← Back to the Lab
/ ESSAY·10 MIN READ·LONG-FORM
/ LONG-FORM

EU AI Act Compliance Checklist: What Funds Ask Before Close

The EU AI Act diligence checklist a GP can run on a portfolio company before closing — risk tier, logging, FRIA triggers, and real deadlines.

EU AI Act Compliance Checklist: What Funds Ask Before Close
/ TL;DR

The EU AI Act diligence checklist a GP can run on a portfolio company before closing — risk tier, logging, FRIA triggers, and real deadlines.

Here is the short version: once you have signed, an AI Act problem in a portfolio company is your problem too. So this is the checklist I would run on a target before closing — six passes that surface the obligations, the live deadlines, and the one finding that changes a deal outright. I build AI systems for funds, and the diligence conversation now routinely opens here. The checklist below is built from the same obligations and deadlines our EU AI Compliant knowledge base tracks against the Official Journal — not a generic regulatory-timeline recap.

Use it as a question set, not a lecture. Every step is something you can hand to a GP or an associate and get a usable answer back before signing.

IWhy this belongs in diligence now, not after close

Start with the money, because that is the part a partner will actually feel. An internal AI Act readiness assessment, done ahead of diligence, runs roughly €10,000–50,000. A third-party conformity assessment that a company is forced into later — because it walked into diligence unprepared and the gap surfaced after the deal — runs €50,000–200,000, according to reporting collected by Pebblous. Same work, up to four times the price, now sitting on the books of a company you own.

That asymmetry is the whole argument for pulling this check forward. And you are not early to it — you are catching up. Preply's Head of Legal, Ivanna Honina, has said European investors have started checking AI Act conformity from the earliest stages of a deal. A monthly VC-focused legal briefing, Ashurst Perkins Coie's Venture Matters (produced with Deutsche Börse), goes further: it recommends funds run an AI inventory and risk classification on portfolio companies, document the governance measures behind any high-risk system, budget for conformity assessments, and write AI Act compliance directly into diligence checklists and transaction documents.

If you want the regulatory backdrop first, I wrote a fuller piece on what changed on schedule and what the Digital Omnibus actually delayed. This checklist assumes you already know the shape of the law and want the fund-specific questions.

IIThe two dates that actually matter after the Digital Omnibus

The single most common mistake I see is treating the AI Act as one far-off deadline. It is not. The Digital Omnibus on AI (Regulation (EU) 2026/1744), which entered into force on 27 July 2026, moved the high-risk regime back — and nothing else. The transparency rules it did not touch are already live.

So in diligence you are really checking against two clocks. One is already ticking (transparency and general-purpose AI). The other is the delayed high-risk deadline, which matters precisely because the infrastructure behind it takes months to stand up.

The European Commission is clear that Article 50 reaches any AI system placed on the EU market, or whose output is used in the EU, regardless of where the provider or deployer sits. A US-incorporated startup with EU users is not outside the net.

IIIStep 1 — Is the target a provider, a deployer, or both?

Ask this first, because it decides which half of the law applies. A provider builds, trains, or substantially develops the AI system. A deployer uses someone else's. The obligation sets are different, and — as investor-side diligence guidance notes — many startups genuinely have not worked out which they are, or realise they are both.

Get them to say it plainly: Are you building and training models, or calling someone else's API? Then watch for the Article 25 trap. Under Article 25, a deployer can be reclassified as a provider — and inherit the full provider burden, including conformity assessment — in three situations:

  • They put their own name or trademark on a high-risk system already on the market.
  • They make a substantial modification to a high-risk system that stays high-risk.
  • They change the intended purpose of a system (including a general-purpose AI system) so that it becomes high-risk.

That third one catches a lot of "we just wrap GPT" companies. The moment they repurpose a general model into, say, automated candidate screening, the compliance story is no longer the model vendor's — it is theirs, and therefore yours.

IVStep 2 — Run the AI system inventory and risk-tier it

This is the pass that produces the actual artefact. List every AI system and feature the target runs — including embedded third-party models — what each one does, and which tier it lands in: unacceptable (Article 5), high-risk (Annex III or Annex I), limited-risk (transparency only), or minimal. Venture Matters treats this inventory-and-classification as the baseline step, and I agree: without it, every later question is guesswork.

For each system, capture:

  • What it does, and who it affects (users, workers, applicants, patients).
  • Whether it touches a regulated sector — employment, credit, insurance, education, law enforcement, migration, critical infrastructure.
  • Whether it does biometric processing or makes consequential decisions about people. Those are the features that most often pull a system into high-risk.

One result on this pass is not a condition to negotiate — it is a stop sign. Unacceptable-risk practices are banned outright under Article 5: social scoring, real-time public biometric surveillance (narrow exceptions aside), subliminal manipulation, and exploitation of vulnerable groups. A private-equity compliance guide from Raphael Nagel puts it bluntly: if diligence surfaces one of these, "the acquisition conversation changes entirely." This is the single finding that can kill a deal rather than just add a covenant.

