MLOps is the set of practices for taking machine learning models from development into production and keeping them working there: managing training data, training and versioning models, deploying them, monitoring their performance and retraining when they degrade. It was built for predictive models such as fraud scores and demand forecasts, and many of its habits carry over to AI agents.
Why does MLOps matter?
Before MLOps, models were often built once by a data scientist, handed to engineers, and left running until someone noticed they had stopped working. MLOps brought software discipline to that process: reproducible training, version control for data and models, automated testing, controlled releases and monitoring for drift.
In regulated firms, MLOps also supports model risk management. Supervisors expect to see a model inventory, validation before use, ongoing monitoring and change control. US banks work to the interagency SR 11-7 guidance; MLOps is how those expectations become routine.
What does MLOps cover in practice?
- Data management: versioned training data with known lineage.
- Training pipelines: repeatable runs that produce the same model from the same data.
- Model registry: every model version, its metrics and its approval status.
- Deployment: packaging and releasing models, with rollback.
- Monitoring: prediction quality, input drift and performance over time.
- Retraining: a controlled route to update the model when performance falls.
What does an example look like?
A US telecoms operator uses a churn model to predict which customers are likely to leave. The MLOps team retrains it monthly on recent data, compares the new version with the live one on a holdout set, and promotes it only if it performs at least as well. Monitoring shows when input patterns shift, such as after a price change. The model produces a score; a retention team decides what offer to make. Operations of that kind are set out in AI agents in US telecoms operations.
How is MLOps different from LLMOps and AgentOps?
MLOps assumes you own and train the model. LLMOps usually works with a model someone else trained, so it focuses on prompts, retrieval, evaluation and provider changes rather than training pipelines. AgentOps goes further: it governs what an agent does in the business, its bounds, its held actions and its authority.
The biggest difference is what the system produces. A predictive model outputs a number and a person or rule acts on it. An agent acts. MLOps monitoring asks "is the score still accurate?" Agent governance must also ask "should this agent have taken this action, and who allowed it?"
What carries over to agents, and what does not?
Carries over: versioning everything, testing before release, monitoring for drift, controlled rollback, and a registry of what is running.
Does not carry over: the assumption that you fix problems by retraining. For agents built on a provider's model, the fix is usually a change to the brief, sources or bounds, followed by recertification. We never update model weights for a customer, for exactly this reason: qualification, not training. In e-commerce, for instance, an agent handling returns is corrected by changing its returns policy source and its refund limits, not by retraining; the setting is described in AI agents in US e-commerce operations.
Which terms sit close to it?
Model risk management: the governance MLOps supports. Model drift and data drift: what MLOps monitors. Feature store: a shared store of model inputs. See our guide to deploying AI agents for how these practices apply once agents act.
Questions people ask
Do you need MLOps to run AI agents?
You need its habits: versioning, testing, monitoring and controlled change. You may not need its training infrastructure if your agents use a provider's model without fine-tuning.
Is an AI agent a model under model risk rules?
Supervisors increasingly treat AI systems, including agents, within model risk frameworks. Expect to keep an inventory, validate before use and monitor ongoing performance, whatever the system is called.