EU AI Act for Nonprofits, NGOs and Charities

Probably minimal-risk — and most of your stack carries no mandatory obligations at all. Fundraising analytics, donor segmentation, the grant-writing assistant, the website chatbot, the back-office automation: none of these are listed in Annex III of Regulation (EU) 2024/1689 — the EU AI Act — and none are prohibited under Article 5. So the heavy high-risk regime (risk management, technical documentation, conformity assessment) does not attach to the ordinary nonprofit toolkit. Your real exposure is narrow and specific: beneficiary-eligibility scoring that can become high-risk in one defined situation, hiring AI under Annex III point 4, the Article 4 literacy duty that applies to every organisation, and Article 50 disclosure when you run a chatbot.

The mistake budget-light teams make is over-reacting — building a conformity programme a charity does not need — or under-reacting and missing the one or two surfaces that do bite. This guide does the classification by use, names the duties that actually apply, and keeps the response proportionate. Start with how minimal-risk classification works, because that is where most of your systems land.

---

The Act applies to nonprofits — there is no charity exemption

There is a persistent hope that the EU AI Act spares the third sector. It does not. The Act binds providers and deployers of AI systems regardless of legal form or profit motive — a registered charity, a foundation, an NGO and a social enterprise are all deployers when they use an AI system in the course of their activity. What changes the obligation is not your tax status; it is what the system does.

That is the good news, framed correctly. Because nonprofits rarely build or operate the use cases Annex III lists, the practical burden is usually small. You inherit a handful of light, horizontal duties and — in two specific situations — one heavier one. Classify by use, document the reasoning, and you are done.

Where you sit in the value chain

Almost every nonprofit is a deployer (an organisation using an AI system under its own authority), not a provider (the organisation that develops a system and places it on the market). Deployer duties under Article 26 are materially lighter than provider duties under Article 16 — and they only engage in full for high-risk systems. For the minimal-risk tooling that makes up most of a charity's estate, Article 26 does not impose a conformity programme at all. If you are unsure which role you hold, provider versus deployer draws the line.

---

What most of your stack is: minimal-risk

Run the inventory and the pattern is consistent. The following are minimal-risk — no mandatory AI Act obligations beyond the horizontal literacy duty:

• Fundraising and donor analytics — segmentation, propensity-to-give scoring, campaign optimisation. These target organisational behaviour and marketing, not an Annex III use case.
• Communications and content — a generative assistant drafting newsletters, social posts, grant applications and reports.
• Operations and back office — scheduling, expense categorisation, document summarisation, translation.
• Programme analytics — dashboards measuring reach and impact, where no automated decision is made about an individual's access to a service.

None of these appear in Annex III, and none are caught by Article 5. The GDPR still governs any personal data you process — donor profiling needs a lawful basis and transparency under GDPR — but that is a parallel regime, not an AI Act high-risk duty. Minimal-risk status switches off no GDPR obligation, and for many charities GDPR is the heavier of the two.

---

The one place beneficiary scoring becomes high-risk: Annex III point 5(a)

Here is the surface that matters most, and it is narrow. Annex III point 5(a) makes high-risk an AI system intended to be used by, or on behalf of, public authorities to evaluate the eligibility of natural persons for essential public assistance benefits and services — including healthcare — and to grant, reduce, revoke or reclaim such benefits.

The trigger is the public-service-delivery context, not the act of scoring in the abstract. A great deal of nonprofit work is delivered on behalf of a public authority: a charity running a council-commissioned homelessness service, a foundation administering a government-funded hardship fund, an NGO assessing eligibility for a publicly financed support programme. If an AI system decides or materially shapes who gets an essential benefit or service in that delegated-public-function setting, it can fall within Annex III point 5(a) and become high-risk.

When it does not trigger

Most beneficiary decisions a charity makes do not hit 5(a):

• Privately funded support, decided on the charity's own authority with no public-authority delegation, sits outside the public-assistance wording.
• A system that merely flags or ranks applications for a human caseworker who makes the real decision can be exempt under the Article 6(3) filter — see below.
• Impact measurement, waitlist management and outreach targeting that allocate organisational effort rather than an individual's entitlement are not eligibility determinations.

The line is whether an AI system determines a natural person's access to an essential public benefit or service on a public authority's behalf. Where it does, treat it as high-risk; where it does not, document why and move on. Social-benefit eligibility AI covers this use case in detail.

The Article 6(3) filter — your most useful tool

Even an Annex III system is not high-risk if it poses no significant risk of harm because it does only a narrow procedural task, improves a previously completed human activity, detects decision patterns without replacing or influencing the human assessment, or performs preparatory work. Only one of the four conditions is needed. So a triage tool that surfaces applications for a human to assess — without making or steering the decision — can be filtered out.

