Why this matters
Report commentary looks ideal for AI: the same structure every month and readers who want the headline. The guide to AI for analysis and reporting names the trap: narrative explains, and the data in a report usually does not. A tool will describe what moved, which can be checked, and then say why, which usually cannot. A workflow that runs every month also needs the ownership, change control and audit that the automation governance guide sets out, because it will keep running after the person who built it has moved on.
The firm and its monthly report
Tallybrook Accountants is an invented firm serving sole traders and small companies. Its marketing manager spends a day each month writing commentary for the partners, who use it to decide the paid budget. A marketing assistant has tried pasting the report table into an AI assistant.
Tallybrook: September report table
| Channel | Spend | Enquiries (Sep) | Enquiries (Aug) | Cost per enquiry | Consultations | New clients |
|---|---|---|---|---|---|---|
| Google Ads search | £2,400 | 60 | 48 | £40 | 24 | 9 |
| LinkedIn ads | £1,200 | 15 | 20 | £80 | 5 | 1 |
| Organic search | £0 | 70 | 66 | Not applicable | 21 | 7 |
| Email newsletter | £0 | 12 | 14 | Not applicable | 6 | 3 |
| Referral partners | £0 | 18 | 17 | Not applicable | 12 | 6 |
| Total | £3,600 | 175 | 165 | £20.57 | 68 | 26 |
Notes: enquiries come from the analytics export and channels use last-click attribution. The CRM records 184 enquiries for September. The site release log shows the enquiry form tag was missing for three days mid-month.
The assistant's AI draft (excerpt): "Enquiries rose 12% month on month to 175, driven by the new LinkedIn campaign. Google Ads cost per enquiry fell to £38. Organic search delivered 40% of new clients, so we recommend moving LinkedIn budget into SEO."
Designing the workflow
Treat the workflow as a chain: data in, checks, draft, review, sign-off, record. Check inputs before drafting, because a model projects a broken tag as if it were the business. Let the draft describe and a person explain. The guide to AI for marketing automation adds the rule for AI that drafts: separate drafting from acting, so nothing reaches the partners without a person's approval.
Task
- Check the AI draft line by line against the table and list every error by type.
- Write the input checks that must pass before the tool drafts anything, and what happens when one fails.
- Write the drafting rules: what the tool receives, what it may and may not write, and the template it follows.
- Assign the operator, reviewer and approver, and say what each checks.
- Specify the audit trail and the inventory entry for the workflow.
- List the cases in which a person writes the commentary instead.
- Choose the measures that show whether the workflow is worth keeping.
Worked answer
The draft. Every sentence has a defect, and none would be caught by proofreading.
Tallybrook: checking the draft against the table
| Draft statement | What the table shows | Error type |
|---|---|---|
| Enquiries rose 12% | (175 − 165) ÷ 165 ≈ 6.1% | Wrong figure |
| Driven by the new LinkedIn campaign | LinkedIn enquiries fell from 20 to 15 | Invented cause, contradicted by the data |
| Google Ads cost per enquiry fell to £38 | £2,400 ÷ 60 = £40 | Wrong figure |
| Organic search delivered 40% of new clients | 7 ÷ 26 ≈ 27%, on last click | Wrong figure; attribution model unstated |
| Move LinkedIn budget into SEO | No evidence in the report for either side | Recommendation outside the tool's job |
The analytics total of 175 also sits 9 below the CRM's 184, about 5% short, which matches three days of missing form tracking. The month should not be narrated until that is resolved.
Input checks. Before drafting: every source has delivered a complete month; enquiry totals in analytics and the CRM agree within a tolerance the manager sets from past months; no tracking incident is open in the release log; definitions match the firm's own, so an enquiry and a new client mean what the partners think. A failed check stops the run and goes to the manager, who either fixes the data or writes the month by hand with the gap stated. This month fails.
Drafting rules. The tool receives the table, the definitions sheet, last month's approved commentary and a fixed template: headline, movements against last month and the same month last year, channel notes, open questions. It may state what moved, by how much and against which comparison, and must quote only figures in the table. It must name the AttributionAssigning conversion credit to touchpoints. model whenever it credits a channel, mark any cause as a question for the reviewer, and make no budget recommendations. Prompt and template are versioned; a change is logged and re-run against a regression set of past months whose approved commentary is known.
Roles. The marketing assistant operates: runs the checks and the tool. The marketing manager reviews: traces every figure to the table and rewrites causes after checking breakdowns. A partner approves. Nobody approves their own draft.
Audit trail and inventory. Each run logs the data snapshot, check results, prompt and model version, raw draft, reviewer's edits, defects by type and sign-off. The inventory row records purpose, owner and deputy, trigger, dependencies, measure, status and last review date, and the workflow is reviewed quarterly.
Human-written cases. Any failed input check; a movement well outside normal variation with no known cause; a definition or tracking change during the month; commentary that recommends moving budget; and the first few months, run in parallel with the manager's own version until the defect rate is known.
Measures. One KPIA metric tied directly to an objective. per level: defects per draft found in review, review time against the old writing day, and whether the partners' decisions still rest on figures reproduced from source. Cost per enquiry (CPACost of each conversion.) belongs to the report, not to the workflow.
Common mistakes
- Narrating a broken month. Input checks come first; fluent commentary on missing data is worse than none.
- Letting the draft explain. Causes are hypotheses until a person checks a breakdown.
- Operator as reviewer. The person who prompted it reads what they meant.
- No regression set. A prompt edit quietly changes every future month.
Next, take the checkpoint on AI for automation and analysis or return to the intermediate AI in marketing level.
