Compliance due diligence checklists are drafted differently from one institution to the next, but the questions that reach AI companies are remarkably stable. Investors do not ask whether you are compliant; they ask how something was done, what evidence exists, and when it was done. The quality of the answer turns on three things: whether the evidence can be produced on the spot, whether the account is consistent across documents, and whether completed, in-progress and not-started work are kept apart.
1. Where the questions come from
There are usually three questioners, at different depths.
Pre-investment risk control. Run by a fund’s internal risk team or external counsel, focused on whether risk affects valuation and deal structure. Questions cluster around material violations, services launched without filing, and the provenance and ownership of data.
The listing standard. Brokers and lawyers conduct a pre-review against IPO criteria, asking whether anything constitutes a substantive obstacle. Questions reach down to filing numbers, change filings and penalty records.
The acquirer. Working backwards from post-closing liability, the buyer asks who bears historic exposure. Questions land on contract terms, indemnities and representations.
The same fact gets a different answer in each setting. Before investing, “in progress, expected within three months” may be enough. Under the listing standard it usually is not, and acceptance receipts and a timetable are required. Establish which setting you are in before answering.
2. Eighteen questions in five groups
Group one: licensing and filing
| Question | What is examined | Answer that satisfies | Recurring defect |
|---|---|---|---|
| Has algorithm filing been completed? | Entity, number, order of launch and filing | Number verifiable, entity matches the operator, filing before launch | Filing through an affiliate; new functions added without a change filing |
| LLM filing or registration? | Whether the service has public-opinion or mobilisation attributes; whether an overseas model is called | Route decision documented with reasons | A filing was required but only registration was done |
| Is there a value-added telecom licence? | Whether ICP and EDI cover the actual business | Licensed scope matches the revenue model | ICP filing only, while charging on a commercial basis |
| Does the generative AI service meet admission requirements? | Service type, public availability | An admission assessment record and conclusion | Launched without assessment |
Group two: data and personal information
| Question | What is examined | Answer that satisfies | Recurring defect |
|---|---|---|---|
| Where does training data come from? | Proportions and legal bases across self-collected, public, purchased and user data | A provenance list with authorisation documents | An oral assurance that it is “all public data” |
| Was consent obtained for personal information? | Notice text, consent records, separate-consent scenarios | Versioned privacy policy; consent traceable | Only the current policy survives; no consent logs |
| Is user data used for training? | Default state, switch, opt-out | Off by default, or clearly disclosed with an opt-out | Terms reserve improvement rights; no switch exists in practice |
| Has a PIA been carried out? | Four triggers, timing of the report, remediation closed | Report exists; remediation accepted | Done after launch; no remediation record |
| Is there any cross-border transfer? | Recipient, mechanism, separate consent | One of assessment, standard contract or certification in place | Overseas model API called without any assessment |
Group three: intellectual property
| Question | What is examined | Answer that satisfies | Recurring defect |
|---|---|---|---|
| Who owns the model weights and training data? | Proportions and chain across in-house, licensed and open-source | Complete ownership documents; traceable chain | The base model licence restricts commercial use |
| Is open-source licensing compliant? | Licence types, copyleft terms, mode of distribution | A licence inventory and evidence of compliance | AGPL components in a closed-source distribution |
| Are employee inventions and collaborative output allocated? | Inventor agreements, joint-development clauses | Assignment clauses cover models and data | University or external collaboration output left unallocated |
Group four: content and safety
| Question | What is examined | Answer that satisfies | Recurring defect |
|---|---|---|---|
| Is generated content labelled? | Explicit and implicit labels, timing | Labels in the interface and in metadata | A text notice in the interface only, no metadata label |
| What is the content moderation mechanism? | Process, disposition records, human review | Written policy, logs, a review step | Reliance on the model’s own filtering alone |
| How are minors protected? | Age identification, minors mode, function limits | Identification and routing, with an appeal channel | A self-declared adult checkbox at registration |
Group five: litigation, penalties and contracts
| Question | What is examined | Answer that satisfies | Recurring defect |
|---|---|---|---|
| Any regulatory enquiries or penalties? | Timing, cause, remediation | Full disclosure with remediation evidence | Only closed matters disclosed |
| Any pending litigation or complaints? | Type, amount, progress | A schedule with estimated exposure | User complaints and small claims omitted |
| Are data terms in major contracts clear? | Supplier DPAs; customer confidentiality and data terms | Contract schedule with key-term summaries | DPA missing, or silent on training use |
Of the eighteen, three are pressed most often: the provenance of training data, whether user data trains the model, and whether the filing entity matches the operator. Their common feature is that once an answer fails to match the documents, every other answer is re-examined.
