How to use Learning Engine after experiments

Learning engine from experiments only works when finished tests leave a clear result behind. The Learning Engine turns completed work into learning events, then proposes source-linked recommendations from similar past actions. It never invents unsourced claims. Think of it as institutional memory with receipts. Chat opinions do not enter the graph unless they sat on a finished record somewhere else in Nautis. Empty Recommendations usually mean experiments, decisions, and deals are finishing without recorded results. The module is quiet at first on purpose. Pattern matches need your history, not a library borrowed from other startups.

Why learning engine from experiments stays quiet at first

Open Strategy & Planning, then Learning Engine. You do not create events on that page. Events arrive when you finish work elsewhere with a real result: a GTM experiment with a result, a Decision Loop outcome, a closed Investor CRM deal, a hire or reject in Jobs, or a completed Outcome Workflow.

Illustrative empty state (composite only): Recommendations stay blank until the first GTM experiment is marked with a result and the engine can compare similar past actions.

Each learning event carries the action, Expected vs Actual, Metric, a distilled Lesson, a suggested Next step, and a View source link. When Expected and Actual diverge, write the gap in plain words inside the source experiment or decision review. The Lesson is sharper when the source already admits surprise. Browse Learning Events for the raw record. Browse Recommendations for proposals waiting on Accept or Reject. Stat chips show Pending, Events, and Worked. Click any row for detail and use refresh when you just closed an experiment elsewhere and the list looks stale.

Close GTM experiments with a result worth remembering

In GTM Planner, open a workspace Experiments tab. Experiments are growth or messaging tests with a hypothesis, Channel, Phase, Metric, and optional budget. Status moves through Draft, Testing, Won, Lost, and Learnings.

Illustrative close (composite only): move Testing to Won, write the learning in the details dialog, confirm the Metrics tab has a matching snapshot, then open Learning Engine after the event appears. Channel and Phase tags help later filters. A Lost paid-social test in Launch should not silently advise a Pre-launch content experiment without you noticing the mismatch in Sources.

AI Suggest can propose experiments from strategy and uses AI credits. It does not replace writing the result when the test ends. Shared Access on a GTM workspace controls who can close experiments inside that plan. Name who owns experiment close-outs in the GTM Shared Access group. Unowned Testing cards are the most common reason learning engine from experiments looks broken. People missing from Shared Access cannot finish tests you expect to feed the engine, so Recommendations stall for social reasons, not model reasons.

Draft status is for design. Testing means the bet is live. Won, Lost, and Learnings are finished states that can feed the engine. Leaving cards in Testing for months starves both GTM metrics and Learning Engine.

Read Learning Events before you trust a pattern

Open Learning Events. Filter by Action type such as Campaign, Pricing change, Investor outreach, Product launch, Hiring, Runway action, Experiment, or Decision. Filter by Outcome: Win, Loss, Mixed, or Inconclusive. Start most weeks on Experiment plus Win and Loss. Add Mixed when the Metric moved but the business still felt wrong. Use Inconclusive sparingly so it does not become a junk drawer.

Click an event for the lesson and Similar past actions with a % match score, their outcomes, lessons, and source links. Reject mental math that ignores weak matches. Sort your attention toward Experiment action types when you are studying growth bets. Mixed Hiring and Experiment sources in one recommendation deserve extra skepticism.

Illustrative read (composite only): a new pricing experiment shows three similar past actions at high match that underperformed when discounting started too early; Next step suggests holding list price through week two.

If AI polishing is off or credits ran out, events still arrive with raw details. You may lack AI-polished lessons, but the source links remain the point. Lessons use a short background AI call and a small amount of credits when that toggle is on. Turn it off and you still keep Expected vs Actual and the View source path.

Treat Recommendations as advice with receipts

When patterns are strong enough, open Recommendations. Each item shows a summary, rationale, and a Sources box pointing at exact records. Accept confirms you may later record whether following it worked. Reject marks it rejected. Nothing executes automatically either way. Teams sometimes fear Accept because they think it launches work. It does not. It only commits you to a later honesty check via Record realized outcome.

Illustrative judgment (composite only): Accept a recommendation backed by four Experiment sources with solid match scores; Reject one whose sources are mostly unrelated Hiring events.

Recommendation statuses include Proposed, Accepted, Rejected, and Superseded. Later, use Record realized outcome on an accepted or rejected item, add an optional note, and choose It worked or Didn't work. That answer becomes a fresh learning signal.

Check Sources every time. If the receipts do not convince you, Reject. Superseded items mean a newer proposal replaced an older one. Do not chase every Superseded row; focus on Proposed and on Accepted items that still lack a realized outcome. When you Reject, write a private note in your ops doc about why if the Sources were weak. Future you will forget, and the engine cannot store a reject reason beyond the status.

Keep Decision Loop and Outcome Workflows honest as feeds

Decision Loop reviews and completed Outcome Workflows also create learning events. When you record Better than expected, As expected, or Worse than expected on a decision, or you complete an outcome workflow, those finishes can show up here. Do not skip the review step and then blame Learning Engine for silence. Outcome Workflow Completed means steps finished setup work. Still record whether the Goal's Target landed if you want that finish to teach the engine something real. Builders who only generate plans in Outcome Workflows without finishing Tasks create Completed workflows that teach the wrong lesson. Finish the board work that the Goal named, then let the event land.

Illustrative habit (composite only): Monday clear Decision Loop Review due rows; Friday close any GTM experiment that left Testing; then open Learning Engine Recommendations.

Hiring outcomes and closed investor deals feed the same engine. Mention them only as sources of events; this post stays on experiment-driven learning. If a recommendation cites mostly Investor outreach while you asked about onboarding experiments, Reject and wait for better Experiment density.

Separate institutional memory from wishful strategy docs

Strategy tabs and Roadmap plans describe intent. Learning Engine stores what happened. Keep both. Do not paste Lessons into Strategy as if they were positioning statements without checking Sources.

Illustrative split (composite only): Strategy still says "usage-based pricing"; Learning Events show two Lost pricing experiments on usage-based packaging; Recommendations urge a pause until a new Metric definition exists.

When a recommendation conflicts with a favorite narrative, read the % match panel aloud with a cofounder. The point of learning engine from experiments is to make disagreement about evidence, not about volume.

Run a light weekly pass so the module stays awake

Once a week, open Needs-shaped work in this order: finish any Testing experiment that already has numbers, clear Pending recommendations, and Record realized outcome on older Accepted items.

Illustrative weekly pass (composite only): close one Lost channel test with a written learning; Accept one Experiment-sourced recommendation; mark last month's Accepted pricing tip as Didn't work with a note. Put the pass on the same calendar as metrics hygiene. A Metrics tab without recent snapshots makes Expected vs Actual theater.

Matching uses only your organization's records. Other companies never train your engine. Investor CRM wins and passes matter, yet a company that never closes GTM experiments will still see a thin Experiment filter. Balance the diet of sources. Early quiet weeks are normal until you have enough finished experiments with results. Aim for a steady drip: at least one closed experiment or reviewed decision most weeks beats a quarterly dump of vague Learnings statuses with no Metric.

How to use Learning Engine after experiments is mostly discipline at close-out: a status, a metric, a written learning, then an honest Accept or Reject when Recommendations appear. How to use Learning Engine after experiments stays useful only while close-outs stay boring and complete. Fancy dashboards without Won, Lost, or Learnings statuses are decoration. Close the test, then open Learning Engine the same day when you can. Close the loop the same week so Pending recommendations do not age into folklore.