The hard limit: any system that profiles natural persons is always high-risk, with no Article 6(3) escape. If your eligibility tool builds a behavioural or risk profile of applicants, the filter is closed to you. Document the assessment either way, and where the system stays high-risk, register it under Article 49.

---

Hiring AI: Annex III point 4

The second high-risk surface is the same one every employer faces. Annex III point 4 makes high-risk AI used for recruitment and selection (point 4(a) — targeting job ads, screening or filtering applications, evaluating candidates) and for employment decisions in the working relationship (point 4(b) — promotion, termination, task allocation, performance evaluation based on individual behaviour).

If your charity uses an AI CV-screener, an automated video-interview scorer or an algorithmic ranking tool to hire staff or volunteers in an employment-like relationship, that system is high-risk on the same footing as any company's. The Article 6(3) filter can apply — a tool that only parses a CV into fields, or de-duplicates applications, may qualify as a narrow procedural task — but profiling candidates closes it. See recruitment-screening AI and HR and recruitment AI compliance for the full obligation set.

A practical note: most charities recruit at low volume and use a human-led process, so a high-risk hiring system is the exception, not the norm. Inventory it; do not assume it.

---

The two horizontal duties every nonprofit carries

Two obligations apply to your minimal-risk estate regardless of tier.

Article 4 — AI literacy

Article 4 requires providers and deployers to ensure a sufficient level of AI literacy among staff and others operating AI systems on their behalf, taking account of their technical knowledge, experience and the context of use. It has been in force since 2 February 2025 and applies to all AI, including the minimal-risk fundraising and admin tools. For a nonprofit this is light: a short, documented briefing on what your tools do, their limits, when a human must check the output, and basic data-protection awareness. It is a competence-and-records duty, not a certification. AI literacy requirements covers what "sufficient" looks like.

Article 50 — transparency for chatbots and synthetic content

If you run a chatbot that interacts with supporters, beneficiaries or the public, Article 50(1) requires you to tell people they are dealing with an AI system unless that is obvious. A short label at the start of the conversation satisfies it. If you publish AI-generated images, video or audio in campaigns, Article 50(2) marking (provider duty) and Article 50(4) deepfake disclosure (deployer duty) apply; the content-marking obligations land on a fixed date of 2 December 2026. Most Article 50 duties apply from 2 August 2026. None of this carries a technical-documentation or conformity burden — it is disclosure. See Article 50 transparency obligations.

---

Worked example: a mid-sized NGO classifies its stack

Consider Brückenhilfe e.V., a German social-services NGO with about 90 employees and roughly €6M in annual income, of which a third is council-commissioned. It runs five AI systems: a donor-analytics tool, a generative grant-writing assistant, a website chatbot, an AI CV-screener for hiring caseworkers, and an eligibility-scoring tool used inside a council-commissioned hardship fund to decide which applicants receive emergency grants.

Walking each system to a tier

The donor-analytics tool and the grant-writing assistant are minimal-risk — not in Annex III; GDPR governs the donor data. The website chatbot is limited-risk, carrying the Article 50(1) disclosure. The CV-screener is high-risk under Annex III point 4(a) — it screens and ranks candidates, which closes the Article 6(3) filter through profiling. The hardship-fund eligibility tool is the one that stops the review: because Brückenhilfe administers the fund on behalf of the council and the tool determines access to an essential public benefit, it falls within Annex III point 5(a) and is high-risk — and as it profiles applicants by risk, the Article 6(3) filter is unavailable.

The proportionate outcome

Of five systems, two are high-risk (CV-screener, hardship-fund tool) and three are not. Brückenhilfe's AI Act workload is therefore: the Article 4 literacy briefing across all staff; the Article 50(1) chatbot label; and a full high-risk programme — risk management (Article 9), data governance (Article 10), technical documentation (Article 11), human oversight (Article 14), and Article 49 registration — for the two high-risk systems only. Crucially, as an SME-scale organisation, any fine is capped under Article 99(6) at the lower of the percentage or the fixed amount. The privately funded grants the charity decides on its own authority stay outside 5(a) — a distinction worth documenting, because it keeps the high-risk scope to exactly two systems instead of the whole programme.

---

The nonprofit use-case-to-risk map

The only routinely high-risk rows are the public-delegated eligibility tool and the hiring tool. Everything else is minimal- or limited-risk plus GDPR.

The penalty tiers, with the SME cap

Article 5 breaches sit at the top — up to €35,000,000 or 7% of total worldwide annual turnover, whichever is higher, under Article 99(3). Most other breaches, including high-risk and Article 50 duties, are up to €15,000,000 or 3% under Article 99(4). Supplying incorrect or misleading information to authorities is up to €7,500,000 or 1% under Article 99(5). For SME-scale and start-up organisations — which most charities are by turnover — Article 99(6) caps the fine at the lower of the two figures. See the Article 99 penalty regime.

