The AI Due Diligence Checklist: What to Run Before You Buy or Integrate Third-Party AI

A recruitment agency buys a SaaS tool that auto-ranks candidate CVs. Run no diligence and the agency discovers, after signing, that it has taken on a high-risk AI system under Annex III point 4 with no conformity evidence, no built-in human oversight, and a contract that flows none of the provider's obligations down to it. The fix is cheap before the signature and expensive after. AI due diligence is the buyer-side workflow that surfaces all of that before money changes hands.

This page sets out the four-phase runbook a buyer should run before purchasing or integrating any third-party AI system or AI feature under Regulation (EU) 2024/1689: triage how deep to go, map who carries which duty, demand the documents that prove compliance, and close out a defensible go/no-go at the procurement gate. It is a starting framework to adapt to a specific buy, not legal advice.

---

What an AI Due Diligence Checklist Is (and What It Sits Between)

AI due diligence is the sequenced set of buyer-side activities that turn a purchase intent into a defensible go/no-go decision. It is not the vendor's answer sheet. The vendor fills in a questionnaire; you run diligence on what comes back. The two are not the same artefact, and conflating them is how buyers end up trusting an attestation that no document supports.

Due diligence is the buyer's pre-purchase workflow, not the vendor's answer sheet

Think of three instruments and one runbook. The AI vendor questionnaire that collects the evidence gathers the vendor's answers. The scoring matrix that helps you score the vendor's answers into a go/no-go rating rates them. The contract locks remediation in. Due diligence is the ordered runbook that sequences all three plus the steps only the buyer performs — triage, exposure mapping, and sign-off. It wraps the others; it does not duplicate them.

Who runs it: procurement, legal, security and the AI governance owner

Diligence is cross-functional. Procurement owns the gate and the timeline. Legal owns the role analysis and the contract red lines. Security owns logging and data-flow verification. The AI governance owner — whoever holds the AI inventory — owns the classification finding and files the completed checklist as the record. One of them has to own the go/no-go call, or it never gets made.

"Vendor" is not a legal category

Regulation (EU) 2024/1689 binds providers (Article 16), deployers (Article 26), importers (Article 23) and distributors (Article 24). "Vendor" appears nowhere as a duty-holder. Running diligence is how a prospective deployer discharges its own Article 26 obligations before deployment — verifying the provider did what the law requires, rather than assuming it.

---

Phase 1 — Triage: Decide How Deep the Diligence Needs to Go

Most buys need a light touch. A few need the full file. Triage decides which — and it keys off intended use, not the price tag. A €200/month tool that screens job applicants warrants more diligence than a six-figure analytics platform that touches no protected decision.

Is there actually AI in the product, and is it material?

First question: is there AI in the product at all, and does it materially inform a decision? A spell-checker is AI and irrelevant. A model that ranks loan applicants is AI and decisive. Triage depth tracks materiality.

First-pass risk screen — Article 5 first

Screen Article 5 before anything else. If the use is a prohibited practice — social scoring, untargeted facial-image scraping, emotion recognition at work or in education, certain biometric categorisation — diligence stops at an automatic no-buy. These prohibitions have applied since 2 February 2025. No conformity file rescues a prohibited use.

Then run a first-pass Annex III screen across the eight high-risk domains, judged on your intended use, not the vendor's marketing label. A high Annex III screen escalates to full diligence; a clear minimal-risk tool gets a proportionate light review.

Record every triaged tool, even the low-risk ones

Link the depth decision to the AI inventory so every triaged tool is recorded — even when it clears as low-risk. A register with gaps is a register an inspector does not trust. Confir's deterministic, rule-based engine produces the same triage finding from the same intake every time, so the depth decision is reproducible rather than a matter of who ran it.

---

Phase 2 — Map Roles and Exposure: Who Carries Which Duty, and the Article 25 Flip Risk

The headline question this phase answers: who is legally on the hook, and could buying this tool quietly turn us into the provider? The questionnaire and scoring matrix touch role allocation only in passing. Diligence has to settle it.

Confirm the provider/deployer split

Provider obligations sit under Article 16; deployer obligations under Article 26. The vendor placing the system on the market is normally the provider. You, putting it into use under your authority, are normally the deployer. If you cannot say with confidence which role you occupy, resolve it before you score anything — every later step depends on the answer.

The Article 25 provider-flip — the trap that reclassifies the buyer

Article 25 is the diligence trap that catches buyers. You become the provider — inheriting the full Article 16 stack — if you put your own name or trademark on a high-risk system, make a substantial modification under Article 3(23), or change the intended purpose so the system becomes high-risk. The transfer is total: conformity assessment, Annex IV documentation, CE marking, EU database registration. A flip does not require retraining. Rebranding the tool as your own or extending it to a new use case can cross the line on its own.

