“Why did sales change last week?” is not one question an AI agent can answer from a dashboard. It is a request to compare several plausible explanations: media, merchandising, demand, tracking, inventory, and normal variation. Give the agent enough context to investigate those explanations. Do not give it permission to choose the budget response.
The six records an agent needs
1. The business outcome and its definition
Start with the outcome that matters: sales, new-customer revenue, contribution, CPA, MER, or another defined measure. Record how it is calculated, which source owns it, its date window, and known coverage limits. An agent cannot reconcile a platform claim with an order record when “revenue” means different things in each place.
2. Campaign roles and media assumptions
Give the agent the job each campaign has, the channel-specific targets, and any accepted tradeoff. A campaign above its usual CPA may be weak. It may also be a loss leader, a demand-creation effort, or a deliberate test. The agent needs the plan before it calls a variance a failure.
3. Your own commercial calendar
Record promotion dates, offer terms, inventory constraints, price changes, launches, and merchandising decisions. A sales increase during a major sale may be driven largely by the offer and the resulting brand demand, not by the new video launched that same week.
4. External-demand context
Use only context the business can explain and maintain: seasonality, category or brand-interest signals where relevant, weather for a weather-sensitive category, and credible records of major competitor activity. These signals do not prove why sales changed. They keep the agent from treating every shift as an advertising story.
5. Granular performance breakdowns
Let the agent inspect device, network, placement, audience, ad, geography, and landing-page breakdowns where the records support them. A campaign can look weak because one component has absorbed more spend while underperforming. The next step may be to inspect that component, not to judge the whole campaign.
6. Known gaps and the decision boundary
Label broken tracking, identity gaps, delayed order feeds, missing costs, and unresolved definitions. Also name what the workflow may do: summarize, compare, flag, and recommend a check. It should not autonomously pause media, change budgets, alter offers, or change targeting.
A better weekly investigation
Ask the agent to show the period-over-period change, then test the explanation against history and available breakdowns. At low volume, one volatile week is rarely enough to justify a broad response. Look for consistency: did the same change appear in the outcome, the relevant supporting measures, and the parts of the account that should be affected?
A useful output has four parts: what changed; the evidence it used; plausible explanations that still fit the record; and the next bounded check. If Meta reports strong sales while the corresponding GA4 path looks unexpectedly weak, the next check may be tracking or identity—not an immediate channel verdict.
Where the causal claim stops
Historical comparisons can show that a promotion, a category-demand signal, or a media measure has often moved alongside sales. They do not isolate which factor caused the change. An agent can use that history to generate a more credible hypothesis and show what information is missing. It cannot transform observational data into causal proof by sounding certain.
For a consequential allocation decision, use the appropriate measurement step: inspect visible journeys with multi-touch attribution, review the portfolio with MMM, or design an incrementality test when the question is whether a defined intervention created a business effect.
The context layer needs an owner
A maintained marketing context layer gives the workflow its working definitions, campaign roles, promotion history, known gaps, and decision rules. A data foundation and a reporting system make the source records traceable. One accountable lead should approve the operating version, distinguish facts from hypotheses, and decide what action the team takes.
That is the useful role for an agent: assemble evidence faster, identify a narrow anomaly, and preserve the reasoning for review. The person responsible for the budget remains responsible for the decision.
FAQs
Does an agent need every possible data source before it can help?
No. Start with a bounded report and the records needed for one live question. Mark what is missing. A manual export can be enough for an early investigation when access, privacy, and review are controlled.
Can Google Trends or weather tell the agent why sales changed?
No. They may provide demand context worth comparing with the business’s own history. They do not establish the cause of a sales change by themselves.
When may an agent recommend pausing an ad?
It can surface a component that merits review. A named operator should inspect the evidence, alternative explanations, and downside before pausing it.