Iterative AI Mode
Conversational analysis with human-in-the-loop approval
Summary
Iterative mode is the fast, back-and-forth workspace for exploring a question, refining it, and checking the AI’s work before you share it. It keeps short-term memory inside the thread, returns quick previews for tables/charts to keep latency and cost low, and lets you review & approve results before they appear in the Message window. Scalars (e.g., “Sales for 1H 2025 was X”) run against full data; preview tables/charts are lightweight. Once you’re happy, approve and (optionally) send the work to Data Studio/GenViz to build polished analytic content that runs on the full dataset.
How to get here (UI path)
Workspace → Channel → Iterative AI
Open your Workspace
Choose the Channel bound to the right Dataverse
Click the Iterative AI tab/mode
Ask your question to start a thread
Human-in-the-loop workflow (review → approve → share)
Ask a question and iterate: follow-ups, filters, drilldowns.
Review the answer in the thread and (where available) check the Building panel for logic/SQL transparency.
Approve to commit the result to the Message window (nothing shows to others until you approve).
After approval you can:
Publish as a Message to the channel
Pin important insights
Share a view-only link
Clone the Message for another audience
Open in Data Studio / GenViz to build cards, summary tables, and charts (full-data execution)
Data scope & performance model (preview vs full)
Scalars = Full data. Examples: totals, averages, single-value KPIs (“Sales for 1H 2025”, “Avg tariff rate last quarter”). These are computed on the full Dataverse so the number you quote is authoritative.
Tables & charts = Preview by default. To keep latency low and the cost of iteration minimal, tabular and visual results render from a sampled/subset preview while you’re exploring. This is ideal for trying different filters, time ranges, or group-bys rapidly.
Promote to full data when ready. Once the preview looks right and you approve, you can:
run the query on full data, or
hand off to Data Studio/GenViz, where the visualization/metric is materialized against the full dataset and saved as reusable analytic content.
Clear indicators. The UI differentiates Preview vs Full so you always know the data scope behind a result.
What Iterative is best for
Rapid exploration when you don’t know the exact question yet
Multi-turn refinement (e.g., “Explain margin dip → by region → last 6 months → call out anomalies”)
Quick “sanity-check” views before promoting to a full run
Multiple concurrent threads within a channel (rename, pin, revisit)
Prompting patterns that work
Start broad, narrow down: “Show landed cost by supplier over the last 12 months → add YoY → filter to >$250K purchases → group by category.”
Ground definitions: “Use landed cost = base + tariff + freight; treat Tier-1 as ‘priority suppliers’.”
Ask for preview sizes: “Return a preview table (10–20 rows) with Supplier, Orders, Landed Cost, Tariff %.”
Validate assumptions: “List the business rules you applied and show the SQL/logic you generated.”
Promote when ready: “Looks good—run on full data and summarize in 3 bullets.”
Quality checks before you hit Approve
Metric definition matches your business lexicon (e.g., landed cost formula)
Filters & time range are correct (currency/units too)
Grouping & sort reflect how you’ll present the result
Preview vs full—decide if it’s still exploration (keep preview) or ready to promote to full / open in Data Studio
When to move to Data Studio / GenViz
You want production-grade cards, summary tables, or charts
You need full data across heavy joins, longer time windows, or governed measures/dimensions
You plan to reuse the metric/visual in messages or other channels
Managing threads
One topic per thread to keep the short-term memory clean
Rename meaningful threads, pin important ones
Clone a Message to tailor for a different audience without losing the original context
Troubleshooting
Answer looks off → confirm you’re in the right Channel/Dataverse; restate key definitions; ask to show assumptions
No rows returned → relax filters or extend the date range; request a wider preview
Slow response → reduce scope (recent period, fewer dimensions) or stay in preview until ready for full data
Last updated