Ask three machine-learning teams how they revisit their goals and you will get three different answers. One team says, “We look at them when the quarter ends.” Another says, “When something breaks, obviously.” The third admits, “Honestly, nobody has opened that planning doc since we kicked off.” All three are common, and all three leave money and attention on the table. In this piece I want to compare these habits side by side with a fourth option that is cheaper than it sounds: a monthly sit-down where the team rereads its written AI and AIOps goals and asks whether each one still fits.
Nothing exotic is needed. You need the goals in writing, a calendar slot and an honest conversation. Let me show you how the options stack up, using a made-up but realistic example so the trade-offs feel concrete.
The running example
Picture a small platform team that supports an anomaly-detection model for server metrics. At the start of the year they wrote down four goals. First, cut false alarms from the model so on-call engineers trust it. Second, retrain on fresh data more often. Third, label a backlog of past incidents. Fourth, try a larger model architecture to catch subtler patterns.
Three months later, things look different. The false-alarm problem was largely tamed by a threshold change. A data-pipeline migration made retraining awkward. The labelling backlog grew because nobody owned it. And the larger architecture turned out to cost far more to serve than anyone expected. Which of those goals deserves effort now? That is the question each habit answers differently.
Option one: the quarter-end review
Quarterly reviews are familiar and they come with a built-in audience, which is good for accountability. The downside is distance. By the time the quarter ends, our example team has spent up to twelve weeks pushing on the larger-architecture goal even though its serving cost made it unattractive in week four. The review is thorough, but it finds the problem late.
Quarterly reviews also tend to be about reporting upward, not about deciding. People prepare slides, defend progress and avoid saying “this goal is dead.” If you rely on them alone, goals tend to survive a lot longer than they deserve.
Option two: fix it when it breaks
Reactive revisiting has one real virtue: it responds to genuine pain. When a pipeline breaks or a model drifts, the team talks about priorities with energy. The cost is that you only ever discuss what is on fire. Slow-moving goals, like the labelling backlog, never become urgent enough to get a hearing, so they rot quietly until someone notices a gap months later.
There is a subtler problem too. After an incident, goals get rewritten in a hurry, often in the shape of whatever just went wrong. That produces a list shaped by the last outage and not by what the team actually wants to build.
Option three: set and forget
This is the cheapest in effort and the most expensive in results. The goals stay on the wiki, the roadmap stays aspirational and people quietly work on whatever feels sensible. Strong engineers can make this work for a while, but nobody knows which goals are real. New hires read the doc and chase aims the team abandoned long ago.
Option four: the monthly reread
Here the team blocks thirty minutes once a month. Everyone reads the goal list, and for each line asks a small set of questions. Is the reason we wrote this still true? Is the cost or difficulty what we expected? Has something else made it more or less important? The answer lands in one of three buckets: leave it, reshape it, or retire it.
Apply that to our example. The false-alarm goal is mostly done, so it moves to a lighter maintenance note. Retraining stays but is reshaped: instead of “retrain more often,” it becomes “make retraining possible after the migration.” The labelling backlog gets an owner or is explicitly cut down to the incidents that matter. The larger architecture is retired for now, with a note about what would change the decision. Within half an hour, the team has stopped paying for a goal that no longer makes sense.
A side-by-side view
- Speed of catching drift: the monthly reread finds it within weeks, the quarter-end within months, the reactive habit only after trouble, and set-and-forget may never find it.
- Effort per round: the monthly reread is the lightest per session; the quarterly review is the heaviest because of the slides.
- Quality of decisions: reactive reviews skew toward the latest incident, quarterly ones toward what looks good in a report, and the monthly reread toward whether the goal still fits.
- Emotional cost: retiring a goal feels smaller when you do it often and early, because less sunk effort is attached.
It does not have to replace the other rhythms. Many teams keep a quarterly review for the big picture and use the monthly reread as the quiet preparation that makes the quarterly one shorter.
What an AI or AIOps team should actually reread
The goals worth rereading are the ones that can quietly drift. In AI and AIOps work, that usually means a few categories.
- Data goals. Collection, labelling and cleaning aims depend on sources that change. A feed may be retired, or a schema may shift under your feet.
- Model goals. A target such as “better recall” may have been right for last quarter’s incident mix and wrong for this one’s.
- Operations goals. Alert routing, retraining schedules and cost ceilings depend on your infrastructure, which evolves.
- People goals. Skills the team wanted to build, like learning a new framework, can lose relevance if the stack changes.
If you are looking for general context on how AI reaches into everyday business decisions, there is a readable overview in a piece about AI and better business choices on this site, and the questions in it pair well with a goal review.
Running the session
Keep it boring on purpose. Share the document ahead of time. Have one person read each goal aloud and one person take notes. Give each goal at most two minutes, and park any deep debate on a separate thread. End by writing down the decisions in a sentence each, then circulate them. The aim is to leave the room with a shorter, truer list.
A tip from teams who do this well: write the reason next to each goal when you first create it. A line like “we want this because on-call people ignore noisy alerts” makes the reread far easier, because you are checking whether the reason still holds rather than whether the goal feels exciting.
One option for learning beyond your own field
Team goals often include learning something outside the daily stack. If you are curious how learners in other professions organise coursework, an online learning platform for aviation professionals is a website you could browse for a look at how one niche profession approaches continuing study; it presents itself as an independent platform. Whatever you read, the monthly reread applies equally to personal study goals.
When the comparison stops mattering
It is tempting to treat these four options as competitors, but the real point is that goal lists decay and some regular rhythm must fight the decay. A team that runs the monthly reread and skips everything else still beats a team with an elaborate quarterly ritual and no honest conversation. Conversely, a team that holds long reviews but never writes anything down will repeat the same debates forever.
If your team is curious about the practical skills side of this, the roadmap of skills for a new DevOps engineer is a decent companion for the people goals in your list, since skill aims are some of the easiest to let slide.
So pick a date, bring the list and ask the plain question on every line: does this still fit? Your models will keep changing underneath you, and your goals should be allowed to change with them.