The operational problem
Ask Maria. If Maria is on leave, wait. If she has left, reconstruct the answer from email threads, a shared drive and the memories of three busy colleagues. That is not a knowledge base. It is a continuity risk disguised as helpfulness.
This brief is for operations leads and senior practitioners. It applies A Practical AI Roadmap for NGO Leadership Teams to one workflow, rather than repeating the full roadmap. The objective is not to turn professional judgement into fixed instructions. It is to stop routine, repeatable work being rebuilt whenever a colleague changes role.
Turnover affects more than recruitment cost. ALNAP’s humanitarian-agency research describes unplanned turnover as damaging learning, efficiency and the ability to continue programmes, in part through loss of institutional memory. A 2023 systematic review of turnover-related knowledge loss reaches the same practical conclusion: organisations need preventive mechanisms matched to the knowledge at risk, not a generic repository. The useful question is therefore not what the organisation knows in the abstract, but which repeatable routes become unsafe or slow when this person is absent.
What a real week looks like
On Monday, a new coordinator receives a referral from a school. They know the service is meant to acknowledge it within two working days, but cannot find the current route. A former colleague’s checklist is in a personal folder. The referral form has been revised twice. The safeguarding escalation contact in an old PDF is no longer right.
Tuesday becomes a chain of interruptions. A senior practitioner answers the same questions between sessions. An administrator forwards a twelve-month-old email because it contains a useful template. By Wednesday the coordinator has completed the task, but not in a way anyone can repeat with confidence. On Friday, another colleague asks where the first-call script lives. The week has created four more fragments of undocumented practice.
The loss is not only minutes. It makes new starters cautious where they should be clear, and casual where they should pause. In helping professions, a concise route with an explicit escalation point is safer than confidence built from a plausible old answer.
What the evidence already says
Eurofound reports that social services continue to face staff shortages, difficult conditions and uneven digitalisation. About one fifth of social-care workers in the EU never use digital devices at work. An intranet that only the confident few can navigate can therefore deepen, rather than solve, the continuity problem. The test is whether a new starter can use it during a pressured day, not whether it looks complete to its authors.
The Charity Digital Skills Report 2026 reports that 79% of surveyed UK charities use AI, rising from 76% in 2025 and 61% in 2024. Only 28% report a digital strategy in place, down from 44% in 2025, while 56% identify lack of skills or technical expertise as their largest AI barrier and 35% do not trust AI tools. The lesson is not that every charity needs AI search. It is that a tool cannot supply ownership, source quality or a review cycle that the organisation has not designed.
| Evidence pattern | Operational implication |
|---|---|
| Turnover interrupts learning and institutional memory | Identify the routes that would fail if one person left tomorrow. |
| Social services have uneven digital access | Test pages with less-digital staff, not only the people who created them. |
| Charity AI adoption exceeds strategy | Start with source control and an owner before adding conversational search. The 2026 edition reports 79% AI use but only 28% with a digital strategy. |
| NIST advises documenting and governing generative-AI use | Record what content is indexed, who can change it and what the system must not answer. |
Where technology helped - and where it stalled
Technology helps when it reduces a real retrieval problem. A short, well-tagged playbook can stop a senior colleague being interrupted five times for the same answer. Versioned templates can stop people adapting an obsolete document. Search can surface the current referral route faster than a shared-drive maze, and drafting support can turn a practitioner interview into a first version of a page.
It stalls when the source material is weak. A search layer over duplicate folders merely makes stale content easier to find. An AI answer that is fluent but cannot point to a current page makes uncertainty harder to see. The 2026 charity evidence is relevant here: high use sits beside a strategy gap, skills concerns and distrust. Capacity is not restored if nobody owns the page after launch.
Do not confuse a knowledge base with a replacement for supervision. The useful system says, “Here is the current process, last reviewed on this date, owned by this role.” It also says, “This does not decide safeguarding, clinical judgement, eligibility or exceptions.”
Search is not a substitute for an owner
Much of the “ask Maria” problem here has nothing to do with AI. If nobody owns the referral-route page, a better search tool just retrieves an unowned page faster. If a template was revised twice without anyone updating the escalation contact, the fix is a review date and a named editor, not a model. McKinsey’s 2025 State of AI survey found that most organisations see no material return from AI because they layer it onto an unchanged process rather than redesigning the workflow around it, which is exactly the risk of pointing a search layer at a shared drive nobody has tidied: it retrieves the wrong version with more confidence, not less.
The pattern worth an agentic layer sits across the whole shelf, not one search query. Across dozens of pages and a rotating team, the same practitioner keeps answering the same five questions, or a page keeps going stale after its owner changes role, and nobody notices until a new starter is misled by it. A system that reads which pages are queried most, which have no recent review date, and which owner has left, then prompts a named operations lead before the shelf rots further, is doing analytical work a static wiki cannot. SAS’s 2026 research on agentic AI return describes the useful pattern as automation that scales a practitioner’s judgement rather than substituting for it, which is the difference between a page that explains the escalation route and a chatbot that guesses one. A vendor pitching conversational search over an unmanaged drive, without learning that a departure-test discipline already finds the right five pages, will sell a search box that finds stale answers faster.
