Brief
Dynourkin Shoes is an invented online retailer of children's shoes with an online fitting guide and staffed chat. Much of its year is taken in the weeks before the school term starts, when parents buy in a hurry and exchange sizes just as fast.
The analyst writes queries by pasting customer rows into a personal AI account. A generated tag for the fitting guide went live last month without a preview, and nobody can say which of the revenue figures in the board pack came from generated SQL. Support wants an assistant to answer sizing and delivery questions in chat and drafts for its agents. The agency wants broad match, broad audiences and an add-to-cart optimisation event, citing the platforms' own reports, and wants all three live before the season.
This plan runs AI in production for a year inside the operating model from the advanced AI in marketing level: what ships, what is tested, in what order and how each result is read.
Inputs and constraints
Dynourkin Shoes: the team, the year and the AI use register
Team: the owner; a marketing manager; an analyst 3 days a week; a customer service lead with 4 agents, 7 in season; a fitting specialist who owns the fitting guide; an operations manager who owns delivery, returns and exchanges; a freelance developer with publish rights on the tag container; a paid media agency.
Year: the back-to-school season runs from mid-July to the second week of September and carries 41% of annual revenue; a spring sale runs for 2 weeks in April; returns and exchanges are accepted within 30 days.
| Register row | Tool | Owner | What the row shows |
|---|---|---|---|
| Revenue and channel queries | The analyst's personal AI account | None recorded | Rows pasted with parents' names, postcodes and children's foot measurements |
| Fitting guide tag | An approved assistant; published without preview | None recorded | Live for 5 weeks |
| Product page markup | An approved assistant, written page by page for the 60 best sellers | None recorded | A review rating on 22 pages that show no reviews |
| Board pack revenue | Unknown | None recorded | 2 of last quarter's 3 monthly figures differ from the shop platform's report, by 9% and 14% |
Tracking plan row: fitting_guide_complete, with size_result and fit_width, sent once when the guide's result
shows. The live tag sends fitting_complete on each of the guide's 4 steps, with a parent_email parameter taken
from the "save my child's size" form.
Dynourkin Shoes: contact reasons and the help centre
| Contact reason | A month out of season | A month in season | Content behind the answer | Owner |
|---|---|---|---|---|
| Which size for my child | 520 | 2,100 | Fitting guide and size charts | Fitting specialist |
| Will these fit wide or narrow feet | 180 | 640 | Wide-fit article, 2 versions live | None named |
| Will it arrive before term starts | 90 | 1,300 | Delivery policy | Operations manager |
| Where is my order | 300 | 1,500 | Read-only order lookup | Operations manager |
| Exchange for another size | 260 | 1,200 | Exchange policy | Operations manager |
| Faulty shoes | 60 | 150 | Guarantee terms | Customer service lead |
| Complaints | 40 | 210 | None | Customer service lead |
- Product-page and fitting guide chat: about 3,000 visitors open it a month out of season; 26% order within 30 days; contribution after returns and exchanges averages £14 an order.
- Agents' target is chats closed per hour. The saved reply library holds 38 replies, 12 with no owner.
Dynourkin Shoes: the ad accounts and the agency's proposals
| Campaign | Set-up | A month out of season |
|---|---|---|
| Google Ads, core search | Exact and phrase; Smart Bidding, target cost per conversion £9 | 900 conversions; most within a day of the click, some up to 2 weeks later |
| Google Ads, brand | Brand keywords | 1,400 conversions |
| Meta, prospecting | Interests in parenting and school preparation; purchase optimisation; customers excluded | 520 purchases reported, 450 matched to shop orders |
| Meta, wide-fit range | 3 ad sets sharing the budget; purchase optimisation; learning limited | 30 purchases, 260 adds to cart |
The agency's proposals: core search to broad match; prospecting to Advantage+ audience; the wide-fit campaign to add-to-cart optimisation. All three to go live in the first week of July.
