MLOps

in English: machine learning operations (MLOps)

advancedwidespread

MLOps is the collection of methods and tools used to reliably deploy, monitor, and continuously update machine learning models.

In plain terms

Building an AI model is like cooking a good meal at home. MLOps is the restaurant kitchen: the same dish must be served every day to hundreds of guests with identical quality, and any spoilage must be detected immediately.

In practice, this is a combination of working methods and software tools. These ensure the model moves from the experimental lab to real users and continues to function well over months and years.

A separate specialty is needed because a model does not break down like a traditional program. It does not stop with an error message but quietly gives increasingly worse answers because the world has changed since it was trained.

MLOps therefore adds constant monitoring of data and model quality to the discipline of traditional software operations (testing, version control, automation).

How it works

Data preparation: the team collects, cleans, and version-controls the training data so that it can be precisely determined later which model learned from what.
Training and experiment tracking: every training run is logged along with settings and results, so the best version can be reproduced at any time.
Verification and deployment: the model undergoes automated tests (accuracy, speed, fairness), then is gradually released to real users.
Observation: in live operation, the model's accuracy is continuously measured, and whether the nature of incoming data has changed (this is called data drift).
Re-training: if performance deteriorates, the system signals or automatically initiates a model update, and the cycle begins again.

An everyday example

Your bank's fraud detection system stops an unusual foreign purchase on your card one day. For this model to recognize new tricks by fraudsters while avoiding unnecessary blocks, the MLOps process ensures this in the background: it monitors the hit rate and regularly retrains the model on fresh data.

Who is affected

developerYou must build and maintain the automated process (pipeline) through which the model moves from experimentation to live operation. → Code

entrepreneur, executiveThis determines whether your AI project remains at the demo stage or delivers lasting business value.

bank customer, investorThe accuracy and currency of the models behind credit assessment and fraud screening directly affect your finances. → Money

industrial worker, engineerQuality control or predictive maintenance models on production lines remain reliable only under regular supervision. → Industry

A common misconception

MythOnce the model is finished and performs well on tests, the hardest part of the work is over.

In realityThe bulk of the work comes only after this: live operation, monitoring, and updates typically require more resources than building the model itself. Without supervision, even the best model will silently become obsolete.

Limits and risks

MLOps does not fix a bad model or bad data: it only ensures you detect the error in time.
Setting it up is expensive and requires expertise; a small company purchasing a ready-made AI service usually does not need its own full MLOps system.
The tool market is fragmented and changes rapidly, making good selection difficult; while major cloud providers' ready-made packages are convenient, they easily lock users into a single vendor.
Automatic retraining can be risky: if incorrect or manipulated data enters the system and is not checked by a human, the model can degrade unnoticed.

Similar, but not the same

DevOps – DevOps connects traditional software development and operations, while MLOps adds the management of data and trained models to this.
LLMOps – LLMOps is the branch of MLOps tailored for large language models, focusing on tracking prompts, costs, and the quality of generated responses.
AIOps – AIOps means using artificial intelligence to support the operation of IT systems, so there AI is the tool, not the thing being operated.

Related concepts

Broader

machine learning, DevOps, AI infrastructure

Narrower

LLMOps, model monitoring, data drift, experiment tracking, modellregiszter

On the edge

AIOps, DataOps, AI governance, model deployment, Continuous Integration and Delivery, cloud service

A brief history

The term emerged in the software world in the late 2000s following the DevOps model and spread in the late 2010s. One of the field's foundational texts is a 2015 study by Google researchers, which pointed out that in a production machine learning system, the model itself is only a small fraction of the code, while the rest serves data handling, operations, and supervision.

Where things stand

By the end of 2026, MLOps is established practice at larger companies that run their own models, and the big cloud providers offer ready-made toolkits for it. Attention is shifting toward operating language models and AI agents, which has given rise to separate tools for LLMOps and agent observability.

Status: 2026 November

Also known as

MLOps · ML Ops · machine learning operations · machine learning operations · machine learning operations · MLOps · model operations · ModelOps

Sources

Google Cloud – MLOps: Continuous delivery and automation pipelines in machine learning

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.