Non-EU providers and GPAI in the stack

If the provider is established outside the EU, check for an authorised representative under Article 22 — a non-EU provider of a high-risk system must appoint one. Where the tool is built on a foundation model, identify the upstream GPAI model provider and whether systemic-risk applies (the Article 51 rebuttable presumption bites at 10^25 FLOP training compute). GPAI provider obligations under Articles 51–55 have been in force since 2 August 2025. If you are unsure whether the buy makes you a deployer or quietly a provider, work out whether you are the provider or the deployer before scoring, and find shadow AI staff adopted without procurement so diligence covers tools already in the building.

---

Phase 3 — The Documentation You Should Demand Before You Sign

This is where attestations meet evidence. Ask for the document, not the claim. "We are compliant" carries no weight without the file behind it. For a high-risk system, the evidence pack below is what you demand and what each item proves.

Transparency, logging and oversight you can verify

Confirm Article 50 disclosure duties where the tool is a chatbot, an emotion- or biometric-categorisation system, or generates synthetic content. The content-marking and watermarking obligation under Article 50(2) lands 2 December 2026. Verify that Article 12 automatic logging exists and is accessible to you, and that Article 14 human-oversight measures are genuinely built into the architecture. Oversight "on paper" that the system cannot actually support is a red flag, not a checkbox.

The hard stop, and proportionate diligence below high-risk

If the provider cannot produce any Annex IV documentation or evidence conformity for a high-risk system, that is a disqualifier, not a follow-up item. Treat refusal to share or even summarise as a breach of the provider's Article 13 transparency duty. For minimal- and limited-risk tools, a proportionate subset — instructions, data-processing terms, security posture — is usually enough. Diligence depth follows the Phase 1 triage.

---

Phase 4 — Procurement Gate and Contract Red Lines

Diligence is worthless until it closes into a decision and the decision lands in the contract. This is the procurement-action layer the questionnaire and the scoring matrix stop short of.

Turn findings into go / conditional / no-go

Convert the file to one of three outcomes:

• GO — proceed and file the diligence record.
• CONDITIONAL — proceed only with contractual remediation of named gaps.
• NO-GO — triggered by any hard disqualifier: an Article 5 prohibition, missing conformity for a high-risk system, no built-in oversight, or unreported serious incidents.

Contract red lines every AI buy should carry

Four clauses earn their place in any AI contract. First, the vendor warrants its provider status and its Article 16 obligations for the system as supplied. Second, the vendor gives advance notice before any change that could trigger an Article 25 flip or a re-conformity assessment under Article 43(4). Third, Article 73 incident-notification timelines flow down to you — reports run within 15 days of awareness, 2 days for widespread infringement or critical-infrastructure disruption, and 10 days where a death has occurred. Fourth, you hold audit rights over the Annex IV summary after a significant change.

Tie the decision to penalty exposure so it has teeth

Deployer and most-obligation breaches carry fines up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)). Prohibited-practice breaches reach €35,000,000 or 7% (Article 99(3)). Supplying incorrect or misleading information to authorities runs to €7,500,000 or 1% (Article 99(5)). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the fixed amount or the percentage. Schedule re-diligence at renewal and on any material change. File the completed checklist as the procurement-gate artefact and seed it into the AI register. The instruments this gate runs on are the AI vendor questionnaire that collects the evidence and the deployer vendor-assessment framework.

---

Worked Example: Running the Checklist on a Recruitment-Screening Vendor

Meridian Talent GmbH, a 280-person recruitment agency in Munich, is buying "ScreenFlow", a SaaS tool that auto-ranks candidate CVs, from a US-based vendor. Here is the checklist applied end to end.

Phase 1 — triage. Automated CV ranking influences hiring shortlists. That is squarely within Annex III employment use (point 4), and because the tool profiles natural persons, the Article 6(3) exemption cannot apply. Meridian escalates to full diligence. Full stop.

Phase 2 — roles. The vendor is the provider (Article 16); Meridian is the deployer (Article 26). Because the vendor is non-EU, Meridian checks for an Article 22 authorised representative. Meridian confirms it will not rebrand or modify ScreenFlow, so no Article 25 flip — but it adds a pre-change notice clause anyway. Deploying in the workplace triggers the Article 26(7) duty to inform workers' representatives before use.

Phase 3 — documentation. The vendor supplies Article 13 instructions and an Annex IV summary, but provides no EU database registration ID and thin bias-testing disclosure. Because conformity evidence does exist, this is a conditional, not a disqualifier.

Phase 4 — gate. Outcome: CONDITIONAL. Before signature, the contract must add the registration ID, disaggregated performance metrics, Article 73 incident flow-down, and pre-change notice. Meridian files the completed checklist and seeds the record into its AI inventory. The deadline note (below) shaped Meridian's timing — it dated the diligence against the adopted 2 December 2027 deferral date.

---

When the Obligations Bite: Deadlines That Shape Diligence Timing

By when must this diligence be done, and which deadlines can a buyer actually rely on? Diligence timing depends on the answer, and not everything is delayed.

What is already in force

Fixed and live: Article 5 prohibitions and Article 4 AI literacy since 2 February 2025; GPAI obligations (Articles 51–55) since 2 August 2025. A new 2 December 2026 deadline adds the CSAM and "nudifier" prohibition plus the Article 50 content-marking and watermarking obligation.