Constraints you must work within:
- Nothing is launched, changed or tested in the back-to-school season, and every test is read before the season starts.
- Customer records go only into tools the data protection lead has approved. Any data rule in the plan is a starting point for that lead to confirm, and nothing in it is legal advice.
- The shop platform is the record of orders and revenue; platform-reported figures are not.
- The assistant says it is automated, and a person is one request away throughout every conversation.
Task
- Gate AI-written technical work. Write the gate every query, tag and piece of markup passes before it ships, built on the AI-written code review checklist, so whatever a Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. writes is proved against something independent before it reaches a report. Then apply the gate to what is already live: the analyst's practice is stopped and handled under the AI use policy; each board pack figure is rerun and reconciled as the routine for checking AI-written SQL sets out; the fitting guide tag is rebuilt to its plan row and previewed under both consent choices following the tracking code routine; and the markup becomes a mapping from template fields with every rating removed where reviews do not show. Give each use its row, owner and tier in the AI use register.
- Scope and gate the assistant. Place every contact reason by risk, keep complaints and customers who may be vulnerable with people, and design the handover and the disclosure before choosing what the assistant answers. Fix the wide-fit article and the unowned replies first. The AI customer service launch checklist is passed before any customer sees the assistant.
- Split saved replies and drafts. Use the saved replies and AI drafts comparison to split the delivery promise, the exchange policy and fitting questions between an owned Saved ReplyA pre-written answer a support agent inserts and then edits for the customer in front of them, keeping facts and tone consistent across the team. and drafts, naming each owner, and replace the chats-per-hour target before drafts arrive.
- Design the assistant test. Write the plan in the experiment hypothesis template from the AI chat assistant test: its primary metric, each Guardrail MetricA metric an experiment must not worsen beyond an agreed bound, watched beside the primary metric that decides the test. with its limit, the margin and the decision rule written before launch, and the reading after the exchange window.
- Plan the automation tests. Write one plan per proposal on the measure its page sets: the broad match expansion test under the Smart BiddingGoogle Ads' automated bid strategies that set each auction's bid from conversion signals. target, the audience targeting breadth test and the optimisation event test, the wide-fit ad sets consolidated first. The guide to AI in ad platforms is why the platforms' own reports decide none of them.
- Build the calendar. Order the gate, the launch, the drafts, the chat test and the three platform tests so each result is read alone and none falls in the season. Record each decision, with its check date and what would reverse it, in the marketing decision log.
- Write the plan. One document with a page per part, naming its owner, calendar slot and reading, with the working behind it appended.
Evaluation criteria
| Criterion | What a strong answer shows | Common gap |
|---|---|---|
| AI-written technical work | Requires a written definition before any request, field names and invented rows in an approved tool, reading and testing against an independent total, release through the usual rules and the verified version saved with its prompt | Generated work trusted because it runs |
| Assistant scope and launch | Sorts every contact reason by risk, keeps complaints and customers who may be vulnerable with people, designs handover and disclosure first and passes the launch gate before any customer sees the assistant | Scope chosen by volume and the handover designed last |
| Drafts and saved replies | Gives policy answers to owned saved replies and varied product questions to drafts from complete content, lets drafts draw on the library and leaves commitments to agents | One method for every question |
| The assistant test | Judges assistant-first chat on contribution per chat visitor, net of returns and the cost of answering, with accuracy, handover and repeat-contact guardrails and a rule written before launch | Deflection reported as the result |
| Platform automation tests | Tests one change at a time, judged on the measure each test page sets, with purchases reconciled to the shop's orders where the page says so, sets learning periods aside, waits for conversion delay and writes each decision rule before launch | The platforms' own reports taken as proof |
| A calendar that can be read | Puts the launch, the chat test and the platform tests in an order where each result stands alone and nothing lands inside the back-to-school season | Everything switched on at once just before the season |
A strong submission also says what the agency hears about its July date and why, and which of last month's problems the gate would have stopped, at which line.
