The operational problem
Funders do not wait for a shared drive to make sense. They expect an account of what happened, what it cost, what changed, and how the organisation knows. The problem is that many teams create that account only when the deadline is close. By then, evidence has spread across calendars, email threads, spreadsheets, personal notes and the memory of people who have already moved to the next urgent task.
This brief is for managers, programme leads, and finance and administration colleagues. It applies the approach in A Practical AI Roadmap for NGO Leadership Teams to the reporting and compliance workflow. The aim is not a faster-looking report. It is a record that is easier to assemble because the work of agreeing definitions, retaining evidence and resolving exceptions happened during delivery rather than after it.
Scope: This is an operational record-keeping and reporting briefing, not legal, accounting, audit or funder-specific advice. Follow grant terms, local controls and professional advice.
The sector context makes this more than a filing problem. The Federation of European Social Employers’ 2023 study reports a Dutch pattern in which care workers spent 36% of their time on administration. Worker representatives in Austria and Belgium described documentation becoming more complex without reliably reducing everyday burden. That does not prove the same proportion in every NGO, but it does explain why a late request for evidence becomes an evening task: the organisation is already trying to deliver support while keeping the administrative trail intact.
The 2026 edition of the Charity Digital Skills Report gives the wider technology picture. As of 2026-07, 79% of surveyed UK charities report using AI, including 92% of large charities, yet only 28% have a digital strategy in place. Adoption has risen from 76% in 2025 and 61% in 2024 while strategy has fallen from 44% in 2025 and 50% in 2024. The relevant warning is not that charities should avoid AI. It is that a licence count cannot substitute for a controlled reporting route with a named owner.
What a real week looks like
The deadline is ten working days away. On Monday, a programme lead looks for the outcome framework agreed at the start of the grant. Finance has a spreadsheet but needs to know which version of the attendance figure is final. A colleague remembers collecting useful feedback in a workshop, although the consent wording for using a quotation is somewhere in last year’s correspondence. None of these are unusual failures. They are ordinary pieces of a programme that were never joined into a reporting trail.
By Wednesday, the programme lead is emailing staff for files they thought had already been shared. Finance is sensibly refusing to certify a number that does not match the ledger. The person who knows why an activity changed is on leave. Thursday produces a polished narrative, but it contains a provisional total and a story that cannot yet be used. Friday is spent reconciling, correcting and apologising for the chase. The visible work has been done. The invisible work is version hunting, duplicate entry and converting spoken understanding into evidence.
Eurofound’s 2023 account of social services describes staff shortages, difficult conditions and uneven digitalisation across the sector. A reporting process therefore has to work on a disrupted Tuesday, not just in a quiet demonstration. If it depends on one confident user, an immaculate folder structure or unpaid catch-up, it is not an operating process.
What the evidence says about reporting and AI
The Charity Digital Skills Report 2026 is especially relevant to this workflow because it includes funders as well as charities. It reports that 63% of charities use AI for administration and project management, while 45% use it for grant fundraising. Those figures make it unsurprising that funders are encountering AI-shaped bids and reports: 58% of funders say AI is changing the bids they receive. Yet 55% do not know whether applicants use AI. That is not an argument for concealing use. It is a reason to make the provenance of claims clearer, because a fluent document will not help a funder judge whether its numbers, quotations and explanations can be traced.
The same report records a difficult implementation environment. Squeezed finances are the largest reported barrier for 63% of charities, 56% identify a lack of skills or technical expertise as the largest AI barrier, and 35% say they do not trust AI tools. Staff training is the top funding need for 44%, while only 17% have received dedicated digital funding. A team that adds a drafting tool without training, source controls and review time may simply create a new task for the person already trusted to check everything.
Charity Commission guidance on grant-making makes the more durable point: monitoring should be proportionate to risk, with clear terms, reporting and evidence. For an NGO, that means a pack should be a by-product of delivery. A dated attendance record, approved expenditure, a consent decision and an explanation of a variance should be available before the final narrative is drafted. The report is then an accountable interpretation of evidence, not a hurried act of persuasion.
Where technology helped - and where it stalled
A controlled evidence register can help. It can show whether the required artefacts exist, link an outcome to the approved source, flag an expired consent decision, and create a first narrative draft from verified facts. Used this way, technology removes the search and formatting work that contributes little judgement. It can also show a programme and finance lead the same status, which is more valuable than two polished documents with competing totals.
It stalls when it makes an uncertain source sound settled. A model cannot decide which of two figures is final, whether an activity is within grant terms, or whether a story may be used. It can also create a new confidentiality problem when staff paste identifiable participant information into a consumer service because the approved route is slow. In that version of the workflow, writing begins faster but review becomes longer, documentation expands and no time is restored to direct work.
The honest test is narrow: did the pilot remove a real handoff, re-entry task or search task? If the answer is no, an attractive report may be concealing the same workload. The 2026 sector evidence is a useful corrective here. Eighty-one per cent of charities report digital progress, but progress is not the same as capacity returned to people.
A review date is not a use case for an agent
Not every gap in this workflow needs an AI system watching it. A lapsed consent decision is usually a missing owner and a forgotten renewal date, not a model problem. A provisional invoice that will not match the ledger is an authority question for finance’s existing sign-off; no tool changes who may certify it true. Some evidence registers fail simply because a funder template still asks for a field the organisation stopped collecting two grant cycles ago. Gartner’s 2025 analysis of agentic AI projects makes the same point at scale: many use cases marketed as agentic do not actually require an agentic implementation, and it names “agent washing”, where rule-based tools are rebranded as autonomous agents to justify a sale.
