Brief
Riverlight Charity Trust is a fictional community charity in a riverside town. It runs a food bank, a youth club and a befriending service that pairs volunteers with isolated older people. Its income comes from individual donors, local businesses, grant funders and community fundraising events.
The marketing and fundraising team is small and leans on volunteers. Over the past year, people have started using AI tools on their own: some for harmless drafting, some in ways that worry the chair of trustees after a volunteer mentioned pasting donor names into a chat tool. The trustees have asked for a written policy before the next appeal.
Your role is the charity's communications manager. The deliverable is the policy itself, written for the people who will follow it, plus a short register of current uses that shows how the policy applies to each. Regulatory remarks in your policy should be stated as general direction; the trustees will take their own advice on anything with legal weight.
Inputs and constraints
Everything below comes from a short survey of the team and a look at the tools in use. Treat it as your evidence and note where you would want more.
Riverlight Charity Trust: the team, the tools and current AI use
| Fact | Detail |
|---|---|
| Team | Communications manager, fundraising officer, 1 part-time designer, about 6 regular volunteers |
| Donor and supporter records | Held in a donor CRM; no customer data platform |
| Tools in use | Email platform with an AI subject-line feature; a free consumer chat tool used by several people; a business AI assistant on trial whose contract forbids training on inputs and sets a 30-day retention |
| Budget for new tools this year | £0 beyond continuing the trial |
| Current use, from the survey | Who | Tool |
|---|---|---|
| Drafting appeal emails and thank-you letters | Fundraising officer | Free chat tool |
| Social captions for events | Volunteers | Free chat tool |
| Summarising monitoring data for grant reports | Fundraising officer | Business assistant |
| Drafting beneficiary stories from caseworker notes | Communications manager | Business assistant |
| Generated photos of "a family at the food bank" for an appeal | Designer | Image generator |
| Personalising thank-you letters by pasting donor names and gift amounts | Volunteer | Free chat tool |
| First-pass translations of the food bank leaflet | Volunteer | Free chat tool |
| Email subject-line suggestions | Fundraising officer | Email platform |
Constraints you must respect:
- The policy must fit on a double-sided page, in plain language, with any detail in an appendix.
- Volunteers must be able to follow it without training beyond a short briefing.
- No new paid tools this year; the business assistant trial can continue.
- Beneficiaries include older people and young people; their stories carry the highest risk.
- The trustees approve the policy and must be named as the route for exceptions.
Task
Work through the steps in order. Each produces a section of the policy or its appendix.
-
Build the register. List every current use with its owner, tool, the data it sees and its tier, using the tiers in the accuracy and review workflows guide. Apply the sorting you practised in the recruitment agency exercise.
-
Set approved and banned tasks. Write the tasks the team may use AI for, the tasks it must not, and the tools approved for each. Base the content tasks on where the AI for content guide says AI fits and where it fails, and say which current uses stop today.
-
Write the briefing rule. Say what every prompt for published work must contain: reader, question, sources, voice sample and the instruction to mark gaps rather than fill them. Point staff to the content brief template and, for images, the creative brief template. Reuse what worked in the travel insurer briefing exercise.
-
Set review by tier. For each tier, state who checks, against what, and whether a second person is required. Include the fact and claim check for appeals that state what a gift achieves, and the rule that the person who prompted does not approve their own published work.
-
Write the disclosure rules. Using the disclosure and data privacy guide, decide when Riverlight tells supporters and beneficiaries that AI was involved, with particular care over images of people and stories about real beneficiaries. Apply the "would anyone feel misled" test you used in the estate agency exercise.
-
Write the data rules. List the data that never enters a prompt, the data that may enter only the approved business assistant, and the minimum fields for tasks such as thank-you letters. Say where personalisation at scale belongs instead of a chat window.
-
Name owners and the exceptions route. Name an owner for each use and say who signs off exceptions, using the approval matrix in the approvals and compliance reviews guide as the model for triggers and reviewers.
-
Set the measures and the review date. Choose how the trustees will know the policy is working, and when the register and policy are next reviewed.
Checking the review load is sustainable
Appeal emails: 4 a month, each needing a fact and claim check of about 20 minutes, so 80 minutes a month. Beneficiary stories: 2 a month, each needing a caseworker check and the beneficiary's consent, about 45 minutes each, so 90 minutes a month. Social captions: about 40 a month, sampled at 1 in 5 by the communications manager at 2 minutes each, so 16 minutes a month. Total review time: 80 + 90 + 16 = 186 minutes, a little over 3 hours a month for the communications manager and fundraising officer between them.
Your figures may differ if you make different, stated assumptions. What matters is that the review your policy requires is something the team can actually do every month, or the policy will be skipped in its first busy week.
Evaluation criteria
Review your own policy against the table, then ask someone outside the team to read it if you can.
| Criterion | What a strong answer shows | Common gap |
|---|---|---|
| Register | Every surveyed use listed with owner, tool, data and tier | Only the uses the author thinks of as "AI"; the email platform's feature missed |
| Approved and banned tasks | Clear lists tied to where AI fits; donor pasting and generated beneficiary images stopped today | A general "use AI responsibly" with no list |
| Briefing | Required contents for every published prompt, with the templates named | Prompting left to individual taste |
| Review by tier | Who checks what, against which source; a second person for published work; specialist or caseworker check for beneficiary stories | One review step for everything, or none for captions |
| Accuracy of appeals | Every statement of what a gift achieves traced to the charity's own figures | Impact claims drafted by the tool and kept because they sound right |
| Disclosure | Rules for images of people, beneficiary stories, bylines and any chat tool, phrased plainly | No rule, or a blanket label on everything that tells readers nothing |
| Data | A short never-list, an approved-tool rule, minimum fields, personalisation kept in the donor CRM or email platform | Data rules that a volunteer could not remember |
| Owners and exceptions | A named owner per use; trustees as the exceptions route, with decisions recorded | Ownership left to "the team" |
| Measures | A register-based KPIA metric tied directly to an objective. such as the share of uses with an owner and a passed review, plus incidents logged | Measures of output volume |
| Usability | Fits on a double-sided page; a volunteer could follow it after a short briefing | A long document nobody reads |
A strong policy also protects the charity's voice. Stories about real people are what supporters trust, and that trust, like a page's E-E-A-TExperience, Expertise, Authoritativeness, Trust., comes from first-hand experience: caseworkers and beneficiaries supply it, a tool cannot. Thank-you letters personalised at scale belong in the email platform or a customer data platform (CDPUnified first-party customer profiles.) with the minimum fields, not in a consumer chat window.
Submission checklist
Before you call the policy finished, confirm each item:
- The register covers every use in the survey, including AI features inside existing tools.
- Approved and banned tasks are listed, with the approved tool for each approved task.
- The uses that stop today are named, with what replaces them.
- Every published prompt has required contents, and the brief templates are referenced.
- Each tier has a reviewer, a source to check against and a rule on self-approval.
- Disclosure rules cover images of people, beneficiary stories, bylines and any chat tool.
- The never-list of data is short enough to remember, and special-category data is on it.
- Each use has a named owner, and trustees are the route for exceptions.
- The measures can be read from the register, and the review load has been worked out.
- Regulatory remarks are stated as general direction, with legal questions left to the trustees' advisers.
- The policy fits on a double-sided page in plain language.
When the policy is complete, mark the lesson as done on the beginner AI in marketing level page. If a criterion exposed a gap, the quickest revision is the matching exercise: the recruitment agency task sort, the travel insurer brief and review or the estate agency review and disclosure rules.
