Malaysia AI market, policy & data
AI Vendor Due Diligence in Malaysia: The 12 Questions Legal and Compliance Teams Should Ask Before Signing (With a Scoring Sheet)
· By AIHQ Team

Twelve questions that decide whether an AI vendor contract is safe to sign
A department head signs a twelve-month AI contract in March. Six months later, the vendor quietly swaps the model behind the assistant, the monthly fee triples under a usage clause nobody read, and the training data your team uploaded is now inside a model you cannot audit. None of that is unusual. Most AI contracts in Malaysia are still negotiated against a generic software template, and AI vendors know it.
The work of catching those gaps does not belong only to Legal. You know the workflow, you know which data your team actually handles, and you will be the one explaining the incident if it goes wrong. What you need is a way to turn a sales conversation into something a lawyer can review in twenty minutes.
That is what the twelve questions below do. Each one maps to a specific clause you should expect in the contract, and each answer gets a score: pass, flag or block. Score the vendor before you sign, send the scored sheet to Legal, and the conversation shifts from "does this look good?" to "what do we need to fix in the paper?"
How the pass / flag / block scoring works
Keep it mechanical. The point is not to grade the vendor's personality, it is to surface the clauses that need intervention before signature.
- Pass — the contract already addresses it in writing, with specific language. No further action.
- Flag — the intention is there verbally or in a proposal, but the clause is missing, vague, or one-sided. Escalate to Legal as a redline request.
- Block — the vendor will not commit in writing, or the answer contradicts something else you were told. Do not sign until it is resolved.
A rough threshold that works in practice: zero blocks, and no more than three flags on contractual clauses (items 2, 3, 5, 7, 8, 9, 11 below). If a vendor accumulates blocks on data deletion or training-data use, walk away. Those two are the ones that stay expensive long after the contract ends.
Part 1 — Data and model clauses

Mapping vendor data clauses to the workflows your team actually runs.
Q1. What exactly happens to our prompts and documents?
The clause to look for: a data-use and confidentiality provision that distinguishes between processing your data to deliver the service and using it to improve or train models.
Ask the vendor to point to the sentence that says your inputs are not used to train shared or foundation models. If they say "we do not train on enterprise data" but the clause only covers "customer personal data", flag it — internal SOPs, pricing sheets and process maps usually are not personal data and fall outside that promise. Score pass only when the contract excludes training on your content, not just your customers' data.
Q2. Who else touches the data, and where does it sit?
AI stacks are layered: the vendor, their cloud host, one or more model providers, an observability tool, maybe a translation service. Look for a sub-processor clause with (a) a current list, (b) an obligation to notify you before adding a new one, and (c) a right to object.
Also ask for the hosting region in writing. "Asia Pacific" is not an answer a Malaysian regulator, auditor or internal risk committee can work with. Get the country and the region code. Note that a data processing agreement (DPA) that fails to name the model provider is a block for regulated workflows, no matter how good the demo was.
Q3. Can the model change under us without notice?
This is the clause most often missing, and it is the one that quietly breaks a workflow. If the vendor swaps a model version, output quality, tone, refusal behaviour, latency and cost per task can all shift overnight. Your team built a process around the old behaviour.
Look for a model change notification clause: advance notice, a description of what changes, a version pinning option, and a reasonable window to test before the change lands in production. If the contract says the vendor "may update the service at its discretion", that is a flag at minimum; if you are running an automated approval or customer-facing workflow on it, treat it as a block.
Q4. What training data or third-party sources sit underneath?
Ask what the model was trained on at a category level, and whether the vendor will indemnify you for intellectual property claims arising from its own model and training data. Look for an IP indemnity clause that actually names the AI output, not just the software.
You are not going to audit a foundation model. You are checking whether the vendor has a documented position and whether they will stand behind it contractually. "The model provider handles that" is an answer about someone else's contract, which means it is not in yours.
Part 2 — Contract clauses you need in writing
Q5. What audit rights do we actually have?
Many AI contracts grant audit rights over the vendor's office in a way that never touches the model, the logs or the evaluation results. Ask precisely: can we request an independent audit report annually, and can we review the logs of prompts and outputs that our own users generated?
Look for an audit clause that includes (a) the type of evidence you will receive — SOC 2 Type II report, penetration test summary, evaluation results — and (b) a frequency. Supporting evidence with no obligation to hand it over is a flag.
Q6. What is the incident reporting timeline, and who tells us?
Get a number. Look for a security and AI incident clause with a defined notification window (72 hours is a commonly negotiated outer bound, tighter for sensitive workflows), a named contact role rather than a generic support inbox, and a definition of "incident" that includes model behaviour failures — a chatbot leaking another customer's document, for example — and not only data breaches.
If the clause covers only "unauthorised access", ask Legal to widen it. Model-level failures rarely look like break-ins.
