What is AI engineering?

Short answer

AI engineering is the discipline of building, shipping and running software products whose behaviour depends on AI models, today most often foundation models used through APIs. The core of the job is the system around the model: framing the problem, choosing the approach, assembling context, evaluating outputs, and keeping quality, cost, latency and reliability under control. AI engineers usually adapt existing models rather than train new ones.

Written for software engineers, students, managers and recruiters who need a precise definition of AI engineering and how it differs from neighbouring roles.

By Dr. Dulan Dias, co-founder and CTO of Tellme AIPhD in AI, Queen’s University Belfast · Author of From Model to Product · Updated 4 November 2026

Key facts

Carnegie Mellon's Software Engineering Institute defines AI engineering as a field of research and practice that combines systems engineering, software engineering, computer science and human-centred design to create AI systems that meet human needs. (Source: SEI, Carnegie Mellon University: AI Engineering)
A June 2023 essay on the Latent Space newsletter, The Rise of the AI Engineer, described the AI engineer as a software engineering specialism and noted that someone could succeed in the role without ever training a model. (Source: Latent Space: The Rise of the AI Engineer)
Chip Huyen distinguishes applications on traditional ML models, which involve more data annotation, feature engineering and model training, from applications on foundation models, which involve more prompt engineering, context construction and parameter-efficient fine-tuning. (Source: Chip Huyen, AI Engineering book repository)
Google Cloud's MLOps guidance states that only a small fraction of a real-world ML system is composed of the ML code. (Source: Google Cloud: MLOps, continuous delivery and automation pipelines in ML)
The 2015 NeurIPS paper Hidden Technical Debt in Machine Learning Systems, by D. Sculley and colleagues at Google, found it common for real-world ML systems to incur massive ongoing maintenance costs. (Source: NeurIPS 2015: Hidden Technical Debt in ML Systems)
LinkedIn's Jobs on the Rise 2026 report, published in January 2026, lists AI engineer first among the fastest-growing roles in the UK, with most jobs in London, Manchester and Cambridge. (Source: LinkedIn News: Jobs on the Rise 2026, Europe)
Chapter 18 of From Model to Product argues that the AI engineer's role has moved from model builder to system designer. (Source: From Model to Product, Chapter 18)

Definitions

Foundation model: A large model trained on broad data, such as a large language model, that can be adapted to many tasks through prompting, retrieved context or fine-tuning instead of being trained for one task.
Machine learning engineering: Building and maintaining models trained on an organisation's own data: features, training pipelines, offline evaluation, serving and retraining.
MLOps: In Google Cloud's definition, an ML engineering culture and practice that unifies ML system development and ML system operation, through automated pipelines, testing, deployment and monitoring.
Evals: Systematic tests of model or application output against a fixed set of examples and scoring rules, used to compare prompts, models and releases before and after they ship.

Where does the term AI engineering come from?

The term has two lineages. Carnegie Mellon's Software Engineering Institute uses it for a research and practice field that brings systems and software engineering to AI, because AI systems that work in controlled settings often fail outside them. Its framing is broad and centres on systems built for mission outcomes.

The everyday industry sense arrived in 2023. An essay on the Latent Space newsletter described the AI engineer as a software engineer specialising in applying AI and using the new model stack well, and argued that shipping AI products needs engineers more than researchers. Chip Huyen's book AI Engineering, released by O'Reilly in December 2024, then framed the discipline as building applications on top of foundation models.

Both senses agree on the core: the work is engineering a dependable system, not inventing a model.

AI engineering vs ML engineering vs MLOps vs data science vs software engineering

The five disciplines overlap, but each starts from a different question. The table shows where each one usually sits.

Discipline	Core question	Typical work	Trains models?
AI engineering	How do we build a dependable product on top of AI models?	Model choice, prompts and context, retrieval, evals, guardrails, cost and latency control, integration	Rarely; adapts existing models
ML engineering	How do we train and serve a model that performs on our data?	Features, training pipelines, offline evaluation, serving and retraining	Yes, routinely
MLOps	How do we ship and run models repeatably?	Pipelines, versioning, CI/CD for models and data, monitoring and retraining triggers	Builds the pipelines that do
Data science	What does the data tell us, and what should we decide?	Analysis, experiments, statistical models, dashboards and recommendations	Sometimes, mostly for analysis
Software engineering	How do we build correct, maintainable software?	Requirements, design, code, tests, deployment and operations	No

What does an AI engineer do day to day?

