So what is an AI engineer? In one line: an AI engineer is a software engineer who builds products on top of AI models that already exist, such as large language models (LLMs) reached through an API, or open models run on your own servers. They rarely train the model. They build what surrounds it: retrieval, prompts, tools, tests, and the limits that keep it safe, fast and affordable.
Chip Huyen, author of the O'Reilly book AI Engineering (2025), defines the discipline as building applications with readily available foundation models: big general-purpose models, like LLMs, that many products build on.
Where did the title "AI engineer" come from?
One essay made the title widely known. On 30 June 2023, the newsletter Latent Space published "The Rise of the AI Engineer" under the byline swyx (Shawn Wang). It predicted that software engineers building on model APIs and open models would settle on one title, the AI Engineer, and that it would likely become the highest-demand engineering job of the decade. That was a forecast, not a measurement. The essay also said it was pointing at a trend already under way, so swyx popularized the title rather than inventing it.
Its key idea was an API line. ML researchers and ML engineers work on the models themselves; AI engineers build products on top of them through APIs, and own the product's data and evals. People cross the line in both directions.
"When it comes to shipping AI products, you want engineers, not researchers." — swyx, Latent Space
Andrej Karpathy replied the same day. He had been calling the role Prompt Engineer, but agreed the name misled people: "It's not just prompting alone, there's a lot of glue code/infra around it."
How is an AI engineer different from an ML engineer or a data scientist?
The roles overlap, so the clearest test is what each one delivers.
| Role | Main output | Typical work |
|---|---|---|
| Data scientist | Insight from data | Analysis, building and testing models, advising decision-makers |
| ML engineer | Models that are trained and served | Training, deploying, tuning and monitoring models |
| AI engineer | Products built on existing models | Prompts, context, retrieval, agents, evals, guardrails |
Huyen describes the difference as where the effort goes: traditional ML work means more data labelling, feature engineering and model training; AI engineering means more prompting, context building and light fine-tuning. The edges blur: Google Cloud's ML engineer certification now expects familiarity with prompt and context engineering too, and LinkedIn says the AI engineer role can also be known as machine learning engineer.
To see where employers draw the line, we read the job descriptions in our own count of 1,768 public company job boards on 29 Oct 2026. Roles with AI Engineer titles (223 at 136 companies) and Machine Learning Engineer titles (220 at 129 companies) were about equally common. A hand check of 40 random AI Engineer roles found 31 centred on building with existing models.
The ads ask for different things, though. AI Engineer postings mentioned agents (79% against 31%), retrieval (56% against 22%), prompts (48% against 9%) and guardrails (34% against 11%) far more often. ML Engineer postings mentioned PyTorch, TensorFlow or JAX five times as often, training models three times as often and a PhD four times as often.
- AI Engineer
- Machine Learning Engineer
Figure 2. Share of postings whose description mentions each term (mentions, not requirements). Terms for training models sit below the dashed line. Source: our own count of 1,768 public company job boards, roles first posted 2 May to 29 Oct 2026, counted 29 Oct 2026.
Product engineers, by contrast, own a whole product and often build AI features inside it; see What is a product engineer? and our comparison of product, forward deployed and AI engineers.
What does an AI engineer actually build?
Back to the help desk, and the five parts in the diagram at the top.
Retrieval: giving the model the right facts
A model doesn't know your manuals. Retrieval searches your documents for passages that match the question and sends them to the model with the prompt. This is retrieval-augmented generation (RAG), named in a 2020 paper by Patrick Lewis and 11 co-authors. In practice it means splitting documents into chunks, turning them into embeddings (lists of numbers that capture meaning) in a vector database, and choosing what to send.
Agents: letting the model use tools
An agent is a model that can act, not only write. Huyen defines an agent by its environment and its tools, and calls RAG a special case whose tool is a retriever. Many agents follow the ReAct pattern from a 2022 paper: reason about the next step, call a tool such as an order lookup, read the result, repeat. The engineering is in the limits: which tools, what the agent may do alone, and when it hands over to a person.
Evaluations: knowing whether it works
Huyen notes that foundation models are harder to evaluate than traditional ML models, because their answers are open-ended. An eval is a test suite for model output: real questions with known good answers, scored on every change by code, people or another model. In Stack Overflow's 2025 survey, developers' top frustration with AI tools, named by 66%, was answers that are "almost right, but not quite". Evals catch those before users do.
Guardrails: keeping it safe
A good checklist to start from is the OWASP Top 10 for LLM Applications 2025. It starts with prompt injection, where text in a question or a document tricks the model into ignoring its instructions, and also covers leaked data, agents with too much power and runaway use. NIST's Generative AI Profile (July 2024) names a related risk, confabulation: confident answers that are false. Guardrails are plain code: filter the input, check the output, limit the tools, log everything.
Cost and latency: keeping it affordable and fast
Every call costs money and time. OpenAI's Batch API, for example, charges half the normal price for results within 24 hours (its docs as of 29 Oct 2026). Speed isn't only an AI engineer's concern: in our count, ML Engineer postings mentioned latency more often (34% against 22%).
What skills and tools does an AI engineer need?
Start with software engineering. On LinkedIn's 2026 US list, the top roles people moved into AI engineering from were software engineer, data scientist and full-stack engineer, with a median of 3.7 years of prior experience. The ads in our count read like product software jobs:
- Python appeared in 73% of AI Engineer postings.
- TypeScript or Node.js appeared in 31%, against 6% of ML Engineer postings.
- An LLM framework such as LangChain, LangGraph or LlamaIndex appeared in 22%, against 2%.
- A PhD appeared in only 7%.
Is there demand for AI engineers?
Yes, on several measures, each with a caveat.
