Progress Reports That Funders and Families Both Understand

The operational problem

One person’s progress should not need three incompatible report formats to be heard. Families need a respectful explanation of what has changed and what happens next. Funders need traceable activity, outcomes and a defensible account of variance. Staff need a record accurate enough to support continuity. When every audience begins with a separate blank page, the final week of a reporting cycle becomes a rewriting exercise rather than a conversation about the work.

This brief applies A Practical AI Roadmap for NGO Leadership Teams to that problem. It does not promise that a writing tool can create trustworthy reporting from weak records. It asks the organisation to agree which evidence is collected once, who checks it, and how it may be translated without changing the claim.

The sector context makes this more than a writing preference. Eurofound’s 2023 research describes social services working through staff shortages, demanding conditions and uneven digitalisation. The associated social-employers study reports a 36% administrative-time pattern among Dutch care workers and warns that technology can add documentation burden. Those figures are sector patterns, not a local baseline. They are enough to justify examining repeated reporting as a capacity issue.

As of 2026-07, the Charity Digital Skills Report 2026 records AI use among 79% of its 807 charity respondents, including 63% using AI for administration or project management. Yet only 28% have a digital strategy in place and 56% name skills or technical expertise as their biggest AI barrier. A team may therefore have a drafting tool while lacking a data dictionary, authoritative record or named reporting owner. The result is faster prose built on the same unresolved confusion.

What a real week looks like

On Tuesday, a key worker writes observations after a review conversation. On Thursday, a manager asks for a funder table showing contacts, goals and outcomes. Friday brings a family update, written with care because a sentence that seems neutral to a funder can sound cold or alarming to a parent. The facts overlap, but dates differ between documents because someone copied an earlier version and updated only one of them.

Before a grant or contract deadline, people hunt through spreadsheets and inboxes, then ask colleagues what happened. The final report can be polished while still leaving a serious question: where did this outcome statement come from, and did the person closest to the work approve it? That question is the practical reason to create an evidence spine before considering assisted drafting.

An evidence spine is a short, agreed set of fields: intended outcome, dated observation, action taken, current status, source and owner. It does not reduce a person to a spreadsheet. It makes the basis for a claim visible, so a family view can use clear language and a funder view can aggregate activity without inventing causation.

Evidence, interpretation and the role of technology

The Charity Digital Skills Report 2026 reports that 35% of responding charities do not trust AI tools, more than double the 2025 level. That is a useful warning for reporting. Trust should not be sought by asking staff to accept fluent output. It is built by showing exactly what source a draft used, what was omitted because the evidence was missing, and where a human reviewer can correct it.

NIST’s Generative AI Profile identifies confabulation, privacy and human-AI configuration risks. In reporting, confabulation is not always a dramatic falsehood. It can be a plausible sentence saying confidence improved when the approved record only says attendance increased. The sentence may sound compassionate, but it silently turns an observation into an outcome claim.

Technology can help once the evidence is sound. It can flag a missing field, retrieve an approved chronology, prepare a first-pass family summary, or turn the same approved fields into a funder table. These are retrieval and drafting tasks. They can reduce repeated formatting and late chases, but they cannot decide what the evidence proves or what a family needs to hear.

It stalls when different systems contain conflicting versions, when the source material is incomplete, or when review is added on top of rather than in place of duplicate work. A faster draft does not shorten a reporting cycle if staff still search three systems and correct invented certainty. The first benefit should be fewer re-entries and fewer last-minute requests, not a higher count of generated reports.

An agent cannot decide what counts as evidence

Deciding what counts as evidence is not an AI question in this workflow, and this brief should say so plainly. Defining terms such as “engagement” and “progress” so they do not shift meaning between authors, and agreeing who checks totals against the contract, are ownership and vocabulary tasks. A team that installs a drafting tool before settling those definitions has skipped the decision that actually determines whether a report is trustworthy, however well the resulting sentences read.

The correction trail this brief asks teams to keep is where a wider-pattern layer has real value. A single missing source or wrong date is a one-off; the same field failing across several reporting cycles, or the same practitioner repeatedly correcting an inferred outcome the evidence never supported, is a pattern worth attention before the next deadline rather than after it. An agentic system that reads the correction log across cycles and prompts the reporting owner that a particular field keeps failing is doing something a one-shot drafting assistant cannot, because it reasons about recurrence rather than producing a single document on request. SAS’s 2026 research on agentic AI return on investment found organisations that keep automation scaling a person’s judgement, rather than substituting for it, substantially more likely to see a return worth having; that is the layer this brief describes, not a generator asked to guess at causation. Gartner’s finding that at least half of generative AI projects are abandoned after proof of concept, usually over weak underlying data, is a fair caution for any vendor proposing to draft straight from an incomplete record.