Q7. How does data come back, and how is it deleted on exit?
Exit clauses are where AI contracts are weakest. Look for a data return and deletion clause that specifies: the format of the export, the deadline (30 or 90 days is typical), the deletion of backups within a stated period, and a written certificate of destruction on request.
One more line worth insisting on: deletion from any fine-tuned model, embedding store or vector index created from your content. A vendor that deletes the source files but keeps your documents as embeddings has not deleted your data in any meaningful sense. If that sentence is missing, flag it; if the vendor says it cannot separate your content from the index, block.
Q8. Who is accountable for decisions the system supports?
Human oversight language in AI contracts is often decorative — "the customer remains responsible for all outputs." That sentence shifts every consequence to you while leaving the vendor's design choices unexamined.
Look for a clause that (a) states where human review is mandatory in the workflow, (b) commits the vendor to flagging low-confidence outputs, and (c) confirms the vendor's own human review of model behaviour. For workflows touching people — hiring, credit, performance, customer eligibility — the absence of any human-in-the-loop commitment is a block, not a negotiation point.
Part 3 — Commercial and accountability terms
Q9. What are the acceptable-use restrictions, and do they match our policy?
Vendors publish acceptable use policies; contracts incorporate them by reference. That means the vendor can change your obligations by changing a web page. Check three things:
- The acceptable use policy is attached as an annex, or changes to it require notice and a right to terminate.
- Its restrictions do not contradict your internal AI policy.
- It does not prohibit the legitimate use case you bought the tool for.
That last point is more common than it sounds. Teams have bought summarisation tools whose terms barred uploading third-party confidential documents — which was the entire job.
Q10. How is usage priced, and what happens when the team actually adopts it?
Get the pricing clause in units you can forecast: per seat, per token, per document, per resolution. Ask for a worked example at your realistic monthly volume, and ask what the overage rate is once you exceed any included allowance.
Look for a price protection clause — cap on renewal increases, or a notice period for pricing changes. A contract with uncapped per-token overage and no volume ceiling is a flag even at a low headline price.
Q11. What service levels are committed, and what remedy exists?
Uptime commitments are standard; meaningful ones come with service credits and an exit right if performance misses for consecutive months. For AI specifically, ask whether there is any commitment on accuracy, latency or evaluation performance for your use case, and what happens if the vendor's own benchmark results do not hold in your environment.
Be realistic — no vendor will guarantee accuracy. But a vendor with no service credits and no termination right is asking you to carry all the risk.
Q12. Who owns the outputs, and can we build on them?
Ask for an explicit clause assigning rights in the outputs generated from your data to you, and confirming you may use them commercially. Separately, ask whether anything in the contract restricts you from building a similar workflow with a different vendor later. Broad non-compete or exclusivity language hidden in an AI tool contract is a block for any organisation planning a multi-vendor strategy.
Tool-vs-tool comparison: what the same question returns
The answers vary enormously by category, which is why the sheet is more useful than the sales deck. This is a general pattern across the market, not a statement about any specific vendor's current terms.
| Due diligence question | Off-the-shelf enterprise assistant | Specialised point-solution tool | Custom-built AI workflow |
|---|---|---|---|
| Training-data exclusion | Usually strong in enterprise tier; verify the free tier is blocked for staff | Mixed; often vague about fine-tuning | Negotiable and explicit, since you define it |
| Sub-processor list | Published, with change notice | Often short or absent | Named in the statement of work |
| Model change notification | Typically none; version changes are silent | Rare | You control the upgrade decision |
| Audit evidence | SOC 2 available on request | Depends on vendor size | Audit is your own internal process |
| Incident reporting window | Contractual, in the DPA | Frequently support-ticket based | Defined in the SLA you write |
| Deletion on exit | Standard, but silent on embeddings | Usually silent | Specified by design |
| Human oversight clause | Generic customer-responsibility wording | Usually absent | Built into the workflow specification |
| Pricing predictability | Per-seat, easy to forecast | Usage-based, variable | Fixed build plus running cost |
Read the table the way a department head should: the enterprise assistant is usually strongest on data terms and weakest on change control; the point solution is convenient but thin on paper; the custom build inverts both. Where off-the-shelf tools run out of contractual detail, organisations often need custom AI solutions with terms written around the actual workflow. If your organisation is still at the tender stage rather than the contract stage, the clause-first approach in this AI procurement in Malaysia tender spec framework will save you a renegotiation later.
Role-by-role breakdown: who owns which question
Due diligence stalls when nobody owns the sheet. Assign it like this.
Department head (you). Owns Q1, Q10, Q12, and the workflow half of Q8. You are the only person who knows what data the team actually handles and what volume adoption will produce. Your job is to answer these from real usage, not from the vendor's example.
