AI Learning
Adaptive AI
Axon learns from how you interact with AI-generated content. As you edit, delete, and accept findings, the system quietly builds a picture of what good analysis looks like for your project — and applies that understanding to every future AI action. No configuration required.
Available on Pro and Enterprise plans.
How it works
Every time you edit, delete, or accept a finding, Axon records a feedback signal. These signals accumulate silently in the background and are used in three ways:
- Guiding new extractions. When processing a new artifact, the AI is shown examples of findings you have previously kept unedited or written yourself. These act as quality benchmarks, steering the AI toward the tone, depth, and focus that works for your project.
- Filtering out rejected content. Findings that are semantically similar to ones you have previously deleted are automatically suppressed before they reach your board. The system remembers what you did not want.
- Building a preference profile. After enough signals accumulate, Axon generates a short natural-language summary of your preferences for this project — for example, that you favour concise operational risks, or that you rarely keep broad strategic observations. This profile is injected into AI prompts to give every AI action a baseline understanding of what you value.
What triggers a signal
| Action | Signal type |
|---|---|
| Accept an AI finding without editing | Positive — used as a quality example |
| Accept an AI suggestion from the suggestion tool | Positive — used as a quality example |
| Edit an AI-generated finding | Mild negative — content was not quite right |
| Delete an AI-generated finding | Strong negative — suppresses similar content in future |
| Create a finding manually | Positive — used as a quality example |
Signals are scoped to each project
Learning is kept separate per project. A competitive analysis project and a product requirements project may call for very different styles of findings, and the system treats them independently. Feedback in one project never influences another.
When does it kick in
Few-shot examples are injected as soon as there are accepted or manually created findings in your project. The negative similarity filter activates once deleted findings have been recorded. The preference profile is not generated until at least 10 feedback events have accumulated, ensuring it is based on a meaningful pattern rather than a handful of actions.
The more you interact with your findings, the more accurate the system's understanding of your preferences becomes. Projects with a rich history of reviewed and curated findings will see the most noticeable improvement in AI output quality.
Tips for best results
- Delete liberally. Every finding you remove is a signal. Deleting weak or irrelevant findings is one of the fastest ways to improve future extractions.
- Write findings manually when the AI misses the mark. Manually created findings carry strong positive weight — the system treats them as the clearest possible example of what you want.
- Be consistent within a project. Mixed signals — keeping some findings but deleting similar ones — can reduce the clarity of the learned preference. The more consistent your curation behaviour, the sharper the profile.
- Give it time. The preference profile requires at least 10 events before it is generated. On a fresh project, the first few ingestions benefit only from few-shot examples. After more active use, all three learning layers work together.