A safer AI-assisted path

1. Choose one recurring report with a stable owner. A quarterly funder update or planned family-review summary is a better first pilot than the most sensitive document in the service. This step has no AI role: the cycle needs a known deadline, a named owner and enough examples for every output to receive full human review.

2. Define the evidence spine before drafting. Agree objective, observation, date, source, action, current status and next step. Define terms such as “engagement” and “progress” so they do not change meaning with each author; this vocabulary work is human-only, as the brief already states, because no assistant can settle what counts as evidence.

3. Create deliberately different audience views. From the same approved fields, a controlled tool may prepare a first-pass family summary and a funder table layout, each clearly labelled as draft. A relationship-aware colleague and the reporting owner must verify tone, totals and confidentiality before anything leaves the organisation.

4. Make review roles specific. The practitioner checks factual and contextual accuracy; the reporting owner checks totals and required fields; a relationship-aware colleague checks family-facing tone. The assistant’s job is to show which fields it used and which it left blank; it does not perform any of these checks.

5. Keep a correction trail. Record material corrections, then let an agentic layer read that log across cycles and flag fields that fail repeatedly or inferences that keep being overturned. The reporting owner decides whether the spine, a template or practitioner training needs changing; the agent surfaces recurrence, it does not alter evidence.

6. Measure the entire ninety-day cycle. Compare evidence cut-off to sign-off, late submissions, duplicate requests and corrections. If draft generation adds a second review queue without removing duplicate entry, the reporting owner should pause or redesign the pilot in response to those measures and any recurrence signals the agent has surfaced, not treat continued drafting as progress.

Where humans must intervene

Humans decide what counts as evidence and whether a statement is fair, proportionate and understandable. The practitioner must not delegate professional judgement to a template. The reporting owner reconciles contractual definitions and totals, while a family-facing reviewer recognises when a technically accurate phrase would be insensitive or misleading in context.

No generated content should be sent externally without approval. Identifiable and special-category data must remain in approved systems and processing arrangements. Where a report concerns safeguarding, health, disability, education decisions or another consequential matter, local governance decides whether assisted drafting is appropriate at all.

Risks and failure conditions

A single generic report can satisfy nobody: it can reduce a person’s story to funder language, or turn nuance into an unsupported outcome. “Data laundering” is another risk, where an unverified note enters a summary and becomes apparently authoritative through repetition. The remedy is not more polished language; it is an accountable route back to the source.

Privacy failures are practical too. A family-facing draft can carry identifiers into an aggregate return, or a staff member under deadline pressure can paste a chronology into a consumer service. If the approved path is too slow to use, that is an operating problem to fix. Do not solve it by weakening data controls.

Measures that matter

• Hours from evidence cut-off to signed-off report. Include chases, drafting, corrections and approval. A shorter drafting stage is not success if the total reporting cycle grows.

• Share of required fields collected once and reused. Sample reports to identify facts requested again for a different audience. This exposes duplication that a writing tool cannot see.

• Corrections by type. Separate factual changes, missing evidence, tone changes and numerical reconciliation. A fall in edits is useful only when reviewers are still checking carefully.

• On-time submission with quality held. Track deadlines alongside a small quality sample. A report delivered on time but later corrected with a family or funder is not a successful cycle.

Decision questions

1. What is the authoritative evidence source for each material claim?
2. Which audience views are genuinely needed, and which have become copies by habit?
3. Who verifies facts, numbers, tone and confidentiality before release?
4. What information is prohibited from an assisted drafting tool?
5. Will the pilot remove a duplicate step within ninety days?

FAQ

Can AI infer outcomes from notes?

No. It may flag a possible pattern for a human to investigate, but it must not turn sparse notes into an outcome claim. The evidence owner decides what can be reported.

What can we change without buying software?

Create a one-page evidence spine, one controlled template per audience and a correction log. These reveal whether a later tool would remove work or merely decorate it.

References

1. David Saliba, A Practical AI Roadmap for NGO Leadership Teams, AIMonger (2026). https://aimonger.com/whitepapers/ngo-practical-ai-roadmap-leadership/
2. Charity Digital Skills Report 2026 (807 respondents; launched 9 July 2026). https://charitydigitalskills.co.uk/report/
3. 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
4. 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
5. NIST, AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
6. Information Commissioner’s Office, Data protection principles. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-protection-principles/
7. HCPC, Our expectations for your record keeping. https://www.hcpc-uk.org/standards/meeting-our-standards/record-keeping/our-expectations-for-your-record-keeping/
8. Regulation (EU) 2016/679 (GDPR). https://eur-lex.europa.eu/eli/reg/2016/679/oj
9. Gartner, “Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes” (2026). https://www.gartner.com/en/articles/genai-project-failure
10. 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