Skip to content

Services

← AI Workflow Audit
Illustrative sample · fictional company

Sample audit: client reporting for a small marketing agency

This transparent example shows the format and level of detail you can expect. It is not a client case study, and the company, data, and workflow are fictional.

Team

Six-person marketing agency

Workflow

Weekly client performance report

Goal

Reduce repetitive drafting while keeping human approval

01 · Current state

What happens today

  1. 01

    Export

    A strategist exports channel metrics from three platforms.

  2. 02

    Consolidate

    Numbers are pasted into a spreadsheet and checked manually.

  3. 03

    Explain

    A first draft is written from scratch for every client.

  4. 04

    Approve

    A lead reviews the narrative, fixes claims, and emails the report.

Main bottleneck: the strategist repeatedly turns the same table structure into a narrative. The risky step is interpreting why a metric changed, because the data alone does not prove a cause.

02 · Proposed workflow

Where AI assists—and where people decide

1

Human prepares verified inputs

The strategist supplies the reporting period, checked metrics, campaign context, and known events. No customer personal data is included.

Human
2

AI drafts a structured summary

A reusable prompt turns the supplied facts into highlights, concerns, and questions. It must label assumptions and avoid causal claims.

AI assist
3

Human verifies every claim

The strategist compares all numbers with source data, removes unsupported explanations, and adds business context.

Human
4

AI adapts approved text

After facts are approved, AI can create a concise executive summary and a client-friendly email draft.

AI assist
5

Account lead approves delivery

A named reviewer owns the final report and decides whether it is ready to send.

Human

03 · Prompt playbook excerpt

Two of the five sample prompts

Draft the factual performance summary

Act as a careful marketing reporting assistant. Use only the verified inputs below. Separate: wins, concerns, and questions for the strategist. Do not infer causes from correlation. Label every assumption and write ‘not enough context’ when the inputs do not support an explanation. [Paste verified, non-sensitive inputs]

Run the pre-delivery quality check

Review the report below as a quality-control editor. List every numerical claim, causal claim, recommendation, and missing source. For each item, mark: verified by supplied data, needs human confirmation, or unsupported. Do not rewrite the report until the checklist is complete. [Paste draft]

04 · Review checkpoints

Non-negotiable controls

  • ✓ Source metrics are checked before drafting.
  • ✓ Unsupported causes are removed or clearly labelled.
  • ✓ No personal client or customer data enters the prompt.
  • ✓ A named account lead approves the final report.

05 · Success measures

Measure the pilot, not a promise

  • • Median drafting time per report
  • • Number of unsupported claims found in review
  • • Percentage of reports approved without a second rewrite
  • • Team confidence score after four weekly cycles

06 · 30-day rollout

Small pilot, clear stop conditions

Week 1

Document inputs and baseline time.

Week 2

Pilot with one internal report.

Week 3

Run two client reports with full review.

Week 4

Compare metrics and decide: adopt, revise, or stop.

Want this level of clarity for your workflow?

Describe one repetitive process and its current bottleneck. No confidential data is needed for the first review.

Request your audit →

ChatGPT capabilities and plan limits can change. Verify current availability in the official OpenAI documentation.