What the Digital Omnibus settled

Stand-alone high-risk Annex III (Article 6(2)): the deferral to 2 December 2027 is now adopted. Product-embedded Annex I high-risk (Article 6(1)): the deferral to 2 August 2028 is likewise adopted.

The status in full: the Digital Omnibus is adopted — the European Parliament passed it on 16 June 2026 and the Council on 29 June 2026. It enters into force on publication in the Official Journal, expected before 2 August 2026, a formality that does not change the dates. The deferral is fixed calendar dates; the standards-contingent "stop the clock" variant was rejected.

Plan against the adopted dates

For buyers, the practical takeaway is to plan and date your diligence against 2 December 2027 for stand-alone high-risk and 2 August 2028 for product-embedded high-risk, now that the Digital Omnibus deferral is adopted.

---

How Confir Helps

Confir runs the diligence runbook as a sequenced workflow rather than a static form. The triage screen records each tool against Article 5 and the Annex III domains; role mapping flags the Article 25 flip risk before you score; the documentation phase tracks each demanded artefact against its Article and marks gaps for the procurement gate. The engine is deterministic and rule-based — no model inference, no hallucination — so the same intake always produces the same finding, and an inspector can see exactly why a tool cleared, was made conditional, or was rejected. Completed checklists seed straight into your organisation-wide AI register.

---

Frequently Asked Questions

What is an AI due diligence checklist?

It is the buyer's ordered runbook for vetting a third-party AI system or AI feature before you buy or integrate it. Under the EU AI Act it sequences four phases: triage how deep the review needs to go, map who is the provider and who is the deployer (Article 16 versus Article 26), demand the documentation that proves compliance, and close out a go, conditional or no-go decision with contract terms. It is the workflow that wraps a vendor questionnaire and a scoring matrix, plus the buyer-only steps neither tool covers.

How is AI due diligence different from a vendor questionnaire or risk assessment?

A questionnaire collects the vendor's answers; a risk assessment weights those answers into a numeric score. Due diligence is the broader procurement workflow that sequences both, then adds the steps only the buyer performs: triaging review depth, mapping the Article 25 provider-flip risk, demanding and reviewing documents, and converting findings into a contract. Treat the questionnaire and scoring matrix as instruments the diligence checklist tells you when and how to use, ending in a defensible, inspectable purchase decision.

What documents should you ask an AI vendor for during due diligence?

For a high-risk system, request the instructions for use (Article 13), an Annex IV technical documentation summary (Article 11), the EU Declaration of Conformity (Article 47), evidence of CE marking (Article 48), the EU database registration identifier (Article 49), and the conformity assessment record (Article 43). Ask for the documents themselves, not attestations — a vendor that cannot produce any Annex IV material for a high-risk system is a disqualifier. For limited- or minimal-risk tools, a proportionate subset is usually sufficient.

Can buying or integrating an AI tool make my company the provider?

Yes. Under Article 25, if you put your own name or trademark on a high-risk system, make a substantial modification as defined in Article 3(23), or change the system's intended purpose so it becomes high-risk, you are reclassified from deployer to provider and inherit the full Article 16 obligation set — conformity assessment, Annex IV documentation, CE marking and EU database registration. This can happen without any retraining; rebranding or extending scope to a new use case can be enough. Diligence should flag this, and the contract should require advance notice of triggering changes.

What are the penalties for deploying a non-compliant AI vendor's tool?

A deployer's own exposure is separate from the provider's. High-risk and most obligation breaches carry fines up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)). Using a prohibited Article 5 practice carries up to €35,000,000 or 7% (Article 99(3)), and giving incorrect or misleading information to authorities up to €7,500,000 or 1% (Article 99(5)). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the two amounts. The vendor's non-compliance does not insulate you.

When does AI due diligence need to be completed under the EU AI Act?

It depends on the obligation. Article 5 prohibitions and Article 4 AI literacy have applied since 2 February 2025, and GPAI obligations (Articles 51–55) since 2 August 2025. For stand-alone high-risk Annex III systems, the Digital Omnibus deferral to 2 December 2027 is now adopted — the European Parliament passed it on 16 June 2026 and the Council on 29 June 2026 — pending only Official Journal publication, expected before 2 August 2026. Plan your diligence against 2 December 2027.

Does an AI due diligence checklist apply to AI features inside ordinary software?

Yes. An AI feature is assessed on the function it performs, not on how peripheral it is to the wider product. Automated CV screening inside an HR platform is an Annex III point 4 high-risk use whether it is a standalone product or a bundled feature, and your deployer duties under Article 26 apply regardless of packaging. During triage, the test is whether there is AI in the product and whether it materially informs a decision — if a vendor cannot tell you whether their AI feature has been classified, that is itself a finding worth recording.

---

Related guides

• work out whether you are the provider or the deployer
• the AI vendor questionnaire that collects the evidence
• score the vendor's answers into a go/no-go rating
• the deployer vendor-assessment framework
• find shadow AI staff adopted without procurement