the national figure. Levels.fyi puts the United States AI Engineer median total compensation at $154,000, so a Bay Area median lands around $185,000 to $215,000, and the same multiplier moves the $215,000 senior figure toward $260,000 to $300,000. Two cautions. The title is applied loosely, so a Bay Area offer can still be ordinary application-engineering pay with an AI label. And the headline numbers come from a handful of employers: Levels.fyi software engineer medians are $880,000 at OpenAI and $820,000 at Anthropic, which is a few thousand jobs, not a market you can plan on.
Sources
Every number on this page traces to one of these. Page checked 16 September 2026.
- levels.fyi/t/software-engineer/title/ai-engineer/locations/united-states
- levels.fyi/companies/google/salaries/software-engineer/title/ai-engineer
- levels.fyi/companies/microsoft/salaries/software-engineer/title/ai-engineer
- levels.fyi/companies/openai/salaries/software-engineer
- levels.fyi/companies/anthropic/salaries/software-engineer
- levels.fyi/companies/scale-ai/salaries/software-engineer
- glassdoor.com/Salaries/ai-engineer-salary-SRCH_KO0,11.htm
- glassdoor.com/Salaries/ai-and-machine-learning-engineer-salary-SRCH_KO0,32.htm
- bls.gov/ooh/computer-and-information-technology/software-developers.htm
- bls.gov/ooh/math/data-scientists.htm
- hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/
- huggingface.co/learn
- huggingface.co/learn/agents-course/unit0/introduction
- deeplearning.ai/courses/
- deeplearning.ai/courses/ai-agents-in-langgraph
- aws.amazon.com/certification/certified-machine-learning-engineer-associate/
- learn.microsoft.com/en-us/credentials/certifications/azure-ai-engineer/
- coursera.org/specializations/deep-learning
- coursera.org/courseraplus
- course.fast.ai/