An AI engineer spends most of the day on the system around the model. A typical week moves through these jobs:

Frame the problem: what the feature must do, how good is good enough, and whether AI is needed at all.
Choose the approach and model: rules, a small classifier, a hosted LLM or an open-weight model, weighed on quality, latency and cost.
Assemble context: prompts, retrieval over the organisation's data, structured output schemas and tool definitions.
Evaluate: build test sets, score outputs, catch regressions before release and study failures.
Integrate and harden: APIs, timeouts, retries, fallbacks, caching, rate limits and guardrails.
Operate: monitor quality, latency, cost per request and drift, and handle incidents.
From practice: as founding engineer at Tellme AI, Dr. Dulan Dias built the production MVP end to end within 4 to 6 months, covering a Next.js front end, a FastAPI and MongoDB back end, an admin portal, CI/CD and AWS infrastructure. Model work, such as optimising computer vision to cut battery drain on low-end devices by over 50%, was one strand among many, which is the usual balance of the role.

How did large language models change the role?

Before foundation models, most applied AI meant training a model on your own data, so the scarce skills were labelling, feature engineering and training. With foundation models the model arrives ready, and the work shifts to adapting it: prompts, context construction and, where needed, parameter-efficient fine-tuning.

Three things changed most. Evaluation became the central skill, because open-ended outputs rarely have one correct answer. Cost moved from training to inference, so every call has a price that scales with users. And the barrier to a demo fell, so more teams reach production with little experience of what breaks there.

What did not change is the system around the model. Google's MLOps guidance notes that only a small fraction of a real-world ML system is the ML code, a point made in the 2015 paper Hidden Technical Debt in Machine Learning Systems. Chapter 18 of From Model to Product applies it to the role: “The future AI engineer is not only a model builder. They are a system designer.”

What AI engineering is not

Not prompt writing alone. Prompts matter, but without evaluation, retrieval and production engineering they are guesses.
Not research. Inventing architectures and training frontier models is research and ML engineering; AI engineers use the results.
Not only LLMs. Classifiers, computer vision and plain rules remain the right tool for many decisions; see rules vs machine learning vs LLMs.
Not the same as agentic AI. Agents are one application pattern within it, covered in what agentic AI is.

Is AI engineer a recognised job title in the UK?

AI engineer is now a mainstream job title in the UK. LinkedIn's Jobs on the Rise 2026 report, published in January 2026 and based on job starts from January 2023 to July 2025, lists AI engineer first among the UK's fastest-growing roles, with most jobs in London, Manchester and Cambridge.

LinkedIn also lists machine learning engineer as another name for the role, a sign of how loosely employers draw the line. Read the responsibilities rather than the title: a post about prompts, retrieval and evals is AI engineering whatever it is called. How to become an AI engineer covers the route in.

Frequently asked questions

Is AI engineering the same as machine learning engineering?

AI engineering and machine learning engineering overlap but start from different places. ML engineering usually starts from an organisation's data and a model to train, then serves and retrains it. AI engineering usually starts from a capable pretrained model and a product to build, then adapts, evaluates and operates it. In small teams one person often does both.

Do AI engineers train models?

AI engineers rarely train models from scratch. Most of their work adapts existing foundation models through prompts, retrieved context and structured outputs, and some fine-tune smaller models when evaluation shows prompting is not enough. Training classical models, such as a cheap classifier that routes requests, is still part of the toolkit and often the cheaper answer.

Is AI engineering just prompt engineering?

Prompt engineering is one part of AI engineering, not the whole of it. A good prompt still needs the right context retrieved, outputs validated against a schema, an evaluation set that proves it works, monitoring once it ships, and fallbacks for when the model fails or the provider is down. Those surrounding pieces take most of the effort.

Do you need advanced maths for AI engineering?

Applied AI engineering needs less maths than research does. You need working statistics for evaluation: sampling, variance, precision and recall, and how to tell a real improvement from noise. Linear algebra helps with embeddings and similarity search. Calculus and optimisation theory matter for training new models, which AI engineers rarely do.

Is AI engineering part of software engineering?

AI engineering is best understood as a specialism within software engineering. It keeps every software discipline, including APIs, testing, version control, CI/CD and operations, and adds probabilistic outputs, evaluation, data and context management, and per-call model costs. Engineers who are strong at the software half usually find the AI half the quicker one to learn.

Will AI engineering still be needed as models get better?

Better models remove some work, such as elaborate prompt workarounds, but not the system around them. Products still need problem framing, data access, evaluation, security, cost control and reliability, and more capable models tend to be given more consequential tasks, which raises those stakes. The tools will change faster than the engineering problems.

What programming languages do AI engineers use?

Python is the default for AI engineering because most model libraries, evaluation tools and data tooling are written for it. TypeScript is common for the application layer, especially in web products, and most providers ship SDKs for both. SQL matters for retrieval and analytics. The language matters less than knowing how to build reliable services in it.

Sources

About the author

Dr. Dulan Dias is co-founder and Chief Technology Officer of Tellme AI in Belfast; founder of EchonLabs, grown from 2017 to more than 60 people before its acquisition in 2026; holder of a PhD in AI and data science from Queen’s University Belfast; and the author of From Model to Product.

Who is Dr. Dulan Dias?