3. Three principles for answering
First, give evidence, not conclusions. “We are compliant” carries no information. An effective answer is a filing number, a page reference in a report, a date range that can be exported from logs. Diligence counsel form a view on verifiable material, not on adjectives.
Second, separate completed, in-progress and not-started. Run the three together and the other side will assume nothing is finished. Written as “completed; in progress (expected in month X, acceptance receipt attached); not started (reasons and plan)”, the account is more credible, not less.
Third, keep the account consistent. The privacy policy, the PIA, the DPAs and the filing materials describe one data flow. Four accounts across four documents are harder to explain than a missing document. Run an internal cross-check before diligence begins rather than waiting for the other side to find the discrepancy.
4. Twelve items to have ready
| No. | Item | Purpose |
|---|---|---|
| 1 | Algorithm and LLM filing numbers with public notice screenshots | Evidence of admission |
| 2 | Value-added telecom licence | Evidence of operating authority |
| 3 | Privacy policy versions with effective dates | Evidence of notice and consent |
| 4 | Consent log samples, including separate consent | Evidence that consent is traceable |
| 5 | Personal information protection impact assessment | Evidence that high-risk processing was assessed |
| 6 | DPA schedule and sub-processor list | Evidence that processing relationships are documented |
| 7 | Training data provenance list and authorisations | Evidence of lawful provenance |
| 8 | Open-source component and licence inventory | Evidence of licence compliance |
| 9 | Interface screenshots and metadata samples for labelling | Evidence that labelling duties are met |
| 10 | Content moderation policy and disposition log samples | Evidence that moderation operates |
| 11 | Minors protection plan | Evidence of handling for that scenario |
| 12 | Schedule of regulatory communications, penalties and litigation | Evidence that exposure is disclosed |
None of the twelve needs to be a polished document, but each must be retrievable within one business day of the question. The response that most damages trust in a diligence session is “I will need to ask a colleague on the technical side”.
5. After a problem is found
Finding the problem is not the end of the matter; how it is handled determines the outcome. Three routes, in order of workability.
Remediate. Where it can be completed in the short term, give a firm timetable and an owner, and describe the interim measures. Change filings, re-executed DPAs and supplementary labelling usually fall here.
Disclose. Where it cannot be completed before closing, put it in the disclosure letter with the facts, the scope of impact and the steps taken. Disclosure brings the uncertainty onto the table for both sides to price, rather than leaving the buyer to find it after closing.
Structure. Where the impact is significant, make it a condition to closing, a price adjustment or a specific indemnity. This protects both sides: the seller buys certainty at a contained cost, and the buyer covers the residual risk through the price mechanism.
The worst course is to describe unstarted work as underway. Diligence turns on credibility, and one inaccurate answer invites re-examination of all of them at a cost far above honest disclosure.
Compliance diligence is not an examination. It is how the other side decides whether the risk can be priced. A risk that can be priced is part of the transaction; only a risk that cannot be priced becomes an obstacle.
Related reading: negotiating term sheets for AI companies, the AI product launch compliance document pack, data processing agreements in AI, personal information protection impact assessment, and the financing and equity and data compliance service pages.
Related reading: Eight weeks before a financing round Last updated: 17 September 2026.
Frequently Asked Questions
Which five groups does investor diligence cover?
Licensing and filing, data and personal information, intellectual property, content safety, and litigation and contracts. Eighteen questions across these five cover most follow-ups.
What are the three answering principles?
State only verifiable facts; separate completed, in-progress and not-started; attach an evidence number to each answer. Only all three together make an answer reviewable.
What are the most common weak answers?
'Presumably done', 'refer to the official position', 'the file is with a colleague'. These read as unverifiable and invite a full re-examination.
How should a missing filing be answered?
State the status, plan and owner honestly, with a timetable and interim measures. Do not claim it has been submitted; one inaccurate statement taints every other answer.
What happens after an issue is raised?
Assess the scope of impact and propose remediation; where material, disclose proactively and reflect it in closing conditions, price adjustments or a specific indemnity.
How much material is enough?
Prepare against each of the eighteen questions; about twelve core documents usually cover most follow-ups. Numbering and indexing matter more than volume.