TL;DR
Planning Mode turns broad, open-ended product analytics questions into an explicit, editable plan you approve before the Analytics Agent starts working. It is built for diagnostic deep dives — retention drops, conversion regressions, onboarding investigations — where wrong assumptions waste the most time. Users trigger it with the /plan slash command or natural-language phrasing like 'how should we investigate…' inside the Analytics Agent.
Planning Mode is a new capability inside Mitzu's Analytics Agent. It converts a broad, open-ended product analytics question into an explicit, editable plan that you review and approve before any query runs.
It is built for the question types where most analytics work actually happens — diagnostic deep dives, root cause investigations, impact analysis. Questions like "why is week-2 retention dropping for paid users" don't have a single right answer; they have a right approach. Planning Mode makes that approach visible before the agent commits to it, so the wrong assumption gets caught in seconds rather than after a finished analysis.
For teams running event data in Snowflake, BigQuery, Databricks, Redshift, or ClickHouse, Planning Mode adds a deliberate approval step on top of the agent's existing workflow — for the moments where deliberation is worth a few extra seconds.
1. What is Planning Mode?
Planning Mode is a mode of the Analytics Agent that splits a complex question into two phases: plan, then execute. In the plan phase, the agent examines existing insights, dashboards, cohorts, and the semantic layer, then proposes a sequence of 3–7 analytical steps in plain language. No SQL runs. No charts render.
You see the agent's intended approach, and you approve, edit, or rewrite it.
Once approved, the plan executes as a live progress widget — each step ticks off as the agent works through it — and the run concludes with a synthesised report and recommended next actions.
Underneath the plan, the same architecture applies that powers the rest of the Analytics Agent: the agent assembles analysis specifications (funnels, retention, segmentations, journeys), and a deterministic query engine turns those specifications into SQL. The plan you approve is what runs.
2. Why a plan-first mode matters for product analytics
Most descriptive product analytics questions are well-defined: how many sign-ups last week, weekly active users by country, conversion rate on the checkout funnel. The agent can run those directly because there is one obvious right interpretation.
Diagnostic questions are different. "Why did retention drop last month" can mean a dozen different investigations: which retention curve, which cohort, which breakdown, which time window, which comparison baseline. Pick a wrong interpretation and the answer is technically correct and analytically useless — you investigated the wrong thing.
Planning Mode addresses that gap. By forcing the interpretation to be explicit before any query runs, it removes the most common failure mode of AI analytics: confidently producing a polished answer to a question the user did not actually ask.
3. When to use Planning Mode (and when to skip it)
| Question shape | Example | Planning Mode? |
|---|---|---|
| Descriptive — single metric, well-defined | How many signups did we have last week? | Skip. Ask the agent directly. |
| Descriptive — segmented, well-defined | Show me weekly active users by acquisition channel. | Skip. Ask the agent directly. |
| Diagnostic — root cause | Why is week-2 retention dropping for paid users? | Use Planning Mode. |
| Diagnostic — impact | Did the new pricing page move trial-to-paid conversion? | Use Planning Mode. |
| Diagnostic — deep dive | What's behind the drop in onboarding completion last month? | Use Planning Mode. |
| Hypothesis validation | Do users who hit feature X retain better than those who don't? | Use Planning Mode — methodology choices matter. |
| Ambiguous scope | How are our enterprise accounts behaving? | Use Planning Mode to make scope explicit. |
Rule of thumb: if the answer depends on choices the agent will have to make for you — which cohort, which window, which comparison — Planning Mode pays for the extra approval step. If the question already specifies those choices, skip it.
4. How to trigger Planning Mode
There are two ways to enter Planning Mode inside Mitzu's Analytics Agent:
- Slash command. Type
/planfollowed by the question. Example:/plan why is retention dropping for our paid users? - Natural-language phrasing. Wording like "how should we investigate…", "what's the best way to look at…", or "can you put together a plan for…" triggers Planning Mode automatically.
Planning Mode is available in the in-app Analytics Agent. Full reference and screenshots are in the Planning Mode documentation.
5. The Planning Mode workflow, step by step
Step 1 — Discovery
Before proposing anything, the agent examines what's already in the workspace: existing insights, dashboards, cohorts, and the semantic layer. This grounds the plan in real data and previous analyses, so the proposed steps reuse what the team has already established.
Step 2 — Plan proposal
The agent returns a plan as 3–7 plain-language steps. No SQL runs, no charts render. Each step describes what the agent intends to look at and why — for example, "Check retention by acquisition channel to isolate whether the drop is channel-specific" or "Compare cohorts before and after the v2.3 release to test the deploy hypothesis."
Step 3 — Review and iterate
Plans are conversational artifacts. You can approve as-is, ask the agent to drop a step that is irrelevant, narrow the scope to a specific dimension, add an analytical breakdown the agent didn't think of, reorder the steps, or request a complete rewrite. The agent responds with an updated plan, and the cycle repeats until the approach matches the question.
Step 4 — Execution
Once the plan is approved, the agent works through the steps in order. Progress shows as a live todo widget — each step ticks off as the analysis completes. Charts, tables, and intermediate findings render inline as the agent goes.
Step 5 — Findings and next actions
When the run finishes, the agent synthesises a report: what it found, which steps led to which conclusions, and recommended next actions. From there you can save insights, define a cohort from a finding, share the analysis, or branch into a follow-up investigation.
6. How to modify a plan conversationally
You don't edit plans by clicking around — you talk to the agent. The plan is a structured object, but the interface is conversational. Common modifications:
- Drop unnecessary steps — "Skip the geography breakdown, we already know it's not regional."
- Narrow scope — "Focus only on paid users who signed up in the last 30 days."
- Add analytical breakdowns — "Also break the retention curve down by plan tier."
