Brief
Marigold Home Care is an invented provider of care at home for older people, run from three branch offices. Its marketing has two jobs that pull in different directions: bringing in enquiries from families arranging care for a parent, and recruiting carers, without whom no new client can be taken on. A marketing manager, a part-time coordinator and a freelance ads specialist do all of it.
AI has arrived without a plan. The ads specialist has left Performance Max and Meta's automated audiences on their defaults. The coordinator drafts social posts and job adverts in a free chatbot. Last month a branch manager pasted an enquiry note, which described a client's diagnosis, into the same chatbot to "tidy it up". The managing director wants the benefits, wants that never to happen again and has asked for a plan that says what the team will use, under which rules, and how anyone will know whether it worked.
Families choose care slowly and on trust. Anything that reads as careless, invented or impersonal costs more here than in most businesses, and the plan has to show it understands that.
Inputs and constraints
Marigold: where marketing time goes each month
| Task | Hours a month | Who | Notes |
|---|---|---|---|
| Carer job adverts and social posts | 16 | Coordinator | Same roles re-advertised across branches |
| Family enquiry follow-up emails | 12 | Coordinator | Templates edited by hand per enquiry |
| Monthly marketing report and commentary | 10 | Manager | Directors read the summary only |
| Website and blog content on care topics | 14 | Manager | Health-related claims need checking |
| Paid ads management | 20 | Freelancer | Performance Max for family enquiries; Meta for carer recruitment |
| Replying to reviews | 4 | Manager | Some reviews mention named clients and carers |
Marigold: the last quarter's paid campaigns
| Campaign | Spend | Platform conversions | What counts as a conversion | Outcomes in the care management system |
|---|---|---|---|---|
| Performance Max, family enquiries | £9,600 | 410 | Any form submit, including the careers form | 120 genuine care enquiries, 31 new clients |
| Meta, carer recruitment | £4,500 | 300 | Application started | 54 completed applications, 11 carers hired |
Tool budget for the year: £6,000, excluding staff time. Average carer recruitment cost the business is willing to bear is set by the operations director at £600 per hire.
Constraints you must work within:
- Client health information is special-category data and never goes into a marketing AI tool. Treat the plan's privacy rules as general direction to be confirmed with the data protection lead, not legal advice.
- The care management system is the system of record for clients, carers and consent; there is no customer data platform (CDPUnified first-party customer profiles.), and the email platform receives contacts through a native connector.
- Claims about care quality, outcomes or ratings must be traceable to evidence before they are published.
- Testimonials and photographs of clients or carers need their written consent, and no synthetic person may be shown as a real client or carer.
- The team cannot absorb more than one new tool at a time.
Task
- Prioritise the use cases. Score each task in the time table on value, review cost and risk, using the method in the AI applications overview, and sort them into adopt now, pilot, defer and refuse. Say why review replies naming clients and any health content are not first in the queue.
- Set the ad platform guardrails. Fix what each campaign optimises towards, using the outcomes the care management system already holds, and set exclusions, URL controls and structure as in the garden tools retailer guardrails exercise. Separate family and recruitment goals. Check the result against the campaign launch QA checklist.
- Design the automation guardrails. For any workflow that uses AI, such as enquiry follow-up drafts or report commentary, say where the model sits, what stays a rule, what it may draft and never send, and the logging and regression set, following the guide to AI for marketing automation and the accounting firm report commentary exercise.
- Set review tiers and roles. Assign each adopted use a tier, source pack, reviewer and approver from the accuracy and review workflows guide. Decide which content is regulated or sensitive and who gives specialist sign-off.
- Write the data and privacy rules. The three data classes, what may enter which tool, vendor terms required and the response to last month's incident, from the disclosure and data privacy guide. Say when AI use is disclosed.
- Evaluate and pilot one tool. Write the requirements, weight and score the candidates, cost them over two years and design a time-boxed pilot with a holdout and a decision rule, following the guide to choosing and evaluating tools and the holiday park tool evaluation exercise. Specify the data flow from the system of record with the data flow and integrations guide, and write the pilot on the experiment hypothesis template.
- Plan onboarding and training. Apply the marketing tool onboarding SOP to the chosen tool, and plan short training for the team and branch managers on the data rules, review tiers and how to report a problem.
- Choose the measures and the scale rule. For each adopted use, one KPIA metric tied directly to an objective. for the outcome, one for the AI component and one for cost and risk, and the return on investment (ROIProfit relative to cost.) calculation that decides whether it scales. Say how results reach the managing director through the monthly marketing report template.
- Write the one-page plan. Every decision above on the one-page marketing plan template, with the detail in appendices, an owner for each part and the dates on which each will be judged.
Evaluation criteria
| Criterion | What a strong answer shows | Common gap |
|---|---|---|
| Use-case priorities | Each task scored on value, review cost and risk, with refusals as well as adoptions and a reason for each | Everything adopted at once, or health content first because it takes most time |
| Ad platform guardrails | Goals rebuilt on genuine enquiries and hires from the system of record, careers traffic separated, exclusions and structure set | Defaults left on and judged by platform conversions |
| Automation guardrails | Models drafting, rules deciding, nothing sent without approval at first, logging and a regression set | An assistant wired to send family emails on day one |
| Review tiers | Tiers assigned at the brief, operator never approving their own output, specialist sign-off for care claims | One review for everything, or proofreading called review |
| Data and privacy | Three data classes, no special-category data in any tool, vendor terms checked, the incident answered with a rule and training, framed as general direction | A ban nobody can follow, or rules with no owner |
| Tool choice and pilot | Requirements before demos, weighted scores with evidence, two-year cost, a pilot with an end date, a holdout and a decision rule | A tool chosen from a demo, piloted until someone likes it |
| Onboarding and training | The SOP followed before real data flows, named owner and deputy, training before editing rights | Accounts shared, training left to the vendor |
| Measurement | A measure per level for each use, return on investment worked from time and outcomes, a rule for what scales and what stops | Time saved claimed without review time counted |
A strong submission also states what the team will not use AI for this year and what evidence at the first quarterly review would change the plan.
Submission checklist
When you have finished, review your plan against the evaluation table, then return to the intermediate AI in marketing level to mark the level complete.
