Recommendations Agent

introduction the recommendations agent enhances copilot by offering suggested actions based on the query you submit and the data returned these recommendations are designed to help you move faster — highlighting next steps you can take without needing to dig deeper on your own how it works after you ask a question in copilot, the recommendation agent evaluates the result and suggests up to three possible actions these actions are based on your data and the context of your question for example you ask “who are my most loyal customers?” copilot returns the segment → the recommendation agent may suggest “create a segment of these users in audiences” or “send this group to meta ads ” about confidence levels each recommendation includes a confidence rating — high, medium, or low — based on how directly the action connects to your original question and the resulting data confidence levels are defined as high a clear, logical next step based on the data (e g , query reveals churning customers → suggest creating a retention segment) medium a reasonable action, but with some inference required (e g , data implies seasonal trends → suggest campaign timing adjustment) low speculative or loosely related (e g , suggesting product changes based on traffic data alone) note chord displays all recommendations returned , regardless of confidence level (up to three per query) how confidence is determined confidence levels are determined by the language model (llm) powering copilot we've prompted the model with guidelines like “prioritize high confidence recommendations over speculative ones ” “err on the side of fewer, higher quality recommendations ” this means the model self assesses whether an action is a strong fit for your query when it's unsure, it’s designed to return fewer or no recommendations, rather than suggest low confidence actions just to fill space learning from your actions copilot also remembers which recommended actions you take and will use that information to improve future suggestions over time, this feedback loop helps the recommendations agent prioritize the types of actions you and your team actually finds valuable, making recommendations smarter and more tailored to your workflows