AI can do the number-crunching on 18 months of timesheets, but the recommendations need a human gut-check.

The analytical structure is consistent — margin by service line, scope creep signals, client-type segmentation — but the specific data shape, project code conventions, and business definitions will vary each time this is run. A one-off setup effort is required to map the data correctly.

The deliverable format (1–2 page summary, charts, 3–5 recommendations) is reasonably crisp, but 'scope creep patterns' and 'best margins' require judgment calls about thresholds and definitions that the task doesn't fully specify. The agent will need to make defensible assumptions and surface them.

The user has the raw data (timesheets, project codes, rates, invoiced revenue), but it must be exported and handed to the agent in a usable format. Chart generation requires a code-capable agent with Python or similar; no live API access is needed, but data prep is a real prerequisite.

A miscalculated margin or misattributed hours could lead to bad strategic decisions — dropping a profitable service line or over-investing in a losing one. The output is a summary, not an automated action, so a human review step before acting keeps the error cost manageable.

Pattern detection and margin math are well within AI capability, but translating findings into actionable recommendations requires knowing the agency's growth goals, team dynamics, and client relationships — context the agent won't have unless explicitly provided.