Skip to main content
SUWANVARA LAWFIRM
Suwanvara Law Firm Co., Ltd.
SUWANVARA LAWFIRM
SUWANVARA LAWFIRM
Suwanvara Law Firm Co., Ltd.
Business Law

Selling AI-Powered Services: How to Write Scope, Human Review and Liability in Client Contracts in Thailand

For agencies and technology providers who use AI to produce client deliverables: what you can and cannot warrant, human review, client data and model training, confidentiality, IP in outputs, liability caps and incident handling.

Suwanvara Law Firm— Business Law TeamNovember 21, 20269 min read
Selling AI-Powered Services: How to Write Scope, Human Review and Liability in Client Contracts in Thailand

The core answer

A company selling work that AI helps produce should spell out four things in the contract: what the AI does in this job, who reviews the output before delivery, how the client's data is used and kept, and who bears what if the output is wrong or data leaks, and up to what cap. A contract that is silent on these points often leaves the client believing the work was purely human, or leaves the provider carrying more risk than intended.

As of 21 November 2026 there is no AI-specific law of general application in Thailand, although drafts and guidance exist and readers should check the current position before relying on this. Meanwhile these contracts sit under general contract law, the PDPA and intellectual property law. The ETDA Generative AI Governance Guideline for Organizations (announced 30 October 2024) is practice guidance for executives and staff, not binding law, but it is a useful frame for showing how an organisation balances benefits against risk.

Who this guide is for

Your stageStart with
Not yet signed, drafting or reviewingSections 1 to 8 and the checklist
Signed and deliveringSections 3, 4, 6 and 8, to check what you do matches the contract
Dispute or incident alreadySections 6, 7 and 8, and keep the documents in the checklist

For the internal question of what staff may paste into AI tools, see the company AI policy and customer data guide. This guide covers the contract with the client only.

1. Must we tell the client we use AI, and where?

Thai law has no general rule requiring disclosure of AI use in every job. But a silent contract invites the argument that the client misunderstood something important, especially where price, timing or quality were agreed on the assumption that people did the work. Write it plainly and truthfully:

  • which steps use AI (drafting, summarising, translating, analysing, image generation, coding) and which are done by people
  • whether the client may opt out of AI for certain parts
  • whether the provider may change tools during the project, without lowering the agreed quality

2. What should the scope say, and what can we warrant?

Good scope separates deliverables from expected outcomes. A provider can warrant that it will deliver to specification and pass the stated checks. It should not warrant things that depend on outside factors or on how AI works, such as 100% accuracy, search ranking, sales or a diagnosis.

TopicWording you can stand behindWording to avoid
QualityWork passes the agreed check list before deliveryNo errors at all
Factual accuracyKey figures and sources checked as listedAI verifies every fact
Business resultsDelivery on the agreed plan and datesGuaranteed sales or ranking
OriginalityThe stated screening process is followedGuaranteed unique and non-infringing
TimeTimeline and number of revision roundsInstant or unlimited revisions

3. How should human review be written?

The ETDA guideline frames Generative AI use as balancing benefits against risk management. Human review should therefore be a defined step with a named owner, not a vague "checked".

  • State which work needs a person's review before delivery, for example content with figures, legal statements, product claims, health or financial information.
  • State the reviewer's role and keep a record of who reviewed what and when.
  • State how many days the client has to accept the work, and whether silence counts as acceptance.
  • State what the provider will not leave to AI without human review.

4. Client data: can it train models, and what agreement is needed?

If the data the client sends includes personal data, such as customer lists, chat logs or images of people, a provider acting on the client's instructions will often be a data processor and the client the data controller. PDPA section 40 sets duties for the processor, including acting on the controller's instructions and putting security measures in place. A data processing agreement should therefore accompany the service contract.

It should cover at least:

  • purpose, data types and duration of processing
  • use for model training: the safe default is no, unless the client authorises it in writing and has the right to do so
  • the AI vendors used (sub-processors) and the client's right to object to changes
  • where the systems are located; if data is processed abroad, check the cross-border transfer rules in PDPA section 28
  • deletion or return of data at the end of the job, and notice to the client of any breach

The terms of the upstream AI vendor you use must not contradict what you promised the client. If you promise no training use but the service tier you bought lets the vendor train on inputs, you carry that risk. The PDPA basics are in the PDPA compliance checklist.

5. How should confidentiality and IP in outputs be drafted?

Confidentiality. State that client data and work are confidential, who may access them (including subcontractors and AI vendors), and that they must not be pasted into public tools. Trade secrets are protected where the owner takes reasonable steps to keep them secret; see customer data and trade secrets.

IP. Copyright protection for output created wholly or partly by AI is unsettled in Thai law, so draft with restraint:

  • the provider assigns or licenses the rights it has, and states the condition, such as full payment
  • say which parts were made by people and which by AI, as far as records allow
  • give no assurance that the output is protected by copyright or does not overlap with others' work
  • keep the provider's tools, templates and know-how separate from the deliverables
  • have the client warrant it has the right to use any third-party images, text or content it supplies

On rights in work done for a client and portfolio use, see permission to show client work in a portfolio.

6. How should accuracy disclaimers and liability caps be written reasonably?

AI can produce text that sounds right and is wrong. A good disclaimer states that limit plainly and ties it to the review steps the provider actually performs, rather than disclaiming everything.