VStep 3 — For anything high-risk: the Article 26 questions to ask

If the inventory turns up a high-risk system, Article 26 is your question bank. Turn each deployer obligation into a direct ask:

  • Do you use the system strictly per the provider's instructions for use?
  • Who holds human-oversight authority — and are they actually trained and empowered to intervene, not just named on a slide?
  • Where you control the input data, is it relevant and representative for the system's purpose?
  • Do you monitor operation and have a route to flag risks back to the provider?
  • Do you retain the system's logs for at least six months?
  • Before any workplace deployment, did you inform workers' representatives and affected workers?
  • Do you have a serious-incident reporting process, and do you tell individuals when a consequential decision was AI-assisted?

Here is the timing nuance that matters for diligence. These duties only bite from 2 December 2027 for standalone Annex III systems (2 August 2028 for Annex I). But the logging pipeline, the oversight roles, and the governance paperwork take months to build. So you are not asking "are you compliant today" — you are asking "what is your current state, and is there a credible path to standing this up well before the deadline." A company with no logging and no named human overseer is not breaking the law yet; it is telling you how much work is still unfunded.

VIStep 4 — Does a Fundamental Rights Impact Assessment apply?

This one is narrow, so ask it as a targeted question, not a blanket one. Article 27 requires a Fundamental Rights Impact Assessment (FRIA) — completed and registered in the EU database before deployment — from a specific group: public-sector deployers, private operators providing public-facing services, and deployers of high-risk credit-scoring or life- and health-insurance-pricing systems (Annex III points 5(b) and 5(c)).

That is a far smaller set than "all high-risk deployers." But if your target is a fintech doing credit scoring, or an insurtech pricing policies, this is squarely on point — and a missing FRIA is a concrete, registrable gap you can price. For most other high-risk companies, the honest answer is that Article 27 does not reach them, and you can move on.

VIIStep 5 — Transparency: what's already live under Article 50

Most diligence checklists I see are still fixated on the delayed high-risk dates and skip this one. That is the mistake, because Article 50 is already in force — and a gap here is a live violation, not a future risk.

Run the transparency pass on anything user-facing:

  • Does the product disclose AI-generated content and deepfakes?
  • Does it flag to users when they are interacting with a chatbot or AI, rather than a person?
  • Does it give notice when emotion-recognition or biometric-categorisation features are in use?

Watch the provider/deployer split one more time. Morgan Lewis notes that the only breathing room the law gives is a four-month transition — to 2 December 2026 — for the provider-side machine-readable marking of generative AI that was already on the market before 2 August 2026. The deployer duties — deepfake disclosure, emotion-recognition and biometric notices, public-interest AI-text disclosure — have applied with no grace period since 2 August 2026. If your target is a deployer and the labels are not there, that is an exposure today.

VIIIStep 6 — GPAI exposure, if they're building on or fine-tuning a model

The last pass is for anyone touching foundation models. Ask whether the target trains or fine-tunes its own general-purpose model — which puts provider-side GPAI obligations on them, and a systemic-risk threshold in play for the largest models — or whether it purely calls third-party APIs. If it is the latter, the question flips: ask them to produce the provider's model documentation, because that is what they will need on file either way.

Mind the dates from the DataGuard timeline: GPAI model obligations have applied since 2 August 2025 for models placed on the market from that date, and GPAI models already on the market before 2 August 2025 have until 2 August 2027 to come into compliance. A company quietly running a pre-2025 in-house model has a 2027 clock it may not have started.

IXWhat happens if the answers are bad

Translate the findings into deal terms, because that is the language your IC speaks. Article 99 sets three fine tiers:

Two things funds get wrong here. First, the SME and start-up provisions are not an exemption. For SMEs, each fine is capped at the lower of the percentage or the fixed amount — a smaller ceiling, not a free pass. Venture Matters makes the same point: sandboxes and proportionate caps do not lift high-risk or GPAI obligations regardless of company size. Second, these numbers belong in the deal, not in a compliance footnote. A real transparency gap or a missing conformity assessment is a line item for valuation and for reps and warranties — the same way you would treat an unpaid tax liability or an unassigned patent.

XTurning this into a standing diligence process

Run once, this checklist catches the obvious problems in one target. But a fund doing repeat deal flow does not need a one-off PDF — it needs the obligation and deadline list kept current, because the next Omnibus will move the goalposts again, exactly as this one did. A static checklist goes stale the moment the Official Journal updates.

That is the problem EU AI Compliant exists to solve: a regulatory knowledge base that tracks the EU AI Act — every article, obligation, deadline, and enforcement action — against the Official Journal, so the questions you ask are current rather than last quarter's. And the first-pass version of this check already fits inside the deal workflow — our Diligence Desk runs a first pass from the public record, right inside your AI assistant, so the AI-risk read starts the moment a company lands in your pipeline.

I build this kind of thing for funds as a custom build on a monthly retainer — a studio that helps funds integrate AI into how they actually work, rather than selling another dashboard. The takeaway stands on its own, though: pull the AI Act check before you sign, tier every system, and treat an unacceptable-risk finding as a stop sign and everything else as a number. The fund that asks these six questions early pays €10k–50k to know. The one that asks after close finds out what the other price looks like.

— Michael Rouveure

/ WORKING WITH BLACK MATTER VC

If this was useful,
you should book a call.

$10k / month. Whatever your fund needs, shipped that month. 30-min intro, no deck — I’ll tell you which three systems I’d ship first.

Or follow along on LinkedIn / X.