- Reorder steps — "Start with the impact analysis, then the cohort breakdown."
- Request a complete rewrite — "Throw this away. Start from a different angle: I want to know whether onboarding changed, not whether retention changed."
7. Why showing the plan changes the trust profile of AI analytics
Most AI analytics tools couple a large language model with a query engine that runs whatever the model produces. The user sees the final chart, not the reasoning. If the methodology is wrong — wrong cohort, wrong window, wrong join — the chart still looks confident.
Planning Mode breaks that pattern in two ways. First, the agent shows the approach before running it, so methodology choices are reviewable up front. Second, the actual SQL is generated by Mitzu's deterministic query engine from the agent's analysis specifications — the same engine that has powered Mitzu's UI for years. The plan you approve is the analysis specification the engine executes.
There's no second translation layer where a hallucination can sneak in.
For data teams accountable to PMs, marketing, and leadership, that combination matters: the approach is reviewable in plain language, and the SQL underneath is reviewable as a verification artifact, not as the agent's authored work. See how Mitzu generates verified, trustworthy SQL for the architecture in more depth.
8. Concrete use cases for Planning Mode
Retention drop investigation
A PM notices week-2 retention dipped this month. "/plan why did week-2 retention drop in November" prompts the agent to propose a sequence: check cohort sizes to rule out noise, segment by acquisition channel, segment by plan tier, compare to the November release timeline, check onboarding funnel completion in the same window. The PM removes the geography step, approves the rest, and gets a synthesised finding in one run.
Post-launch impact analysis
A growth lead wants to know whether the new pricing page moved trial-to-paid. "/plan did the new pricing page launched on May 12 move trial-to-paid conversion" returns a plan that defines the cohorts, sets the conversion window, picks the comparison baseline, and proposes which segments to break down. The growth lead narrows the analysis to one geo and approves.
Onboarding funnel deep dive
An analyst is asked to figure out which onboarding step has the biggest drop-off and why. Planning Mode proposes a multi-step investigation: build the funnel, identify the worst-performing step, break it down by device and acquisition source, compare against an earlier baseline, and surface candidate cohorts to investigate further.
Feature usage and engagement impact
A product lead wonders whether feature X actually drives long-term engagement. Planning Mode proposes: define the feature-user and non-feature-user cohorts, control for tenure, compare retention curves, segment by plan tier, and flag any selection bias the analyst should review before reading the result.
9. Who Planning Mode is for
Planning Mode is designed for the same audiences as the rest of the Analytics Agent, but it earns its keep most for question shapes where the user does not write SQL and cannot independently catch a methodology error after the fact.
- Data analysts and analytics engineers — use it to scope deep-dive investigations without committing to a wrong approach, and to give stakeholders a readable record of how a finding was reached.
- Product managers — use it for retention, activation, and feature impact questions where the scope of the investigation needs to be agreed before the work starts.
- Growth and marketing leads — use it for campaign impact, channel mix, and lifecycle analyses where the comparison baseline drives the conclusion.
- Founders and operators — use it to ask broad strategic questions and review the agent's interpretation before the analysis lands in a board update.
10. Warehouse readiness — the prerequisite
Planning Mode runs inside the Analytics Agent, which means the same prerequisites apply: your event data already lives in a modern cloud warehouse — Snowflake, BigQuery, Databricks, Redshift, or ClickHouse — and Mitzu's Configuration Agent has indexed it into the semantic layer. If you're still capturing events into a third-party vendor silo without warehouse export, Mitzu (and therefore Planning Mode) is not the right fit yet.
Frequently asked questions
What is Planning Mode in Mitzu?
Planning Mode is a capability in Mitzu's Analytics Agent that converts a complex product analytics question into an explicit, editable plan. You review and approve the plan before any query runs, so you can catch wrong assumptions before the agent commits time to a misframed analysis.
How do I trigger Planning Mode?
Type the slash command /plan followed by your question, or phrase the question as "how should we investigate…" or "what's the best way to look at…". Both trigger Planning Mode inside the Analytics Agent.
When should I use Planning Mode versus asking directly?
Use Planning Mode for diagnostic, ambiguous, or multi-step questions — root cause analysis, impact analysis, deep dives, hypothesis validation. Skip it for descriptive metric requests where the interpretation is obvious, like "how many signups did we have last week."
Can I edit the plan after the agent proposes it?
Yes. Plans are conversational. You can drop steps, narrow scope, add breakdowns, reorder steps, or ask for a complete rewrite. The agent regenerates the plan based on your feedback, and the cycle repeats until you approve.
Does the agent write SQL during Planning Mode?
No. The agent never writes SQL — not in Planning Mode and not in regular mode. It assembles analysis specifications (funnel definitions, retention parameters, segmentation breakdowns) and Mitzu's deterministic query engine turns those specifications into SQL. The plan you see is the specification the engine will execute.
How is this different from a chain-of-thought response?
Chain-of-thought is the model showing its reasoning as it produces an answer. Planning Mode is structurally different: the agent stops before any analysis runs and waits for human approval. You can edit the plan before execution, not after. The plan is also grounded in real workspace context — existing insights, dashboards, cohorts, and the semantic layer — not just the prompt.
Does Planning Mode work in Slack or only in the app?
Planning Mode is available today in the in-app Analytics Agent. For surface-by-surface availability, see the Planning Mode documentation and the Mitzu product page.
Is Planning Mode part of the standard Mitzu plan?
Planning Mode is included in the Analytics Agent and available to all workspaces with the agent enabled. See Mitzu pricing for plan details.
Related reading
- Why agentic analytics needs a product-analytics-shaped semantic layer
- How Mitzu generates verified, trustworthy SQL
- AI analytics hallucinations and SQL transparency
- What is agentic analytics?
- Warehouse-native analytics: benefits and how it works
