Upgrade & Secure Your Future with DevOps, SRE, DevSecOps, MLOps!

We spend hours on Instagram and YouTube and waste money on coffee and fast food, but won’t spend 30 minutes a day learning skills to boost our careers.
Master in DevOps, SRE, DevSecOps & MLOps!

Learn from Guru Rajesh Kumar and double your salary in just one year.

Get Started Now!

Why Does the Model Keep Failing? A Five Whys FAQ for AI Ops

Why Does the Model Keep Failing? A Five Whys FAQ for AI Ops
Artificial Intelligence

A recommendation model that was fine last month starts returning odd results, an on-call engineer gets paged, and the quickest fix is to retrain it. Two weeks later the same page arrives. Teams running machine learning in production know this loop well. Repeated “why?” questions are the way out of it, and this FAQ explains how to apply them to AI operations without turning the exercise into a ritual. The examples are deliberately plain, and nothing here depends on a particular product.

What does “asking why five times” mean in AI operations?

It means treating the first explanation of a failure as a hypothesis. When a model degrades, the first explanation is usually about the model: it drifted, it is stale, it needs retraining. Each follow-up question moves one layer outward, from the model to the data feeding it, then to the pipeline producing that data, then to the checks around the pipeline, and finally to the working agreement that decided which checks exist. Five is a convenient stopping point, not a quota.

Can you show a worked example?

Take a ranking model whose click-through quality falls overnight.

  1. Why did quality fall? The model scored many items with an empty value in one input field.
  2. Why was the field empty? An upstream table stopped receiving a column after a schema change.
  3. Why did nobody notice for a day? The pipeline still ran green because empty values are valid values to the loader.
  4. Why is an empty column treated as valid? No expectation was ever written for how often that field may be blank.
  5. Why was no expectation written? The release checklist for pipeline changes covers code review and tests, and says nothing about input contracts.

The first answer would have led to a retrain, which would have trained on the same hollow column. The fifth answer leads to a one-line addition to the release checklist and a blank-rate expectation on the field. Those changes are small, cheap and permanent.

How is this different from a standard post-incident review?

A review often records a timeline, a severity and a list of actions. The five-whys chain is narrower: it is a single path from symptom to a fixable cause, and it can be run in ten minutes. Many teams use it inside a larger review as the section that decides which actions are worth the effort. The two work well together, but the chain is the part that stops the review from ending at “retrain the model”.

Where should the chain stop?

Stop when the answer is something your team owns and can change within a normal planning cycle. “The vendor’s service was slow” is a poor stopping point if you can add a timeout and a fallback; “we have no fallback for that call” is a good one. If an answer describes the weather, the economy, or a person’s mood, step back one question and rephrase it as a process gap.

What goes wrong most often?

  • Answers that restate the question. “Why did accuracy fall? Because the model got worse.” That is circular. Demand an observable fact for every answer.
  • Single-thread tunnel vision. A noisy incident can have two contributing causes. Run one chain per cause rather than blending them.
  • Stopping at a person. If the chain ends in “the engineer forgot”, ask what the routine would have to look like for forgetting to be harmless.
  • No follow-through. Record the single change, its owner and a date to confirm it held.

Does it work for non-technical problems around the model?

Yes, and it is often more valuable there. A labelling team that keeps returning inconsistent annotations might trace the issue from the labels, to the guidelines, to a guideline revision nobody announced, to the lack of a channel for announcements. A stakeholder group that keeps misreading a report might trace it to a chart default, then to the template, then to the absence of a reviewer. The mechanics are identical, and the fixes tend to be even cheaper than technical ones.

How do I explain the method to colleagues outside engineering?

Use a domestic analogy. If a kettle keeps tripping the fuse, you can reset the fuse each morning, or you can ask why: the socket is overloaded, because three appliances share it, because the kitchen has one outlet, because nobody planned the extension. Colleagues grasp that quickly, and it avoids jargon about “root cause analysis” that can make the exercise sound heavier than it is.

Is there a place for outside examples of organised information?

Occasionally, yes. When teams want to see how a large subject can be arranged for easy browsing, they sometimes look at unrelated sites. A medical-tourism style directory such as this site with separate hospital, doctor, blog, forum and procedure sections is a handy example of how clear categories keep a big catalogue navigable. The lesson for model operations is organisational: when everything has an obvious home, a missing item stands out sooner.

How does this connect to wider decision quality?

Better data about outcomes makes every chain easier. Our piece on how AI feeds everyday business decisions describes the kind of downstream consumers that notice trouble first, and it hints at why the reader of a weekly report frequently spots a silent pipeline fault before any monitor does. That observation alone justifies giving report readers an easy way to flag oddities. Similarly, anyone interested in the training side can read our overview of AI changing classrooms and learning tools, where data quality questions appear in a different setting.

What does a lightweight routine look like?

Keep it modest. After any page that repeats, spend ten minutes with the people involved and a blank document. Write the symptom at the top, then add one answer per line, each phrased as an observable fact. Stop at an answer your team owns. Underneath, write one change, one owner and one date. Review the document at the next planning session to see whether the repeat page disappeared. If it did not, the chain probably had a wrong link, and the exercise starts again from the new evidence. That loop, repeated over a quarter, usually does more for reliability than another round of tuning.

What should I take away?

Retraining, restarting and rerunning are all responses to symptoms. They are sometimes necessary, but they rarely end the series of pages. The questions in this FAQ give you a precise way to move from the visible fault to the working agreement that allowed it, and to change that agreement in a way the whole team can see. Start with the next incident that repeats: write the first answer, resist the urge to act on it, and ask why once more.

Subscribe
Notify of
guest
0 Comments
Oldest
Newest Most Voted
0
Would love your thoughts, please comment.x
()
x