  • A statement of AI's limits is reasonable. But an overbroad clause in a standard-form contract drafted by one side may be challenged as unfair to the other, and a court may look at the circumstances. A disclaimer does not always protect the provider.
  • Tie the cap to a sensible measure, such as fees received over a period, and say clearly what it covers. Do not expect to limit liability for intentional acts or gross negligence.
  • Treat data leaks or PDPA breaches separately from ordinary liability because the scale of loss differs. If you set a special cap, agree it expressly.
  • The provider answers for what it delivers even if AI drafted it. Blaming the tool does not help in a negotiation with the client.

Caps, indirect loss and indemnities are covered in the guide on limitation of liability and indemnity.

7. What should the contract say about incidents, leaks and wrong output?

  • Who notifies whom, and by when after becoming aware of the incident
  • Investigation and evidence: usage logs, work versions, list of people with access
  • Correction: number of rounds, rework and timeline
  • Personal data incidents: the processor must tell the controller so the client can meet its own obligations; statutory notice periods may be short, so check the PDPC pages
  • Upstream AI vendor outage or change of terms: whether the provider must find a substitute and whether this counts as force majeure; do not draft it more broadly than is realistic

8. Checklist before sending an AI-powered service contract

  • Scope says which steps AI assists and which people perform
  • What is warranted is separated from outcomes that are not
  • Reviewer, review list and the client's acceptance period are stated
  • A data processing agreement exists where personal data is involved
  • Training use of client data is clearly forbidden or permitted
  • Upstream AI vendor terms checked against the contract
  • Sub-processors and system location are listed
  • IP clause does not promise more than the law supports
  • Accuracy disclaimer and liability cap are reasonable
  • Incident procedure and notification channel are set
  • Review records, work versions and client approvals are kept

Hypothetical example: a content agency for a retailer

This is a hypothetical example, not a real case. Names and figures are illustrative only.

Agency A uses AI to draft advertising copy and answer customer chats for retailer B. The old contract only said "responsible for all content". The AI later drafted wrong product specifications, and B sent them to customers before anyone reviewed them. B demanded that the agency bear everything. Separately, the agency found its AI service allowed the vendor to use conversation text to improve its systems.

A better contract would say that copy containing product specifications must be checked by the agency's staff against data supplied by B before approval, how many days B has to review, that customer chat data is processed under a data processing agreement, and that an AI tier with no training use is chosen. When an error occurs, it is then clear which step was missed and where liability sits. The outcome would still depend on the actual facts and wording.

Common misunderstandings

MisunderstandingCorrection
"Writing that AI may err means we need not answer for the work"A disclaimer helps set expectations, but a clause that is too broad may be challenged as unfair and does not replace actual review
"The AI vendor is liable if the output is wrong"Your client contracts with you, not with the upstream vendor. You answer to the client first
"A paid AI plan makes client data safe automatically"Check the terms on data use, retention and location, then compare them with what you promised the client
"AI-assisted work is certainly the client's copyright once paid"Protection for AI output is unsettled; assign or license what you have and promise no more
"The ETDA guideline is binding law"It is practice guidance without direct binding force, though useful as a governance frame
"Thailand already has an AI law"As of 21 November 2026 there is no AI-specific law of general application; drafts exist, so check the latest position

Further reading

An AI-powered service contract should say what the AI does, who reviews, where client data goes and who bears the loss and how far, and then warrant only what you can actually deliver.

We review and draft service contracts, scopes of work, data processing agreements and the confidentiality, IP and liability clauses for companies that sell AI-assisted services. We are not technical auditors of AI systems and cannot guarantee the result of negotiations with clients or AI vendors. Fees depend on the stage of the matter and the work involved. Send us the documents first and we will confirm the scope of work and fees in writing before we start. Call +66 92 254 2045 or send us the details. See also contracts and documents and PDPA compliance.


This guide is published by Suwanvara Law Firm — a Khon Kaen law firm established in 1986. General information only, not legal advice on a specific matter.

Frequently asked questions

If the contract says AI may make mistakes, is the provider off the hook?+

Not necessarily. A reasonable, plainly worded statement of AI limits helps set the client's expectations. A broad clause that tries to exclude liability in every case may be challenged as unfair, especially in a standard-form contract one side drafted in advance, and it does not help where the provider was grossly negligent or acted intentionally. A court looks at the wording, the real negotiation and the facts of each case.

The client sends us its customer data for AI-assisted analysis. What more must we do?+

If the data contains personal data, the provider will often be the client's data processor. The PDPA sets duties to act on the controller's instructions and keep the data secure. Put in place an agreement covering purpose, data types, any AI vendor acting as sub-processor, where the systems are located and deletion at the end of the job.

May the provider use client work and data to train its own models?+

Only if the contract or consent says so clearly and the data is not confidential or personal data the client has no right to release. The safe default is no training use unless the client agrees in writing. Also check the upstream AI vendor's terms match what you promised the client.

Who owns the copyright in work an AI helped create?+

Thai law has no settled answer. Copyright protection for AI-generated output is unclear, so do not promise the client full copyright. Say the provider assigns or licenses whatever rights it has, identify which parts were made by people and which by AI, and give no assurance that the output is protected.

Does Thailand have a specific AI law that applies to contracts like this?+

As of 21 November 2026 there is no AI-specific law of general application, although drafts and guidance exist and readers should check the current position. The contract is governed by general contract law, the PDPA and intellectual property law. The ETDA guideline is practice guidance, not binding law.

Lawyer-client handshake symbolizing trust at Suwanvara Law Firm

Ready to hear your story

Get an initial consultation with one of our experienced attorneys.