Lantide Data

← Back to Learning Center

Why Analysis Conclusions Can't Stay in Chat

Read time: ~5 minutes · Series: Analyst getting started · Next: Quick vs Project

Monday's follow-up, Friday's dead end

On Friday you asked ChatGPT about funnel conversion; the thread ends with "overall conversion 12.3%." Monday the PM asks: "Is the denominator placed orders or created orders? Do cancellations count?"

You scroll the history—

  • The summary says "abnormal orders excluded" but not the WHERE clause
  • Screenshots show result tables, not whether JOIN caused fan-out
  • No document says "I signed off on this metric definition"

The issue usually isn't a dumb model—analysis never became reviewable artifacts: metrics in bubbles, logic in your head, delivery in Slack.

Three things chat lacks

Missing Why it hurts How Lantide helps
Contract Denominator, grain, time window unclear Plan states goals, stages, checkpoints
Evidence Can't rerun or diff SQL tabs keep pull logic; cache View SQL
Delivery Weekly meetings cite summaries only Report in project with numbers, limitations, HTML export

Lantide doesn't ban conversation—the Agent is in the right panel. The difference: formal conclusions and analysis contracts live on Plan / SQL / Report; chat is collaboration, not the only archive.

Why bother?
Without AI, those "contracts" lived in your head and query history. When AI handles most grunt work, separating contracts from chat becomes essential for traceability and revision.

Why annotations bind to documents

Replying in Slack "change denominator to paid users" often means:

  • Spoken change, Plan unchanged
  • Section A updated, section B still old metrics
  • Three months later nobody remembers "which version we meant"

Lantide annotations select source text and bind to Markdown. After Resolve the Agent patches only cited sections, not full rewrite. Consensus follows text, not memory.

You can revise Plan back and forth in Planning; before you press Execute, the Agent won't start formal pulls—the line between "discussed" and "approved."

Quick exploration OK; formal delivery needs upgrade

Not every question needs a project. Column shapes, try a WHERE, validate a join key → Quick Analysis is enough: Agent creates SQL tabs, runs results, interprets in chat.

But when analysis must:

  • Be signed off by leadership or finance
  • Survive cross-week metric challenges
  • Let colleagues take over and rerun

Focus a project and run Plan → annotate → Execute → Report. See Quick vs Project.

Self-check

  • If this number is challenged in six months, can I point to which SQL, Plan, and Report?
  • If PM changes metrics, does the change live in a document diff or drown in chat?
  • Am I exploring or doing deliverable analysis? The latter belongs in project mode.

Next steps

← Previous: Analyst introduction: what Lantide Data means for youNext: Quick Analysis vs Project Analysis →