A safer AI-assisted path
-
Run a departure test. This step has no AI role: ask each team which five answers would become hard to find if a particular experienced colleague left next week. Capture the route, source documents, exceptions and escalation point from the people who hold the knowledge; do not ask a model to infer what matters from an unmanaged drive.
-
Write short operational pages. Each page should answer one task: when to use it, the steps, required system links, what not to do, owner and review date. An approved drafting tool may turn a practitioner interview or existing material into a first version, but a senior colleague must validate every step before the page is indexed or searched.
-
Separate practice from personal data. This step has no AI role for case content: use de-identified examples and links to approved systems, and do not turn client histories, identifiable case notes or sensitive correspondence into a general knowledge corpus or paste them into consumer tools. A page can explain the route without duplicating the facts of a person’s case.
-
Introduce search only over approved, current pages. Search results should display title, owner, version and review date, with a link to the source. If the system cannot find a reliable source, it should say so and direct the colleague to a named human route.
-
Test with a new starter and a sceptic. This step has no AI role: give each a realistic task and watch where they hesitate. Their failure to find, understand or trust a page is better evidence than a polished search demonstration or a fluent answer with no source.
-
Review one small set quarterly. Retire superseded pages, assign a new owner when roles change and sample whether the process is still followed. Where an agentic layer has flagged the same five questions recurring, or pages going stale after an owner leaves, the operations lead acts on that signal before expanding the shelf; a smaller living set is safer than an expanding archive.
Where humans must intervene
Senior practitioners validate practice, exceptions and escalation language. Supervisors decide whether a colleague is ready to use a process independently. A buddy remains necessary for contextual judgement, relationship-building and the situations no template should compress.
Human review is mandatory where an answer affects safeguarding, a care decision, eligibility, legal obligations or external communication. The organisation must also decide who can edit a page and who can authorise a change. These are operating responsibilities, not configuration details.
Risks and failure conditions
The common failure is shelf rot: a page looks official long after its owner or policy has changed. Other risks include turning unverified notes into apparent policy, indexing restricted client material, and making a chatbot the sole route to an answer. Each problem becomes more serious when a rushed colleague assumes the system is authoritative.
Privacy matters even if the initial work appears administrative. Health, disability and child-related information may be special-category personal data under GDPR. Keep the knowledge base focused on operational method and use approved case systems for case facts. NIST’s generative-AI profile is useful for the discipline of documenting risks, users and human oversight, not as permission to automate judgement.
Measures that matter
-
Top ten repeat questions resolved from a current page. Record the questions that interrupt senior staff most often, then sample whether colleagues can find a source without a private message. A reduced interruption count is meaningful only if the retrieved answer is still correct.
-
Pages with a named owner and live review date. Track this as a proportion of the pilot shelf. A page without either field is background material, not an operational instruction.
-
Time to complete one supervised core task for a new starter. Compare cohorts carefully and keep the task consistent. The target is earlier safe competence, not pushing someone into independent work before they are ready.
-
Repeat clarification requests to senior practitioners. Count a small agreed set of channels for four weeks. If requests move from chat to a search box but errors increase, the pilot has not succeeded.
Decision questions
- Which answers would disappear if one experienced person were absent for a month?
- Can every operational page show an owner, source and review date?
- What categories of data are excluded from the knowledge shelf and any AI tool?
- Where must the system stop and direct staff to a supervisor or safeguarding lead?
- Can a new starter locate and use the path without creating a new local workaround?
- Who has time and authority to retire stale content?
FAQ
Is a shared drive enough?
It can be enough for a very small, disciplined team, but folders alone rarely show which file is current, who owns it or when it was checked. Start by improving those controls before paying for search technology.
Can AI write the playbooks?
It can turn an interview or existing approved material into a draft. A practitioner must validate every step, especially wording about safeguarding, eligibility, consent and escalation.
What should we do first with no budget?
Run the departure test from step one of the safer path on a single team this week: ask which answer a senior colleague gets asked most often, then write just that one page using the format from step two, stored with version history in whatever document tool the organisation already pays for. Do not build a second page until a genuine new starter, not the page’s author, has tried to use it under realistic time pressure and it has held up. A library of untested pages is not progress; one page a new starter can actually follow is.
References
- David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
- ALNAP, Understanding and Addressing Staff Turnover in Humanitarian Agencies (2006). https://alnap.org/help-library/resources/network-paper-55-understanding-and-addressing-staff-turnover-in-humanitarian-agencies/
- S. M. N. et al., Knowledge loss induced by organizational member turnover: a systematic literature review (2023). https://diva-portal.org/smash/get/diva2:1793919/FULLTEXT01.pdf
- 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
- Charity Digital Skills Report 2026 (launched 9 July 2026; 807 respondents). https://charitydigitalskills.co.uk/report/
- NIST, AI 600-1: Generative Artificial Intelligence Profile (2024). https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.600-1.pdf
- CIPD, Induction (updated 2025). https://www.cipd.org/en/knowledge/guides/employee-induction/
- ICO, Special category data. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/
- Regulation (EU) 2016/679 (GDPR). https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Regulation (EU) 2024/1689 (EU AI Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- McKinsey & Company, “The State of AI: Global Survey” (2025 edition). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- SAS, “For Agentic AI ROI, Invest in Human Judgment” (22 July 2026). https://www.prnewswire.com/news-releases/sas-for-agentic-ai-roi-invest-in-human-judgment-302831330.html