How AI Reports Are Created — Background
What Happens When You Generate a Report
When you click Generate, the following happens: Picasi collects all updates from the currently visible portion of your inbox—filtered by the active folder and any filters you’ve set. These updates, along with your AI context (if available), are passed to an AI model.
The model is given a clear task: to summarize, analyze, create a list, write LinkedIn posts, or identify conversation starters. In addition, you can provide an optional focus text to steer the output in a specific direction.
The AI processes all of this in the background—hence the short wait time of 15 seconds to two minutes—and returns a finished text, which Picasi saves as a report.
What Affects Quality
Three factors significantly determine how useful a report will be:
The Update Base: A report based on three updates will inevitably be more superficial than one based on 50. Too little input means too little output. There’s no minimum number, but for an analysis, it makes sense to collect at least one week’s worth of activity.
The AI context: Without your own context documents, the AI doesn’t know who you are. The analysis then remains general and describes what the competitor is doing—without explaining what that means for you. With a well-filled-out AI context, analyses become concrete and actionable.
The Focus Text: Without a focus, the AI generates a general overview. With a focus text like Was kommunizieren unsere Wettbewerber über KI und Automatisierung?, the report becomes targeted. Focus texts are especially helpful when you have a specific question, not just need an overview.
What the AI Can’t Do
The AI generates reports based on the updates you provide—it cannot find or research updates that aren’t in your inbox. If a company has published a relevant post that Picasi hasn’t retrieved yet, it won’t appear in the report.
Analyses do not contain any claims about a competitor’s internal data—only about publicly available content that you’ve collected in Picasi.
Reports Are Not Deterministic
The same input can easily result in slightly different reports when generated twice. This is normal behavior for AI models. If a report isn’t accurate, it may be helpful to refine the focus text or apply different filters and regenerate the report.
How Conversation Starters Are Selected
Conversation starters follow a different selection logic than other report types. Instead of creating a summary, the AI evaluates individual updates for their suitability as a starting point for a comment on LinkedIn.
Scoring: Each update receives a score from 0 to 14. This score consists of a base score (7–9 points) and possible bonuses (0–5 points). Only updates with a score of at least 8 are included in the results—updates with lower scores are excluded, even if this results in fewer than three suggestions. The stated goal is fewer but better suggestions.
Source Diversity: A maximum of one update is suggested per source. Even if a source has been very active, only its best match is displayed. Exception: A second post from the same source may be included if it has a score of 12 or higher.
Number: The result contains zero to five conversation starters. If no updates reach the threshold, the report remains empty—this is not an error, but indicates that there are currently no suitable points of engagement.
When more suggestions are generated: If the inbox contains many updates from various sources and competitors have recently published particularly relevant content, the hit rate increases. During periods of low activity, few or no suggestions are normal.