Legal. Owns Q2, Q3, Q5, Q9, Q12 and the drafting of redlines. Legal should not be asked to interpret what your team does with the tool; feed them the workflow description alongside the sheet.
Risk and compliance. Owns Q6 and Q7, and sets the escalation threshold. They also decide whether an AI data processing agreement in Malaysia needs to reference PDPA obligations explicitly and whether the vendor's incident process can meet your internal timeline.
IT and security. Owns Q2 (technical verification of hosting and sub-processors) and Q5 (evidence review). Ask them to validate the vendor's claims against what they can observe, not just what is written.
The business owner who signs. Owns Q11 and the final block/pass call. Fifteen minutes of their time on the score sheet prevents the pattern where a contract is signed by someone who has never seen the workflow.
For teams scaling beyond a single tool, the same questions recur in a AI framework for the enterprise — clause review is cheaper at the start than at renewal.
Running the score sheet without a legal team on call
Most mid-sized Malaysian organisations do not have an AI-specialist lawyer in the building. That is workable if you sequence it properly.
- Score first, discuss second. Fill in the sheet from the contract documents, not the proposal. Anything you cannot find in writing is a flag by default.
- Send the score sheet with the redline request. Legal reviews a specific ask far faster than a general "please review this contract".
- Time-box it. Fifteen working days from shortlist to signed is realistic for a standard tool; complex or regulated workflows take longer. If the vendor pushes for a signature before the sheet is complete, that is information about the vendor.
- Record the decision. Keep the scored sheet with the signed contract. At renewal, the same document becomes your change log — particularly for Q3, Q10 and Q11.
- Get a legal review before you sign. Everything here is a business and contractual review, not legal advice, and Malaysian AI regulation, PDPA interpretation and sector rules continue to evolve. A qualified practitioner should read the final paper.
Two further cautions worth building into the process. Do not treat a favourable demo as evidence for any of the twelve answers — demos show the happy path, and contracts cover the rest. And do not accept "we will sort that out in the SLA" as a substitute for a clause; a commitment that appears after signature is not a commitment.
Where the sheet produces flags on human oversight, data-use boundaries or employee usage expectations, a responsible AI and governance workshop can give the room shared language before the next negotiation round. The same discipline is useful for your own teams when they build their first internal assistants — the ChatGPT Enterprise adoption patterns risk and legal teams get wrong is a useful companion read.
Across the engagements AIHQ has run, including a twelve-month capability journey with Media Prima Group and leadership briefings for organisations such as Lion Group, MTD Group and MUI Group, the pattern holds: the organisations that negotiate AI terms early spend less time fixing the workflow later. You can review what AIHQ does across training, advisory and implementation if you want the background, but the sheet stands on its own.
What a finished score sheet looks like
A vendor clearing all twelve questions typically looks like this: three pages, one row per question, columns for clause reference, vendor answer, score and action owner. Two blocks and one flag is a normal first pass for an unmodified vendor template — that is not a reason to reject the vendor, it is a reason to negotiate.
What matters is that no block is left open at signature, and that the flags are written into the redlines rather than parked in a meeting note. The department head who brings a completed sheet to the signature meeting is the one whose team gets a contract they can actually live with.
FAQ
Is this AI vendor risk assessment checklist a substitute for legal review?
No. It is a business and contractual screening tool that helps you arrive at legal review with specific, actionable redlines. Malaysian AI regulation, PDPA interpretation and sector-specific rules continue to change, and a qualified lawyer should review the final contract before signature.
How many questions should a vendor fail before we walk away?
There is no fixed number, but a practical rule is zero blocks at signature and no more than three open flags on contractual clauses. Blocks on training-data use or data deletion at exit are usually reasons to walk away rather than negotiate, because both create obligations that outlast the contract term.
Our vendor says their AI data processing agreement already covers everything. Should we accept it?
Read it against the twelve questions first. A vendor's own DPA is a good starting point but often covers personal data only, leaving out internal documents, model change notification and deletion from embeddings. Anything not addressed in writing should be scored as a flag.
How long does AI vendor due diligence usually take for a Malaysian department?
For a standard department-level tool, fifteen working days from shortlist to signature is realistic when the score sheet is completed early. Regulated workflows, or vendors with unmodified contract templates, typically take longer because the redline cycle repeats.
Who should own the due diligence score sheet?
The department head who will use the tool should own it, with legal, risk, IT and the signing business owner each responsible for their assigned questions. Ownership split across four functions without a named lead is the most common reason the sheet never gets finished.
Does a custom AI workflow reduce due diligence effort?
Not effort, but it changes the shape of it. With a custom build the data-use, model change and deletion terms are written for your workflow rather than inherited from a vendor template, so more of the negotiation happens before development starts and fewer surprises appear at renewal.