Dotwave Understanding AI risk warnings
Every operation the Dotwave AI proposes carries a risk assessment, and learning to read it turns the assistant from a black box into a transparent advisor. The risk badge is not there to block you — it is there to tell you how confident the AI is and why, so you can weigh its judgement against your own knowledge of the data. This article breaks down the three risk levels, the fixed structure behind every recommendation, and the specific warnings you will encounter most often.
The three risk levels
Dotwave sorts every proposal into one of three levels, each with a colour and a plain-language label:
- Green — "Looks good": the operation is statistically sound. The AI sees no issue and recommends proceeding.
- Amber — "Review carefully": the operation is valid, but a potential issue is flagged for your attention — for example, a high null percentage in the column being filled. It is not a stop sign; it is a "look before you leap."
- Red — "Not recommended": the AI does not recommend this operation. A typical trigger is a change that would drop more than 50% of your rows. The operation may still be possible, but the AI is advising against it.
The colour maps to how much scrutiny the AI thinks the change deserves before you apply it — green invites a quick confirm, amber asks for a second look, red asks you to reconsider.
The reasoning structure
Behind every proposal the AI builds its reasoning in a consistent order, which is what makes its recommendations easy to audit. Before it proposes, it states:
A restatement of your request, so it is clear what the AI is responding to.
The relevant facts from your profile — counts, percentages, distributions — that inform the decision.
What the AI advises you to do.
Why it advises that, tied back to the data evidence above.
A different approach you might take instead, when one applies.
This structure is why the "Why this recommendation" toggle is worth opening: it shows the evidence and logic in the same order every time, so you can check the AI's work rather than take its badge on faith.
Common warnings and what they mean
A handful of warnings account for most of what you will see, and each points to a concrete data condition:
- Skewed distribution → fill with median: when a numeric column is lopsided, the AI favours the median over the mean because the mean gets dragged by the long tail.
- Not recommended → drops more than 50% of rows: an operation that would delete over half your data earns a red badge, because losing that much is rarely what you intend.
- High cardinality → one-hot capped at 50: a categorical column with too many distinct values cannot be safely one-hot encoded, so Dotwave caps the expansion at 50 unique values.
- Division-by-zero warning: when a calculated or combined column divides by a field that contains zeros, the AI flags it before creating the column.
If the AI says "Do not recommend" and you disagree, type "Do it anyway." The AI will restate the risk once, then propose the operation with a High risk badge.
The risk system is advisory, not authoritarian. You know context the profile cannot capture — that a 60% row drop is exactly right because you are isolating a single segment, for instance. When your judgement is sound, override the warning; when you are unsure, open the reasoning and let the AI's evidence guide you. Either way, nothing changes until you apply it.