- LinkedIn Jobs on the Rise. AI engineer was the fastest-growing US job on both the 2025 and 2026 lists, and in 2026 also in the UK, France, Italy, the Netherlands and Spain. These lists rank growth in jobs people started, not size, and LinkedIn's 2026 lists count ML engineers under the same title.
- LinkedIn job postings. In US postings, ML engineer was the more common title in 2023 and 2024, and AI engineer in 2025 and 2026. LinkedIn says part of this shift may be a change in labels, with companies using "AI engineer" as a catch-all for roles once called ML engineering. Forward Deployed Engineer is the third most common AI job in US postings, and Gen Z makes up more than two-thirds of US hires into AI Engineer and Forward Deployed Engineer roles, according to LinkedIn's AI Talent Divide report. More in What is a forward deployed engineer?
- Hacker News. In our own count of Hacker News "Who is hiring?" posts, the share that mention AI Engineers rose from 0.7% in 2023 to 5.2% in 2026 (January to October of each year), while mentions of ML Engineers went from 3.6% to 3.8%.
- Mention AI Engineers
- Mention ML Engineers
Figure 3. Share of top-level posts in "Ask HN: Who is hiring?" threads that mention each title, January to October of each year. Source: our own count of Hacker News "Who is hiring?" posts, counted 29 Oct 2026. The posts lean towards US startups.
These sources measure different things, so read them side by side, not as one series. Pay: in US LinkedIn postings, the typical AI job lists a base salary midpoint of about $177,000, against $80,000 for non-AI jobs, but listed pay for AI engineer roles has been roughly flat since 2024. Those are US dollars from US postings, not pay in Egypt. Juniors: Indeed found that 71% of the recent rise in US software development postings (May 2025 to May 2026) came from senior roles. More in Tech hiring trends in 2026.
Closer to home, we found no LinkedIn Jobs on the Rise list for Egypt in 2025 or 2026. The nearest, the UAE's 2025 list, ranked AI engineer second.
How do you become an AI engineer?
LinkedIn found that the most common routes in are software engineering, data science and full-stack roles, so start there.
- Learn ordinary software engineering: APIs, databases, tests, deployment. Ship one small web app.
- Build one retrieval feature on real documents that shows the sources behind each answer.
- Write evals before tuning: a few dozen real questions, scored on every change.
- Add one tool to make it an agent, with the least power it needs.
- Attack it with prompt injection and the OWASP list; fix what breaks.
- Measure cost and latency per request, then write up the before-and-after scores.
Prefer owning the whole product? Most steps carry over: see How to become a product engineer.
An AI engineer turns an existing model into a product people can trust: the right facts in, safe actions, proof it works, protection against misuse, and sensible cost and speed. The title is young and blurs into ML engineering; the lasting skills are software engineering and testing what a model says.
How does Nilerobot use AI engineering?
Nilerobot is a software studio in Egypt. We design, build, launch and run products, and AI engineering is one of our capabilities: when a product needs an AI feature, we build it with the retrieval, evals and guardrails above. Our own AI-assisted code goes through the checks in What is harness engineering?
Frequently asked questions
Is an AI engineer the same as a machine learning engineer?
No, but the two overlap. ML engineers train, deploy and monitor models; AI engineers build products on top of models that already exist.
Do AI engineers need a PhD or need to train models?
Usually not. In our own count of public company job boards, 7% of AI Engineer postings mentioned a PhD and 12% mentioned training models, against 29% and 37% of Machine Learning Engineer postings.
Who came up with the title AI engineer?
The essay The Rise of the AI Engineer, published on Latent Space on 30 June 2023 under the byline swyx (Shawn Wang), made it widely known. It said it was naming a trend already under way.
How is an AI engineer different from a product engineer or a forward deployed engineer?
An AI engineer owns the AI part of a product. PostHog describes product engineers as developers who build products for real users and write full-stack code; LinkedIn describes forward deployed engineers as helping organizations put AI solutions in place. Our comparison covers all three.
Sources
- The Rise of the AI Engineer, Latent Space (swyx), 30 June 2023.
- Andrej Karpathy's reply to the essay, X, 30 June 2023.
- AI Engineering (O'Reilly, 2025), its book repository and chapter summaries, Chip Huyen, read 29 Oct 2026.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Lewis et al., arXiv, 22 May 2020.
- ReAct: Synergizing Reasoning and Acting in Language Models, Yao et al., arXiv, 6 Oct 2022.
- Professional Machine Learning Engineer, Google Cloud, read 29 Oct 2026.
- Occupational Outlook Handbook: Data Scientists, US Bureau of Labor Statistics, 27 Aug 2026.
- LinkedIn Jobs on the Rise, LinkedIn News: US 2025, US 2026, UK 2026, Europe 2026, UAE 2025; Jan 2025 and Jan 2026.
- The AI Talent Divide, LinkedIn Economic Graph Research Institute, August 2026.
- LinkedIn research on AI hires and postings, LinkedIn Newsroom, 18 Aug 2026.
- OWASP Top 10 for LLM Applications 2025, OWASP Gen AI Security Project, 17 Nov 2024.
- Generative AI Profile (NIST AI 600-1), NIST, July 2024.
- 2025 Developer Survey: AI, Stack Overflow, 2025.
- AI and Job Postings: From Destruction to Creation?, Indeed Hiring Lab, 8 Jul 2026.
- Batch API guide, OpenAI, read 29 Oct 2026.
- What is a product engineer?, PostHog, read 29 Oct 2026.
- Our own count of public company job boards and of Hacker News "Who is hiring?" posts, Nilerobot, 29 Oct 2026.
Need an AI feature people can trust?
AI engineering is one of our capabilities: we build AI features with retrieval, evals and guardrails as part of the products we design, build and run.
See what we build →