Real estate
A marketing team running project campaigns across Meta and Google is generating performance data faster than it can read it. Spend, impressions, click through, cost per lead, audience segment and creative variant all sit in separate platform dashboards, and lead quality sits somewhere else entirely, in the CRM.
The consequence is not that reporting is slow. It is that the connection between ad spend and booked revenue is only ever assembled retrospectively, once a month, by hand. By the time a campaign is shown to be underperforming on qualified leads rather than raw leads, three more weeks of budget have gone through it.
An analyst agent that reads directly from the advertising platforms and the CRM together, runs the same analysis a performance marketer would run, and produces a written read of what is working and what should be cut. It runs on a schedule and on demand.
The important part is the join. The agent does not report on cost per lead. It reports on cost per qualified lead, because it can see what happened to each lead after it entered the CRM.
Model and surface
Claude, running as a scheduled agent with a dashboard front end for the marketing team and a scheduled email digest for the marketing head.
Connectors
Meta Ads and Google Ads for campaign, ad set, creative and audience level performance. Salesforce for lead status, qualification outcome, site visit booked and closure. Google Sheets or Drive for any offline spend the team tracks outside the platforms.
Skills
The analysis logic is encoded as a skill rather than left to the model's judgement, which is what makes the output consistent week to week. The skill defines the qualification rule the client actually uses, the attribution window, the minimum spend and minimum lead volume below which a campaign is not judged at all, the thresholds that classify a campaign as scaling, holding or cutting, and the specific comparisons to run: creative against creative within an ad set, audience against audience within a campaign, channel against channel at project level, and current period against the trailing four weeks.
Data flow
The agent pulls the period's platform data, pulls the corresponding CRM records, joins them on campaign and lead source identifiers, computes the derived metrics the platforms do not provide, runs the classification rules, then writes the analysis. Anything that fails the join, a lead with no traceable source, is reported as an unattributed bucket rather than silently dropped, because the size of that bucket is itself a data quality signal the team needs to see.
Human review
The agent recommends. It does not change budgets, pause campaigns or edit creative. Every recommendation carries the figures behind it so the marketer can disagree with it on evidence.
Governance
Read-only credentials on both ad platforms. CRM access scoped to lead source, status and outcome fields, with no access to customer contact details, since the analysis does not require them.
Monday morning, the agent pulls the previous week across both platforms. It finds that a Meta campaign on one project is delivering the lowest cost per lead in the account and would, on platform data alone, be the obvious one to scale. Joining to the CRM, it finds that campaign's leads are qualifying at roughly a third the rate of the account average, so its true cost per qualified lead is among the worst.
It flags the campaign, shows both numbers side by side, identifies that the gap sits almost entirely in one broad audience rather than across the campaign, and recommends narrowing rather than cutting. It also notes a creative variant in a different campaign that is being under-served budget despite the strongest qualified lead rate in the account.
The marketing head reads that at 9am with the numbers attached, and makes the call.
1,800 annual hours saved
Accelerates budget optimization using weekly qualified lead data instead of raw costs.
