The operational problem
If finding the policy takes a meeting, the policy is not really available.
This brief is for managers, clinical leads and administrative staff. It applies A Practical AI Roadmap for NGO Leadership Teams to one retrieval problem: finding the current instruction quickly enough to support a real decision.
In one line: Staff cannot find the current policy version when a decision is needed, so they rely on a meeting, a private message or an old file that looks official.
What a real week looks like
On Monday, an administrator needs the current consent route before sending a family form. The shared drive returns five PDFs: two drafts, one scanned copy with no date, a “final final” file and an old appendix. A clinical lead is in a session. The administrator uses the newest-looking document and hopes it is right.
Wednesday brings a safeguarding question. A manager convenes a short call simply to establish which escalation wording applies. By Friday, someone has forwarded the answer by email, creating another plausible but unowned source. The problem is not a lack of policy. It is that source status, owner and review date are invisible at the moment of need.
Eurofound identifies uneven digitalisation and staffing strain across social services. Under that pressure, people will take the quickest available route. A retrieval system must therefore make the safe route easier than the workaround.
What the evidence already says
The UK National Archives’ records-management guidance emphasises that organisations need to manage information across its lifecycle, including retention and disposal. For this workflow, that means obsolete drafts cannot remain in the same search space as live instructions without a clear status. Retention is not the same as operational availability: an old file may need to be kept, but it should not masquerade as current practice at the moment a colleague needs an answer.
NIST’s generative-AI profile identifies risks such as confabulation and information integrity. In policy retrieval, the relevant safeguard is simple: every answer should link to a current source, show its version and tell the user when no authoritative answer was found. The Charity Digital Skills Report 2026 reports 79% AI use among surveyed UK charities, but only 28% with a digital strategy in place. Its 56% skills barrier and 35% distrust figure reinforce the point: a new interface cannot substitute for records governance.
Where technology helped - and where it stalled
Search, metadata and versioned repositories can reduce the minutes spent finding a policy. A well-bounded assistant can translate a plain-language query into links and short extracts from approved sources. It can be especially useful for a new starter who does not know the organisation’s file names.
It stalls when old files are indexed, owners do not review content or an answer sounds more certain than its evidence. A fluent policy summary with no source can be more dangerous than a slow folder search because it hides uncertainty. Technology also fails to restore capacity where staff must still double-check five documents because nobody has retired the ghosts.
The register problem an agent cannot shortcut
Much of what looks like a search problem here is an ownership problem. The five PDFs on the shared drive, two drafts, one undated scan, a “final final” file and an old appendix, do not need a cleverer assistant on top of them; they need a named owner willing to quarantine four of the five. No retrieval technology fixes a document nobody has agreed is current, and pretending it can only postpones the register work this brief already asks for.
Where an agentic layer earns a place is once that register exists and the organisation wants to know where it is still failing. A bounded agent watching search activity across the register can flag policies queried repeatedly with no confident result, review dates that have quietly lapsed, and owners who have not touched their assigned documents in months, then prompt that owner directly rather than waiting for the next safeguarding call to expose the gap. McKinsey’s 2025 State of AI survey found that most organisations layering AI onto an unchanged process see no material bottom-line effect, while the minority who redesign the workflow around it account for most of the measured gain; a search tool bolted onto an unmanaged drive is the layering case. Gartner’s account of generative AI projects abandoned after proof of concept blames the same root cause, weak underlying data, more often than a weak model. A vendor proposing a policy chatbot without first asking who owns each document is selling the layering version. The agentic version this brief describes is narrower: it watches a register that already works and tells a person when it needs attention.
A safer AI-assisted path
-
Create one register of current policies. Include title, owner, approval date, review date, classification and canonical link. This step has no AI role: named ownership and review dates fix accountability before any search layer is added. A tool cannot substitute for agreeing which document is current.
-
Quarantine or retire obsolete versions. Keep records where retention rules require them, but remove them from routine retrieval and label them clearly. A human approves what is archived and what stays live; an assistant should not move documents out of search space without that sign-off.
-
Use a source-first search experience. A bounded assistant retrieves links and short extracts from approved register entries only, showing the exact document, relevant section, version and review date. A short summary is an aid to reading, not the policy itself; the user must still judge whether it applies.
-
Define no-answer and escalation behaviour. When a source is absent, conflicting or overdue for review, the assistant must refuse to guess and route the query to the named owner. Staff receive a clear stop signal, not a fluent best-effort answer.
-
Test realistic questions. Ask staff to retrieve policies in plain language through the bounded search layer, then verify source accuracy and elapsed time themselves. Include less-digital colleagues and new starters so the test reflects real judgement, not model fluency.
-
Review failed searches monthly. Let a bounded agent surface repeated no-result queries, lapsed review dates and owners who have not updated their entries; the register owner then decides what to fix in tags, wording or the policy itself. The agent prompts; it does not rewrite live instruction.
Where humans must intervene
Policy owners approve wording, version changes and retirement. Staff apply policy to the facts in front of them, particularly where safeguarding, clinical judgement or legal duties are involved. An AI tool can quote and point; it cannot create policy, determine eligibility or override professional responsibility.
Risks and failure conditions
Risks include treating an outdated draft as current, giving a confident but unsupported answer, searching personal or consumer systems with case detail, and leaving no audit trail of sources used. GDPR requires purpose limitation and data minimisation. Do not turn a policy query into a case narrative in an unapproved tool.
Measures that matter
-
Median time to retrieve the correct current policy. Time a small set of real questions and verify the result against the register. Speed without correctness is not an improvement.
-
Search results displaying owner, version and review date. This measures whether staff can judge a source, not merely whether they received a text answer.
-
Meetings called only to locate a document. Log this narrow category for a pilot period. The reduction shows whether retrieval has become genuinely self-service.
-
Use of retired documents. Treat each occurrence as a source-control failure and correct the index or archive route.
Decision questions
- Which policies create the most urgent retrieval failures?
- Who approves and reviews each source?
- What content is excluded from routine search?
- What must the system display before staff rely on an answer?
- What happens when the policy is unclear or out of date?
FAQ
Is a chatbot the first step?
No. First make the current source set small, owned and versioned. A chatbot over unmanaged documents increases the chance of a polished wrong answer.
Can policy search include case facts?
Not in a general tool. Keep identifiable case information inside the approved system and use professional judgement to apply policy.
References
- David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
- The National Archives, Records Management Code of Practice. https://www.nationalarchives.gov.uk/information-management/manage-information/planning/records-management-code/
- NIST, AI 600-1: Generative Artificial Intelligence Profile (2024). https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.600-1.pdf
- 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
- 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/
- ICO, Data minimisation. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/a-guide-to-the-data-protection-principles/
- European Data Protection Board, guidelines and recommendations (general guidance index). https://www.edpb.europa.eu/our-work-tools/general-guidance/guidelines-recommendations-best-practices_en
- ISO, ISO 15489-1: Information and documentation - Records management. https://www.iso.org/standard/62542.html
- Gartner, “Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes” (2026). https://www.gartner.com/en/articles/genai-project-failure
- McKinsey & Company, “The State of AI: Global Survey” (2025 edition). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai