Agency reporting automation: a practical weekly AI workflow

Start with the decisions the report supports

Ask who reads the report and what they need to decide. Choose a small set of sections, such as completed work, open issues, and decisions needed. A fixed structure makes it easier to compare weeks and notice missing information.

Collect before summarising

Gather the relevant records for the same time period before asking for a narrative. Keep links or identifiers that let a reader check important statements. If a source is unavailable, show the gap rather than silently producing a complete-looking report from partial data.

Worked example: a weekly agency delivery report

This hypothetical example extends a task-based operations report to a five-person marketing agency. The account lead needs to decide which client deliverables need attention before the next planning call. The inputs are an approved task export, an agreed reporting window and the account lead’s decision log. Use fictional client labels in testing. This is a workflow design, not a customer result.

For fictional Client A, the source records show six deliverables due in the window: four completed, one awaiting client feedback and one blocked by a missing asset. The output should preserve those counts and source IDs. A useful draft reads: “Four of six deliverables are complete. One awaits feedback; one needs an asset. The account lead must confirm who will obtain the asset.” The numbers are invented to show the format.

If the task board is unavailable, the output becomes “delivery status unavailable for this window,” with a collection error for the reviewer. It must not reuse last week’s four completions as current activity. If an advertising export is absent, the report cannot infer campaign performance from delivery tasks. A task marked complete also does not prove the client accepted it.

The account lead checks the six source records, confirms the unresolved owner and decides whether the draft is ready for the client. Distribution remains a separate approved step. Keep the report version, reporting period and reviewer decision together so a corrected draft cannot quietly replace the version someone approved.

Specify the input and output before connecting tools

Croissant recommends a short reporting contract: source system, included accounts, reporting start and end, governing time zone, required fields and collection status. Limit collection to the client records needed for this report. A broad connection to every account creates unnecessary exposure and makes checking the draft harder.

The output contract can contain completed work, blocked work, decisions needed and source references. Separate observed facts from the writer’s interpretation. Record the extraction time and any unavailable sources beside the draft. Calculate totals using explicit rules before asking AI to explain them; ask the reviewer to check both the arithmetic and the meaning.

Anthropic’s context-engineering guidance discusses selecting relevant context and retaining identifiers for retrieval. ^[1] Our reporting contract applies that principle to one bounded business task. It does not establish that an AI narrative will be accurate simply because its inputs are well organized.

Review changes and unresolved questions

Ask the agent to distinguish what changed from what is merely restated. Review numbers against their source and label interpretation as interpretation. Give each unresolved question an owner when possible. The report should point to work that needs attention, not conceal it behind a polished summary.

A reality check before you hit run

Anthropic emphasizes relevant context rather than indiscriminately adding more material. For a weekly report, define the reporting window and source records first. ^[1]

Anthropic’s evaluation guidance supports testing agent outcomes against explicit criteria. For reports, compare totals and conclusions with the source data. ^[2]

Decide whether to buy a workflow or keep the template

If your team already exports consistent data and only needs a fixed table, a spreadsheet template or scheduled export may be enough. A custom workflow becomes worth investigating when collecting context across approved systems and routing exceptions are recurring burdens. Ask for a demonstration using your agreed report structure, including a failed collection step.

For a pilot, compare manually checked reports with the draft output. Record factual corrections, missing sources, review time and unresolved decisions. Write acceptance rules before inspecting the drafts. Anthropic describes task-specific evaluations and outcome checks; use those ideas to test the whole reporting workflow rather than judging prose quality alone. ^[2]

Before buying, ask which systems are included, who maintains the mappings, what happens after a source changes and who can pause collection. Agree that client distribution needs a person’s approval. Bring Croissant the report structure, source tools and the decision your reader needs to make through the existing contact form. Use fictional examples and leave private client records out of the first enquiry.

Implementation checklist

• Define the reader, decision, and reporting window.
• Keep source links alongside important facts.
• Show unavailable or incomplete sources.
• Review figures and assign the next actions.
• Define one output structure and a reviewer for each report.
• Test a missing source, an old status and conflicting reporting periods.
• Separate draft preparation from approved client distribution.
• Compare factual corrections and review effort before expanding.