The operational problem
Creating from scratch every week is sometimes craft. Creating from scratch forever is usually a sign that the organisation has no dependable way to retrieve what it already knows.
This brief applies A Practical AI Roadmap for NGO Leadership Teams to resource creation. It is for educators, therapists and programme leads who repeatedly prepare visual schedules, routine worksheets, activity instructions and parent-facing material. The point is not to make every item standard. It is to distinguish a reusable starting point from a personal draft, then leave professional judgement with the person who knows the learner or group.
The pressure is real. Eurofound’s 2023 account of social services describes staff shortages, demanding conditions and uneven digitalisation. In a related social-partner study, care workers in one Dutch pattern spent 36% of their time on administration, while worker representatives in Austria and Belgium reported that digital documentation could add rather than remove everyday work. That evidence does not measure any particular organisation’s preparation time. It does explain why a twenty-minute search for a nearly identical resource is not a minor inconvenience when it happens across a team.
The Charity Digital Skills Report 2026 gives the digital context. As of 2026-07, 79% of the 807 responding UK charities report using AI, rising from 76% in 2025 and 61% in 2024. Yet only 28% have a digital strategy in place, down from 44% in 2025. Adoption can therefore coexist with a messy shelf, unclear permissions and no decision about who reviews generated material. A resource generator without ownership is simply a faster way to create more near-duplicates.
What a real week looks like
On Monday, a therapist needs a simple visual schedule for an activity that has been run before. A colleague remembers making one last term, but it lives in an unnamed desktop folder. By Tuesday, two new versions exist. One uses language beyond the intended reading level; another carries an old photograph whose reuse and consent status are unclear. Neither has an owner, a review date or a note explaining when it should not be used.
By Friday, preparation has become a blank-page exercise. The practitioner is not short of ideas. They are compensating for a library that cannot tell them what is current, adaptable or safe. A new starter experiences the same problem more sharply: without the informal knowledge of who made a resource and why, they either recreate it or take a risk with an unsuitable version.
This is why the useful unit is not a vast “knowledge base”. It is a small resource family with a common record: purpose, intended audience, accessibility notes, source, owner, review date and limits. A colleague should be able to decide in under a minute whether an item is a sensible starting point. If they cannot, search has not solved the operating problem.
What the evidence says about a safer shelf
UNESCO’s 2023 guidance on generative AI in education and research asks for human-centred use that accounts for privacy, equity, cultural and linguistic diversity, and educator capacity. For this workflow, that is a practical test rather than a broad principle. A fluent worksheet is not evidence that its reading level, examples, images and instructions are appropriate for the people who will receive it.
The Charity Digital Skills Report 2026 also shows why training and ownership belong in the design. Fifty-six per cent of respondents identify lack of skills or technical expertise as their biggest AI barrier, 35% do not trust AI tools, and 44% name staff training as their leading funding need. Those figures do not mean staff should be pushed into trusting a resource tool. They mean leaders should make the permitted workflow comprehensible: what may be adapted, what never enters a consumer tool, and who can say an item is ready for use.
There is a second, less visible issue. A shared resource can preserve outdated assumptions as efficiently as it preserves good practice. The shelf must make currency visible. A review date is not decoration: it tells the next colleague that a resource has been considered recently enough to be used as a starting point, and gives them a route to report what no longer works.
Where technology helped - and where it stalled
Technology can help with the portions of the work that are genuinely repetitive. A controlled shelf can make approved resources searchable by purpose, age band, format and accessibility feature. Accessibility tools can produce a large-print or plain-language variant. Within an approved environment, drafting assistance may turn a de-identified brief into a first-pass adaptation that the practitioner then checks.
Those gains are narrow but worthwhile. They reduce formatting and retrieval work, and they make it more likely that a good existing resource is seen before another is made. They do not decide therapeutic or educational fit. A tool does not know whether a learner’s communication needs, current circumstances or safeguarding context make an otherwise well-written resource unsuitable.
The work stalls when the shelf becomes an uncurated drive, or when a generated output is treated as ready because it looks polished. Staff then spend the saved minutes correcting weak text, checking images, or rebuilding an item they cannot trust. The honest aim is not “more materials”. It is fewer blank-page starts and more time for the practitioner to adapt a known, reviewed base.
A generator does not fix an unowned shelf
The near-duplicate worksheet problem this brief describes, two new versions of the same visual schedule appearing by Tuesday, is not primarily an AI absence. It is a retrieval and ownership failure: nobody labelled the first version, nobody gave it a review date, and nobody had the standing to say which one colleagues should use. A resource generator dropped into that gap does not remove the duplication; it lets the team produce near-duplicates faster and with more apparent confidence than the situation deserves.
The pattern worth an agentic layer’s attention sits one level up, in the shelf itself once it exists. This brief already notes that when several practitioners independently simplify the same instruction, the shelf is telling the organisation something; a bounded agent watching adaptation activity across the resource family can surface exactly that signal, that one item is quietly failing its intended audience, and prompt the shelf owner to update the source before a fourth near-duplicate appears under a new filename. Gartner’s warning about “agent washing”, where ordinary search or drafting tools are rebranded as autonomous agents, and its estimate that only a small fraction of vendors claiming agentic capability are genuine, is directly relevant to how a resource-generator pitch should be read here. Judge it against whether it can watch the shelf and reason about where staff keep recreating the same item, not against whether it can produce a worksheet on demand. McKinsey’s 2025 State of AI findings point the same way: adoption without redesigning the underlying workflow, in this case the shelf and its ownership, rarely shows up as a measurable gain, however fluent the individual outputs look.
