Die Learning-Seite lesen und danach handeln

Learning sits in the AI section of the left sidebar, under Knowledge Base. Open it and you get a page titled Learning with a subtitle describing what you are looking at: recorded outcome signals across AI Autopilot campaigns.

A scope switcher sits at the top right with three options: Workspace, Campaign, and Chat. Workspace is selected by default and aggregates every AI Autopilot campaign in the workspace. Campaign narrows to one campaign. Chat is a different thing entirely and reports what Chat has learned from your approvals, feedback, and prospect replies.

Keep the scopes straight, because they fill from different sources. Workspace and Campaign fill from real outbound sends and the replies they produce. Chat fills only from your verdicts inside Chat, so a busy Workspace scope tells you nothing about whether Chat has learned anything, and an empty Chat scope is normal in a workspace where nobody rates replies.

Under the scope switcher are three tabs: Overview, Outcomes, and Graph. They are three views of the same evidence, from summary to detail.

The same Learning view is also available inside a campaign builder as a tab, which is the faster route when you already have one campaign open and want its numbers rather than the workspace roll-up.

The Overview tab opens with a status banner. In the normal state it reads "Signals collected" and explains the situation in two sentences: outcome signals are being recorded, automatic policy changes remain paused, and the status applies to the selected workspace. At the right of the banner is a Last model update timestamp with the note that it is the most recent signal applied.

Read the paused line carefully, because it is the part people misread. The system is collecting evidence and reporting patterns, and it is not silently changing how your campaigns write. Nothing on this page is being applied to your sending behind your back.

Under the banner is a row of counters. Retained signals, last 90 days is the volume of evidence held in the current window. Recorded observations is how many observations have been logged. Mature outcome observations is the subset whose reply windows have closed, and it is the number that matters most, because everything else is provisional until it grows. Segments covered tells you how many distinct audience segments the evidence spans. Indexed chunks reports indexed content available to the system.

A workspace with hundreds of retained signals and zero mature outcome observations is a workspace that has sent plenty and learned nothing yet. That is a timing state, not a fault. It resolves when the reply windows on those sends close.

A low segments-covered number is worth noticing too. Evidence gathered from one segment describes that segment, so treat the patterns below as advice about the audience you have actually been emailing.

Below the counters is the Current pattern readout, introduced with the same caution as the banner: these are evidence-weighted patterns from the selected workspace, and automatic policy changes remain paused until outcome evidence qualifies.

Two panels sit side by side. What is working holds patterns receiving favourable observed signal. Needs attention holds patterns receiving negative observed signal. Each pattern is a chip written in plain words rather than a metric name, for example a share-an-asset call to action, bullet points, longer copy, opening with a personal signal about the prospect, or asking fewer than one question.

The two panels are often in direct tension, and that is the useful part. Seeing "longer copy" on the left and "very short copy" on the right is a readable instruction: this audience is not responding to two-line emails. Seeing "open with a personal signal about the prospect" on the left and "open with a workflow or process observation" on the right tells you which opener to write next.

Do not treat a chip as a rule to hard-code. A pattern on the left is evidence from your own sends in your own segments, and it can flip as the audience changes. The right way to use it is to adjust your campaign brief in the AI Instructions tab and then come back to see whether the chip moved.

Chips are also the fastest way to spot a campaign brief fighting itself. If your brief demands short copy and this readout has short copy under Needs attention, the brief is the thing to change.

Click the Outcomes tab. Two cards sit at the top. Treatment shows the number of prospects assigned to treatment, with a breakdown under it: how many were assigned, how many executed, how many are mature, and how many have an unknown execution state. Control shows the baseline holdout with the same breakdown.

The holdout is what makes the rest of this tab trustworthy. Without a group that did not receive the learned treatment, a rise in replies could be seasonality, a better list, or luck. The control arm gives the comparison something to sit against.

When outcome windows have not matured, the page says so directly rather than showing a number. The message reads that lift is not yet available, because assigned sends exist but their outcome windows have not matured. A note under the cards accounts for excluded data, including legacy sends that predate the snapshot and prospects that ended up in mixed arms.

That exclusion note is worth reading before you argue with the totals. A workspace with thousands of legacy sends will show a much smaller assigned population here, and the gap is by design rather than missing data.

Under the cards is a Rate by signal chart comparing control and treatment across the outcomes that matter: qualified positive reply, meeting or goal, open, click, unsubscribe, bounce, and out of office. A table below repeats the same signals with treatment, control, and lift columns. Read the negative signals as carefully as the positive ones. A treatment arm with more replies and more unsubscribes is not a clean win.

Click the Graph tab for a visual map of the same evidence. Nodes are patterns, grouped by the categories in the legend: campaign, opener, call to action, length and shape, structure, and semantic. Edges are coloured by verdict, with preferred and avoid as the two states.

The footer explains the interaction and the encoding: you can drag nodes to rearrange, scroll to zoom, and node size and edge width scale with learned weight. Zoom controls and a fit-to-view button sit at the bottom left.

Size and width are the whole point of this view. A large node with a thick edge is a pattern with real evidence behind it. A small node on a thin edge is a pattern the system has barely seen, and reacting to it is reading noise. The chip lists on the Overview tab do not show that difference, which is the reason to come here when a chip surprises you.

The campaign nodes at the centre are the anchor. Following the thick edges out from a campaign shows what that campaign's sends have actually taught, and following avoid-coloured edges shows what to stop doing before it costs you more replies.

Treat the graph as a reading tool rather than a control surface. Nothing here changes your campaigns. Any change you decide on gets made in the AI Instructions tab of the campaign, then shows up here later as new evidence.

The page reports. You act. A short loop keeps that honest.

• Check Mature outcome observations first. If it is zero, read the patterns as early signal and do not rewrite a campaign on them.
• Pick one chip from Needs attention rather than four. Changing several things at once means the next readout cannot tell you which change worked.
• Make the change in the campaign's AI Instructions tab, in the explicit instruction fields, then confirm it landed in the effective brief pane.
• Come back after the send volume has grown and check whether the chip moved between panels.

Two cautions. First, patterns are drawn from the segments you have emailed, and a low segments-covered count means the advice is narrow. Second, automatic policy changes stay paused, so nothing improves because you read this page. The loop only closes when you edit a brief.

The Chat scope follows the same logic with a different input. It fills from your approvals, rejections, and reply ratings inside Chat, and it stays empty until you produce those verdicts. If you want that scope to become useful, rate replies as you go rather than waiting for the page to populate itself.