AI-generated editorial image. Illustrative scene; not a photograph of our team or a client.
A useful client report answers three questions: what changed, what might explain it, and what should happen next. AI can help assemble the narrative, but a fluent summary can conceal a wrong denominator or an unsupported explanation. Build the workflow around evidence, review, and a clear release decision.
Define the report before choosing the assistant
Select one report your team already produces, such as a monthly paid-media review. Identify the client decision it supports: keep the plan, investigate a problem, or approve a proposed change. Set the audience, length, reporting period, and required sections. An executive needs a different document from a campaign operator.
Write a data contract in plain language. List the approved source, account identifiers, currency, timezone, date range, and definition for every key metric. Specify whether conversions follow a platform definition or an agreed business definition. If two systems disagree, preserve the discrepancy rather than letting AI silently pick the more flattering number.
Keep calculation and explanation separate
Export or query the approved data into a controlled table. Use checked spreadsheet formulas or tested calculations for totals and ratios. Include source links or export references. AI may help write a formula, but a reviewer should verify it against a small example with a known result.
In a hypothetical report, one campaign spends $600 for 20 qualified leads and another spends $400 for 10. Combined cost per qualified lead is $1,000 divided by 30, or about $33.33. Averaging the two campaign costs of $30 and $40 would produce the wrong combined figure. No generated prose can fix an incorrectly prepared metric.
Draft a narrative that exposes uncertainty
Provide the verified table, approved definitions, and notes about material changes. Ask for observations, possible explanations, and recommended investigations as separate sections. A budget increase, a tracking change, or a new landing page can affect interpretation; the model should not invent causality from two numbers.
Keep the first version short. A good reporting assistant can say that the evidence is insufficient. It can also identify the one decision that needs client input instead of filling space with generic optimization advice.
Prompt to tryDraft a client performance brief from this verified table and context: [table, definitions, dates, notes]. Separate observed changes, plausible explanations, and proposed next actions. Cite the table row or source reference for every numeric claim. Mark missing or conflicting information. Do not claim causation from correlation, invent benchmarks, promise outcomes, or recommend budget changes as already approved. End with the decision we need from the client.
Build an approval gate that actually blocks delivery
Use a clear sequence: prepared data, generated draft, review, approved version, delivery. The sending step must depend on an explicit approval record for that specific version. A message saying “please review” is insufficient if the automation continues sending regardless.
Give the reviewer a concise checklist and a way to reject the draft. Confirm totals, date comparisons, metric definitions, unsupported explanations, client-sensitive details, and the requested next step. If the report changes after approval, require approval again. Keep delivery permissions separate from draft-generation access where the tools support it.
- Data owner confirms the input period, account, definitions, and totals.
- Account lead checks interpretation, recommendations, and commitments.
- Release owner approves the exact artifact and recipient list.
- A failed or missing approval leaves the report unsent.
Test the unpleasant cases
Run the workflow with a missing data source, zero conversions, mismatched currencies, a stale export, and a rejected draft. The output should identify the problem and stop at the appropriate stage. A blank or zero value must not automatically become a success story.
Use a test destination under your control when validating delivery. Verify that retries cannot send the same report repeatedly, and store enough history to identify what was approved and sent. NIST’s generative AI guidance discusses risks including confident false output; your operational checks should target the particular errors your clients could encounter.
Prompt to tryAudit this draft against the supplied source table, client instructions, and approval criteria: [materials]. List each numeric inconsistency, unsupported explanation, ambiguous comparison, accidental commitment, and missing caveat. Show the exact statement and supporting or conflicting evidence. Return ready for human approval or needs revision. Your response is a review aid and must not trigger delivery.
Schedule only after a successful manual run
A saved reporting prompt does not create a recurring job. Scheduling requires a configured trigger, current access to the sources, an output destination, failure handling, and the approval dependency. Check these in the actual automation platform and record the timezone and next run.
Measure preparation time, review time, corrections, and client usefulness across several cycles. Expand only when the complete process improves. The service your client values is a trustworthy recommendation with a clear next step; fast drafting is one component of that service.
Try it on one real task
- Choose one existing client report and document its decision and metric definitions.
- Generate a draft from a verified sample, then complete both review passes.
- Test a rejected approval and a missing source before enabling any recurring run.
Check before you use the result
- Every number traces to verified data.
- Observed changes and explanations are distinguished.
- Approval applies to the exact report version.
- No scheduler or delivery step is assumed to exist because a prompt mentions it.
Take the next step
Download the Waymaker Agency AI practice pack for these prompts, exercises, and review checklists.
New to AI? Take the free first-workflow course at AI Teachers. Already using it? Practice building a repeatable workflow.
Sources and further reading
Sources support the principles discussed; the examples and suggested workflows are original teaching exercises.
Want this applied to your business? The Growth Fix is one fixed-price session to find where your customers drop off and fix the part that matters most.