A safer AI-assisted path
-
Choose one high-repeat resource family. Start with something that already creates repeated work, such as visual schedules, routine worksheets or a standard parent pack. This scoping step has no AI role: humans must agree which family is worth a shared shelf before any generator or search layer is added.
-
Create a minimum shelf record. Record title, purpose, intended audience or age range, accessibility notes, owner, source date, review date and usage limits. Metadata and status labels are human decisions that make later retrieval trustworthy; an assistant indexes what the owner has approved, it does not invent shelf status.
-
Separate reusable material from identifiable material. Do not use names, identifiable photographs, client stories or case histories as prompts. This boundary has no AI role: consent and data minimisation are human judgements under GDPR, not tasks to hand to a drafting tool.
-
Use AI for preparation, never for final suitability. A controlled tool may suggest a plain-language version or a new layout from an approved template. The practitioner checks content, reading level, cultural fit, accessibility, copyright and safety before use. The human gate is valuable because it tests the material against a real context rather than a generic prompt.
-
Turn adaptation into learning. Record material changes and their reason, then let a bounded agent watch adaptation activity across the family and flag when several practitioners independently simplify the same instruction. The shelf owner decides whether to update the shared base; the agent surfaces the pattern, it does not approve the revised resource.
-
Retire material visibly. An archive route is safer than allowing obsolete items to look current. The named role withdraws items and states where users find the replacement; a retrieval layer can exclude retired status from search results, but only after a human has made the retirement decision.
Where humans must intervene
Practitioners decide whether a resource is educationally or therapeutically appropriate. Managers set the standard for ownership and review, while the person delivering the activity remains responsible for suitability in the moment. Consent, safeguarding, inclusion and copyright questions cannot be delegated to a drafting system.
Human review is also how the shelf avoids becoming restrictive. A standard item should make adaptation easier, not erase the reason a practitioner may need to depart from it. The relevant question is not “did staff use the template?” but “did the starting point leave them more time and attention for the person in front of them?”
Risks and failure conditions
Predictable failures include the wrong reading level, inaccessible formats, stereotypes, reused images without an appropriate basis, copied copyrighted material and old versions presented as current. Consumer AI use creates a separate risk when staff paste identifiable examples into a service that is not approved for that information. The response is not a blanket instruction to be creative without tools. It is a practical, usable approved route.
There is also a maintenance risk. If nobody owns the shelf, staff will stop trusting it, and duplication will return under a new name. Do not buy a generator until the organisation can name the resource family, owner, human gate and ninety-day measure.
Measures that matter
-
Share of routine materials adapted from an approved shelf. Track one defined resource family, rather than every document created. A rising share suggests fewer blank-page starts only when practitioner review still confirms that the material fits its intended use.
-
Time to produce one routine resource. Compare equivalent requests before and after, including search, adaptation and review. A shorter drafting step is not a gain if it introduces lengthy correction or uncertainty about consent.
-
Items with an owner and current review date. This is the shelf’s health measure. Material without an owner should not appear as an approved starting point, because no one has accepted responsibility for its continued relevance.
-
Near-duplicate resources created each month. Review why each duplicate occurred. Missing tags, poor search and lack of trust imply a shelf problem; a genuinely new need may justify a new variant.
Decision questions
-
Which resource family consumes preparation time often enough to justify a shared starting point?
-
What labels let a colleague judge fit, limitations and currency safely?
-
Which data, images and stories are prohibited from the shelf and from prompts?
-
Who approves an adaptation before it is used, and who can retire an item?
-
Within ninety days, will the pilot show fewer blank-page starts without lowering practitioner judgement?
FAQ
Can AI generate activities for children?
It can prepare a draft or variation from a de-identified brief in an approved process. A qualified practitioner must still assess safety, accessibility, fit and professional appropriateness for the individual.
Do we need a new platform first?
No. Begin with a controlled shared location, a small resource family and consistent labels. If staff cannot find and trust five items there, a larger platform will not solve the underlying design problem.
Is this a clinical or safeguarding programme?
No. It is an operational brief about resource retrieval and adaptation. Follow applicable professional standards, safeguarding procedures and local governance.
References
- David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
- Charity Digital Skills Report 2026 (807 respondents; launched 9 July 2026). https://charitydigitalskills.co.uk/report/
- Eurofound, Social services in Europe: Adapting to a new reality (2023). https://www.eurofound.europa.eu/en/publications/all/social-services-europe-adapting-new-reality
- Federation of European Social Employers, Evolving jobs, skills and training needs in the social services sector (2023). https://www.socialemployers.eu/wp-content/uploads/2023/09/report_Evolving_jobs_skills_and_training_needs_in_the_social_services_sector.pdf
- UNESCO, Guidance for generative AI in education and research (2023). https://unesdoc.unesco.org/ark:/48223/pf0000386693
- Information Commissioner’s Office, Data minimisation. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/
- UK Intellectual Property Office, Copyright notice: digital images, photographs and the internet. https://www.gov.uk/guidance/exceptions-to-copyright
- NIST, AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
- Regulation (EU) 2016/679 (GDPR). https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled By End of 2027” (25 June 2025). https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- McKinsey & Company, “The State of AI: Global Survey” (2025 edition). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai