The Refine row

After any report in the Report Library finishes loading, a Refine row marked with a sparkle icon may appear beneath the KPI tiles. It can hold two to four tappable filter chips — things like "Just Daniel Mir," "Compare to previous," or "Same period last year" — alongside short, italic observation notes such as "Sara Aziz rebooks noticeably more than the rest of the team."

The two things in that row look similar at a glance — both are AI-touched, both sit in the same strip — but they work quite differently under the hood, and it's worth understanding the difference, because it's the difference between "a shortcut to a real filter" and "a sentence of AI writing that's been safety-checked before you ever see it."

In this article
  • What happens, mechanically, when you tap a suggested filter chip
  • Why chip wording is never the AI's own invented text
  • How observation notes are checked before they're allowed to appear on screen
  • What you'll see while suggestions are loading, and when none appear at all
Refine row with AI-suggested filter chips and an observation note
A loaded report with the Refine row visible below the KPI tiles, showing three suggested filter chips and one italic observation note.

Filter chips: AI picks which one, never how to phrase it

Tapping a chip re-runs the actual report with that filter applied. It isn't a preview or a suggestion you then have to go set up yourself in the filter panel — it's the same result you'd get by configuring that exact filter manually, just arrived at in one tap instead of several.

The AI's role here is deliberately narrow: it looks at the loaded report and decides which one or two filters would plausibly be interesting to apply next, given what it's seeing. That's the entire decision it's trusted to make. The actual chip label and the short reasoning shown alongside it are not the AI's own sentence — they're rebuilt server-side from real, validated filter definitions and data, using the AI's pick only to choose which pre-defined chip to surface.

Chip labels are never the AI's own wording The AI's role is limited to choosing which filter to suggest. The label and reasoning shown on a chip are rebuilt server-side from real, validated data — the AI never writes or phrases the chip text itself.
1

Load a report

Open any report as usual from the Report Library.

2

Check the Refine row

If suggestions are available, chips appear below the KPI tiles, marked with a sparkle icon.

3

Tap a chip

The report re-runs with that filter applied — the same result as if you'd set the filter yourself.

Observation notes: scrubbed before they ever reach your screen

Observation notes are the italic sentences in the Refine row that call out a pattern worth noticing in the report — they're informational only, not clickable, and there's no filter behind them. This is closer to genuine free-form AI writing than the chips are, since there's no pre-defined library of possible sentences to pick from — which is exactly why it carries an extra safety check that the chips don't need.

Before an observation note is ever displayed, it passes through a validation step that specifically rejects the note if it contains anything that looks like a number, a percentage, or a currency symbol. If an observation fails that check, it's discarded rather than shown with the offending figure edited out — you either see an observation with no numeric content at all, or you don't see one.

A hallucinated figure can't reach the screen Before an observation note is shown, it's scrubbed and rejected if it contains anything that looks like a number, percentage, or currency symbol. This means an AI-invented figure is structurally prevented from appearing in an observation, not just discouraged by instruction.

This is the same design principle used elsewhere in this category — see AI in ClinyPal: Overview — applied to free-form text specifically: rather than trusting an instruction not to invent a number, the system makes it structurally impossible for a number-shaped AI sentence to be displayed at all.

Loading, and when nothing appears

While suggestions are being generated, the Refine row shows a "Looking for suggestions..." shimmer beneath the KPI tiles. Critically, this never blocks or delays the report itself — the report's tiles and table load and are fully usable before, during, and after the Refine row resolves. You're never waiting on the AI to see your actual data.

  • If suggestions load successfully, you'll see chips, an observation note, or both, depending on what the AI found worth surfacing.
  • If the AI call fails, isn't configured, or genuinely has nothing useful to suggest for that report, the Refine row simply doesn't appear at all — the report is unaffected either way.
Refine row loading shimmer beneath report KPI tiles
The Refine row's loading shimmer, shown briefly below a report's KPI tiles while suggested chips and observations are generated.

FAQ

Was this article helpful?
Your feedback helps us improve our documentation.
Rate this article: