Upgrade & Secure Your Future with DevOps, SRE, DevSecOps, MLOps!
We spend hours scrolling social media and waste money on things we forget, but won’t spend 30 minutes a day earning certifications that can change our lives.
Master in DevOps, SRE, DevSecOps & MLOps by DevOps School!
Learn from Guru Rajesh Kumar and double your salary in just one year.
Every few years the AIOps conversation changes its favourite word. First it was correlation, then prediction, and lately it is autonomy. Underneath the vocabulary sits one stubborn habit that has not moved at all: teams keep waiting for the model that is finally accurate enough to trust, and while they wait, the dashboards stay noisy and the pages keep arriving. This state-of-AIOps piece is a plea for a quieter idea, which is that “good enough for the purpose” is a legitimate engineering target and “perfect” is not.
A small story from the night shift
Picture an operations group that spent most of a year tuning an anomaly detector for its payments platform. Every review meeting ended the same way. The detector caught the big incidents, but it also flagged a handful of harmless blips, and somebody always said the false positives had to be near zero before the tool could be switched on for the on-call rota. So the work continued, feature by feature, while the old threshold alarms kept ringing through the night.
Then a quiet incident changed the argument. A slow memory climb ran for days, the old alarms never tripped, and the half-finished detector, running only in a shadow mode nobody looked at, had marked it on day one. Nothing was wrong with the model that morning. What was wrong was the definition of done. (This is an invented scene, offered only to make the pattern easier to see.)
Where AIOps actually stands today
Look at the practical landscape and a few things are fairly settled. Detection of unusual behaviour in metrics is a mature technique, grouping related events into one incident works well when the telemetry is tidy, and summarising an incident timeline for a human is now routine. Root-cause suggestions work as a starting hypothesis, not a verdict, and automatic remediation is mostly limited to narrow actions such as restarting a stateless service.
That map shows where “good enough” is already reachable. A detector that is right most of the time and says why it fired beats silence. The trouble starts only when a team demands the one property no probabilistic system can offer, certainty, and treats every miss as proof the tool is not ready.
The moving target problem
Perfection is slippery in operations for a structural reason. Systems change weekly, traffic patterns drift with seasons and product launches, and a model tuned to last quarter’s behaviour starts decaying the moment it is finished. If your acceptance bar is “no mistakes”, you will never reach it, because the thing being measured keeps moving underneath you. If your bar is “better than what we have now, and cheap to correct”, you can reach it in a few weeks and then keep it healthy.
There is a second cost. The responder at three in the morning cannot use a model that lives in a notebook. A modest detector that is live and a little noisy teaches the team more in a month than a flawless one teaches in a year, because only the live one produces feedback.
How to write a good-enough definition
The practical move is to decide, before the build starts, what the tool is for and what level of correctness that purpose needs. Teams that do this well tend to write three short sentences and pin them where everyone can see them.
- The job: name the single decision the output supports, for instance “tell the on-call engineer which service to open first”.
- The acceptable miss: say what kind of error is tolerable and what kind is not, such as a harmless extra suggestion versus a silent failure on a customer-facing path.
- The stopping point: state the evidence that ends the tuning work, for example a few weeks of side-by-side use in which responders preferred the new output more often than not.
No target accuracy figure borrowed from a vendor slide appears there, on purpose. Such numbers only mean something against your own data and your own consequences, and once you know roughly what a wrong call costs, you can argue sensibly about how much more tuning is worth buying.
Keep the remaining polish optional
Stopping at good enough does not mean abandoning quality. It means ordering the work so that the optional parts are explicitly optional. Once the detector is live for its narrow job, improvements queue up behind real evidence, such as which alerts responders dismissed and which incidents slipped past. None of them blocks the release.
The write-up on how decisions in IT operations get help from machine learning makes the same underlying point from another side: tools earn their place by narrowing a choice, not by removing the human from it. And anyone who has read the piece on calming a flood of operational events with smarter tooling will recognise that the first win is rarely elegant, it is simply live.
An everyday parallel outside tech
The same logic shows up wherever people choose among many options. Take BestLawyerNow, a one-page legal directory grouped by area of work, with a journal and appointment buttons on view. Nobody comparing options there is aiming for the one flawless choice; the sensible aim is a decision that is sound enough to act on, followed by a conversation that fills in the rest. Operations teams face the same shape of problem with alerts and runbooks.
What to do on Monday
If your own AIOps effort is stuck in a tuning loop, here is a small sequence that tends to unstick it without any new budget.
- Pick one detector or one correlation rule that is already roughly working.
- Write the three sentences above, job, acceptable miss and stopping point, and get the on-call group to agree to them.
- Switch it on for a real, limited slice of traffic, with a visible label saying it is a suggestion.
- Review dismissals and misses after a few weeks, then declare the work done for now and move on.
Where this leaves the field
The honest state of AIOps is a set of capable but imperfect tools, plus a set of teams that differ mainly in how they handle imperfection. Some wait for certainty and ship nothing. Others define the purpose, accept a known and cheap kind of error, and learn quickly from real use. Over a year the second group ends up with better models, because they started sooner and kept the loop short.