The pattern across a live grant caseload is different. One missing consent form is an ownership fix; the same gap recurring across six grants with different review dates is where an agentic layer earns its place, reading the pattern, noticing which evidence type is consistently late, and prompting the named owner ten days out rather than waiting to be asked. SAS’s 2026 research on agentic AI return calls this keeping a human “in the lead” over automation that scales their judgement, not replacing it. A consultancy that pitches a reporting agent without first asking whether finance’s sign-off route or the programme lead’s exceptions list already works will sell the same generic module to the next NGO, instead of building on what already functions here.
A safer AI-assisted path
-
Map one recent report backwards. This step has no AI role: the broken handoff must be traced by the people who lived the week. Start with the submitted pack and trace each outcome claim, financial figure and story to the source that supported it, including waiting, duplicate entry, version hunting and after-hours completion. Ask the people who did the work whether the map reflects their actual week before anyone chooses a platform or pilot scope.
-
Define the evidence register and its boundaries. Give every required item an owner, source link, status and review point. An approved register can flag a missing artefact, a stale review date or a field the funder template still asks for but the team no longer collects; the programme or finance owner must confirm each exception before it enters a draft. Decide which material is de-identified enough for an approved drafting tool, and which stays in the case, finance or consent system, with a short rule for what never enters a consumer tool.
-
Pilot one recurring requirement. Choose a monthly activity return, a routine outcome update or a low-complexity grant report, not the most politically sensitive submission. Let the tool retrieve from approved sources and draft a clearly marked first version. Retain links to the underlying evidence so a reviewer can challenge a sentence rather than trusting smooth prose.
-
Set the human gates before the draft moves. This step has no AI decision role: finance signs numerical claims and expenditure, the programme owner confirms delivery and interpretation, and a communications or safeguarding owner approves stories, quotations and consent boundaries, while trustees retain oversight where terms require it. A drafting tool may assemble material up to these gates, but only a named person may move it past them.
-
Review for one reporting cycle before expanding. Compare the pilot with the prior route: time to assemble, corrections, source retrieval and staff experience. If an agentic layer has surfaced the same late evidence type across several grants, use that signal to decide whether to fix ownership, narrow the pilot or expand to a second workflow; do not scale on polished paperwork alone or because one enthusiast built it.
Where humans must intervene
Finance and programme owners must sign numbers and outcome claims because they understand both the calculation and the operational context. Communications, safeguarding or data-protection owners must approve stories and consent boundaries. These gates are not delay for its own sake. They preserve the ability to explain how an external statement came to be made.
NIST’s AI 600-1 guidance is useful at this scale: know the tool, understand its inputs and limits, retain provenance and maintain human oversight for consequential outputs. GDPR does not create an exception because a deadline is close. The approved workflow has to make the safer action easier than the personal-phone or consumer-chatbot workaround.
Risks and failure conditions
The visible failure is a fabricated or untraceable metric. More common are quieter failures: a provisional number copied into the final narrative, an unreconciled invoice, an old consent form treated as permission, or a reporting pack that exists only on one person’s drive. Each weakens the organisation’s ability to answer a reasonable funder question without reopening the entire week.
Do not scale because a tool appears clever. Scale only when there is a named owner, a documented data boundary, a visible approval point and evidence that quality held. If those conditions are absent, the pilot may be an unobserved risk dressed as efficiency.
Measures that matter
-
Hours to assemble a complete reviewed pack, recorded by role. Compare a recent baseline with the pilot, including finance, programme and communications time. This reveals whether work disappeared or merely moved from the programme lead to the checker.
-
Share of required artefacts found through the approved register without inbox search. Test this before the deadline, using ordinary colleagues rather than the person who built the system. A high score shows that evidence is genuinely available, not simply well remembered by one person.
-
Days before the deadline when a complete reviewed draft exists, alongside corrections after submission. Earlier readiness matters only when accuracy holds. A report submitted quickly but followed by corrections has not reduced risk.
-
Unresolved evidence exceptions at the review gate. Count missing approvals, conflicting definitions and uncertain consent decisions. The objective is not zero exceptions by hiding them. It is to surface them early enough for a human decision.
Decision questions
- Which report or return creates the most repeatable reconstruction work, and can we trace its claims to sources now?
- Who owns each evidence type, and which information may never enter an AI tool or consumer service?
- Who can approve numbers, outcome interpretation, quotations and consent, and what happens when they disagree?
- Will a colleague who did not build the pilot find the evidence and use the route safely?
- What will prevent reclaimed minutes being absorbed by more forms, inbox work or an expanding caseload?
- What can we show after ninety days without overstating certainty or hiding exceptions?
FAQ
Who is this brief for?
NGO managers, programme leads, finance and administration colleagues, and anyone who has to turn delivery into a credible external account. For the wider leadership sequence, use the NGO AI roadmap.
Is this about replacing staff with AI?
No. It is about reducing avoidable re-entry, search and chase work while retaining human responsibility for work that affects people.
What is the first step with no platform budget?
Take the last submitted report and trace it backwards with the people who assembled it. An owned folder, a source register and a review rule will often reveal more than a new tool.
Why not just use a consumer chatbot?
These workflows routinely contain sensitive personal information, financial data and consent records. A consumer tool may create confidentiality, provenance and retention problems while producing unverified content. Use approved systems, minimise data, and keep human review around every external claim.
References
- David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
- 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
- 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
- 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
- Charity Commission, Making grants to charities and other organisations (2026). https://www.gov.uk/guidance/making-grants-to-charities-and-other-organisations
- Charity Commission, Regulator publishes new guidelines on grant-making (2026). https://www.gov.uk/government/news/regulator-publishes-new-guidelines-on-grant-making-as-charity-sector-continues-tofeel-financial-squeeze
- ICO, Data sharing: a code of practice. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/data-sharing-a-code-of-practice/
- European Commission, European Care Strategy COM(2022) 440. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52022DC0440
- 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
- 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