---

Deadlines and how Confir helps

What applies, and when

Two duties already bite: Article 5 prohibitions and Article 4 AI literacy have been in force since 2 February 2025. Most Article 50 transparency duties apply from 2 August 2026, with content-marking and watermarking from 2 December 2026.

For your high-risk systems — the public-delegated eligibility tool and any AI hiring tool — the stand-alone Annex III obligations under Article 6(2) apply from 2 December 2027. That date is set by the Digital Omnibus, which is adopted: the European Parliament passed it in plenary on 16 June 2026 and the Council formally adopted it on 29 June 2026. Only publication in the Official Journal remains — expected before 2 August 2026 — a formality that does not affect the dates. The deferral is to fixed calendar dates; the standards-contingent "stop the clock" variant was rejected. The 2 December 2027 date is now the settled deadline for stand-alone Annex III high-risk obligations. See is the AI Act delayed.

What Confir does

Confir registers each AI system a charity deploys and classifies it under Article 5 and Article 6 with Annex III logic from a plain-English intake — separating the minimal-risk fundraising and admin tooling from the one or two systems (public-delegated eligibility, AI hiring) that genuinely reach high-risk. It runs the Article 6(3) filter as an explicit, documented step, derives the deployer role under Article 26, and generates the Article 11 / Annex IV documentation and Article 49 registration only where they are actually owed.

The engine is deterministic and rule-based: the same intake answers produce the same finding, with the rule that fired shown in plain English — reproducible and audit-defensible, which matters for a budget-light team that needs to prove a proportionate decision without a consultant. No model inference, no guesswork.

---

Frequently asked questions

Does the EU AI Act apply to nonprofits and charities?

Yes. The EU AI Act binds providers and deployers of AI systems regardless of legal form or profit motive, so registered charities, foundations and NGOs are in scope when they use AI. There is no charity exemption. What keeps the burden small is that classification is by use: most nonprofit tooling — fundraising, communications, operations — is minimal-risk and carries no mandatory obligations beyond the Article 4 AI-literacy duty. The heavier high-risk regime only engages where a specific Annex III use case is present.

Is beneficiary-eligibility scoring high-risk under the EU AI Act?

It can be, in one defined situation. Annex III point 5(a) makes high-risk an AI system used by or on behalf of a public authority to evaluate eligibility for essential public assistance benefits or services. A charity administering a council-commissioned or government-funded programme that uses AI to decide who receives an essential benefit can fall within it. Privately funded support decided on the charity's own authority generally sits outside 5(a). A tool that only flags applications for a human caseworker may be exempt under the Article 6(3) filter — unless it profiles applicants, which is always high-risk.

Is AI used for hiring at a nonprofit high-risk?

Generally yes, if it screens, ranks or evaluates candidates. Annex III point 4 makes recruitment and selection AI high-risk on the same footing for a charity as for any employer, covering CV-screening, automated interview scoring and candidate ranking. The Article 6(3) filter can exempt a narrow procedural tool — for example, one that only parses a CV into fields — but profiling candidates closes it. Most charities recruit at low volume with a human-led process, so a high-risk hiring system is the exception rather than the norm; inventory it before assuming.

What is the AI literacy duty for a charity?

Article 4 requires deployers to ensure a sufficient level of AI literacy among staff operating AI systems, in force since 2 February 2025 and applying to all AI, including minimal-risk fundraising and admin tools. For a nonprofit this is light: a short, documented briefing covering what each tool does, its limits, when a human must check the output, and basic data-protection awareness. It is a competence-and-records duty, not a certification, and it scales to the size and risk of the organisation.

Do we need to label our charity chatbot as AI?

Yes, in almost all cases. Article 50(1) requires you to tell people they are interacting with an AI system unless that is obvious to a reasonable user. A short label at the start of the conversation satisfies it. If you publish AI-generated images, audio or video in campaigns, Article 50(2) marking and Article 50(4) deepfake disclosure apply, with content-marking duties landing on 2 December 2026. None of this carries a technical-documentation or conformity-assessment burden — it is disclosure, not a high-risk programme.

What are the EU AI Act deadlines a nonprofit should plan against?

Article 5 prohibitions and Article 4 AI literacy have been in force since 2 February 2025. Most Article 50 transparency duties apply from 2 August 2026, with content-marking from 2 December 2026. For any high-risk system you deploy — a public-delegated eligibility tool or AI hiring — stand-alone Annex III obligations originally read 2 August 2026 in the statute; the deferral to 2 December 2027, adopted via the Digital Omnibus in June 2026, is now the settled deadline.

---

Related guides

• How minimal-risk classification works
• Social-benefit eligibility AI under Annex III
• Recruitment and selection AI under Annex III
• Article 4 AI literacy requirements
• Article 50 transparency obligations
• Provider versus deployer roles
• Public sector AI compliance