Affiliate Intelligence

#Intelligence answers which model is usable here

Open Affiliate → Intelligence to inspect Expected Sales model availability and the system’s automatic selection at account, Shop, and market scope. It is not another task queue and does not replace approval in Agent Workbench.

APP CAPTUREThe English light-theme Affiliate Intelligence page with model availability at account, Shop, and market scope.

#Select a scope before reading the result

The same model may have different coverage and availability in different Shops or markets. Select a scope on the left, then inspect:

• which model signals are currently available;
• whether the system can make an automatic selection;
• why a model was not selected or is unavailable;
• whether the data and model state were recently refreshed.

Use account scope for overall capability, Shop scope for a specific operation, and market scope for regional coverage. Do not transfer a verdict from one scope to another.

#Two comparisons answer different questions

#Same Sample count: select better within the existing budget

The model and historical staff selection use the same number of Sample opportunities and compare predicted sales. Distribution charts show which sales bands contribute, rather than only an uplift percentage.

Historical human selections also have predicted values. Do not mix model-predicted totals with human actual sales as if they shared a basis. Actual historical sales are shown separately. Historically rejected Creators have no directly observable post-Sample sales; their predictions are not realized incremental sales.

#Same sales threshold: find potentially missed opportunities

At the same expected-sales threshold, compare historically approved, above-threshold Creators with the model-selected population. This identifies potential opportunities, not an instruction to approve all of them immediately.

The historically implied threshold describes the model cutoff at the historical approval proportion. The current Shop setting is the business reference supplied to the Agent. They are not the same setting, and these model comparisons are not a live randomized A/B experiment.

#Read evidence strength before the uplift

• Outperformance probability: evidence for the statistical comparison, not the success probability of the next Creator.
• Primary range: uncertainty around predicted lift, not a guarantee.
• Data foundation: the strength of the evaluation evidence. Check window, sample coverage, and availability before using a figure for a business decision.

Switching account/region/Shop model views inspects their evaluations. The automatic-selection marker identifies the system's source; viewing another scope does not manually change the model running the business. If a fallback is shown, use its actual source rather than assuming an independent Shop model exists.

#How it supports a human decision

Expected Sales is one piece of decision evidence. A Sample review or other material action should still consider Creator performance, Product fit, historical choices, current business rules, and Approval Policy. If policy requires review, the Agent proposal enters the Workbench; the Intelligence page does not approve it.

When a signal is unavailable, do not substitute a missing value for a prediction. Open the proposal to see the evidence that was actually frozen with that decision. Keep human review in place or adjust policy in Team & Channels when needed.

#Troubleshooting order

1. Refresh and wait for the selected scope to finish loading.
2. Compare account, target Shop, and relevant market scopes.
3. Determine whether one model or the whole scope lacks a usable signal.
4. Open the concrete proposal in the Workbench and verify its frozen evidence.
5. If the issue persists, check Shop authorization, data synchronization, and model configuration.