A new label has become established around the operation of software: "AI SRE". This refers to tools that detect incidents, narrow down causes and, in the best case, take countermeasures themselves. In a recent article, Severin Neumann criticises precisely this buzzword and instead argues for a model from the Cloud Native Computing Foundation (CNCF) – the foundation that is home to projects such as Kubernetes and Prometheus. The core of his argument: reliability tasks should be automated, not people.
What lies behind the abbreviations
SRE stands for Site Reliability Engineering, the discipline that safeguards the reliable operation of applications using engineering methods: measurable target values for availability, automated responses to recurring errors, clearly regulated on-call duties and a post-incident review that focuses on causes rather than culprits.
"AI SRE" promises to fill this role with AI systems. And that is exactly where the criticism begins: a label on a product says nothing about which specific task it takes over and how reliably it does so.
Why the distinction is practically relevant
For operators of websites, portals and business applications, the difference is not a play on words but a question of procurement. Anyone who buys a role buys a promise. Anyone who automates a task can check whether the result is correct.
Specific tasks can be named and assessed:
- Consolidating alerts instead of sending thirty individual notifications during an outage
- automatically compiling the relevant logs, metrics and most recent deployments during an incident
- triggering standard measures such as rolling back a version or restarting a service
- converting recurring routine operational tasks into executable workflows
For each of these tasks, one can ask: How often does it occur? How much time does it cost today? And how can one tell that the automation acted correctly? These questions are considerably harder to answer for a product that simply promises "AI for operations".
AI is an ingredient, not a foundation
Language models are good at condensing unstructured information – for example, summarising long log outputs into a readable summary, or deriving a suggestion for the next check from an incident history. At three o'clock in the morning on call, that is a real gain.
What language models do not bring: a reliable data foundation. If an application does not deliver clean metrics, has no traceable deployment history and configurations are not versioned, even a model can only guess. The effort then lies not in the AI but in the groundwork: observability, i.e. the ability to assess the state of a system from the outside using metrics, logs and traces, is the prerequisite – not the by-product.
Added to this is the question of authority to act. A system that makes suggestions is uncritical. A system that intervenes autonomously in production needs the same safeguards as a human: limited rights, traceable logs, defined rollback paths.
What this means for evaluating offerings
When a provider offers AI-supported operations, a few sober follow-up questions are worthwhile:
- Which task exactly is automated – and what happens with all the others?
- Which data does the system need, and is that data even available in sufficient quality in your own operations?
- Is the system allowed to act or only to make suggestions? If it is allowed to act: to what extent, and how is this logged?
- What does the fallback mode look like if the automation fails or makes the wrong decision?
- Who bears the responsibility for a wrong decision – contractually and organisationally?
Answers to these questions say more about the benefit of a tool than any abbreviation in the product name.
The unspectacular part remains the most important
Reliable operation rarely arises from a single tool. It arises from repeatable deployments, from monitoring that measures the right things, from documented procedures for incidents and from the willingness to genuinely follow up after an incident. This applies just as much to a TYPO3 portal with connected business systems as to a distributed microservice architecture.
These foundations are not very glamorous and are hard to put on a slide. But they are the basis on which automation – with or without AI – can take effect in the first place. A model that addresses reliability tasks instead of replacing roles makes this connection explicit: humans remain responsible, the machine takes over the repetition.
For web and software projects, this means: before deciding on AI-supported operations tools, it should be clarified which operational tasks are currently performed manually and recurrently and whether the necessary data is available in clean form. Anyone who does this groundwork can deploy AI in a targeted way where it actually saves time – and at the same time retains operations that continue to work even when the tool gets it wrong.
Sources
- AI für SRE: Warum Automatisierung das Ziel ist und KI nur ein Werkzeug — heise developer News