Contract Clauses for Outsourced AI Projects
A contract for an outsourced AI project needs three clauses an ordinary software contract never writes: who owns the evaluation set once it is built, exactly what the vendor may do with your data and outputs, and who owns a model or fine-tuned artifact trained on your material. Put all three in writing before signature, not after a dispute.
Book a solution review See the related solutionBy Huy Nguyen (David), Founder & CEO · Updated · 11 min read
This guide is for a CTO, VP Engineering, founder, or the procurement or legal reviewer turning an AI vendor's proposal into a signed contract. It assumes you have already weighed outsourcing itself and read the general IP-protection checklist for assignment, background-IP and confidentiality clauses; this page goes deeper on the three clause types an AI deliverable adds on top of that, whether you contract through a dedicated development team or an offshore development center. It is general guidance, not legal advice.
In this guide
- What does an ordinary outsourcing contract leave out for an AI project?
- The three clauses, compared
- Writing each clause into the contract
- Where AI fits in an outsourced AI contract
- What delivery record exists, and what does not
- Alternatives: who negotiates these clauses for you
- Limits of this guide
- Frequently asked questions
- Next step
What does an ordinary outsourcing contract leave out for an AI project?
A standard outsourcing contract already assigns custom code to you, licenses the vendor's own tools, and binds the team to confidentiality; our IP-protection guide covers that ground. An AI engagement produces three more artifacts that a code-and-confidentiality contract never names: the evaluation set the team builds to judge the system, the record of what the vendor's systems did with your data and outputs while building it, and the model or fine-tuned weights trained on your material. None of the three is "code" in the sense an assignment clause usually covers, so without a clause naming each one explicitly, a vendor can walk away from the engagement still holding all three and call it background know-how.
The three clauses, compared
| Clause | What it must say | Why an ordinary contract misses it | What happens without it |
|---|---|---|---|
| Evaluation-set ownership | The evaluation set, the labelled examples and the scoring logic are deliverables, owned by you, handed over in a format you can reload and re-run | An assignment clause written for "code and documentation" does not mention a labelled test set as a category of deliverable | The vendor keeps the one artifact that proves whether the system works, and a future vendor cannot re-test it without rebuilding it from nothing |
| Data use and training restriction | The vendor may use your data, prompts, files, outputs and feedback only to deliver the contracted service; it may not use them to train, fine-tune or improve any model unless you give prior written approval | A confidentiality clause stops disclosure to third parties; it does not by itself stop the vendor's own systems from training on your material | Your data, or patterns derived from it, can end up inside a model the vendor also serves to other customers |
| Model and fine-tuned-artifact ownership | Any model, adapter or set of weights trained or fine-tuned on your data is yours, or you hold a licence broad enough to run it indefinitely without the vendor, and it is never reused for another customer | IP clauses are usually written around source code; a trained model is a statistical artifact, not source code, so a code-only assignment clause does not reach it | At termination the vendor can keep the only working copy of the system's "brain," or reuse what it learned from your data on someone else's engagement |
Writing each clause into the contract
-
Define "Customer Data" broadly before anything else
Cover every input, prompt, uploaded file, generated output and piece of feedback the vendor's systems touch; a narrow definition ("files you upload") leaves prompts and generated outputs unprotected.
-
Default to no training, make any exception opt-in
Write that the vendor processes Customer Data only to deliver the contracted service and does not use it to train, fine-tune or improve any model, service or prompt library without your prior written approval; if you do permit limited use (for example aggregated, de-identified usage metrics), name the de-identification standard and ban re-identification.
-
Name the evaluation set as a deliverable, not a by-product
State who owns it, that it is handed over at each milestone and at termination in a format you can reload, and who owns the scoring logic used to grade it; this is the artifact our fixed-price AI scoping guide builds through a discovery spike, and the clause is what turns that spike's output into something you keep.
-
Set a deletion and retention scope that names every copy
Backups, caches, logs, development and staging environments, and the vendor's own support tickets, with written certification inside a stated window after termination; "delete your data" without naming these categories usually leaves at least one copy behind.
-
State who owns a model or fine-tuned artifact trained on your data
The buyer-favorable default is that anything trained on your material is yours, or licensed to you broadly enough to run without the vendor, and is never served to another customer; a vendor's own pre-trained, general-purpose model stays the vendor's, licensed to you for the term.
-
Write the handover mechanics for the model itself, not only the code
Which format the weights, adapters or prompts are delivered in, which accounts and API keys sit in your name, and what happens to any copy that remains on the vendor's infrastructure after the engagement ends.
Where AI fits in an outsourced AI contract
Netbase works with the major commercial and open-source AI tools and models, chosen per project rather than standardized on one vendor's stack, which is also why a model-ownership clause should name the trained artifact and the evaluation set rather than a specific vendor's platform, since the underlying model can change over the life of a system.
-
In delivered work
Retrieval-augmented knowledge assistants, document AI and MLOps pipelines
These engagements produced the same three artifacts this guide names — an evaluation set, a data-use record and a trained or fine-tuned component — for clients that are not named; no count, duration or result is published for them.
What delivery record exists, and what does not
In Netbase custom development, the client owns the IP created for it; Netbase's own productized modules and Business Division products stay licensed, not transferred — the same split a model-ownership clause has to make explicit when the "custom work" is a trained artifact instead of application code. Netbase contributors work under NDA, and NDAs and data processing agreements are available to clients on request, which is the contractual base a data-use clause builds on. Fixed-price is Netbase's primary commercial model, with milestone-based, monthly team retainer and KPI-linked terms also offered, and these commercial terms are unaffected by which of the three AI-specific clauses apply; the clauses govern ownership and data handling, not price. The anonymised retrieval-augmented, document AI and MLOps engagements referenced above ran under this same assignment-first, NDA-backed practice, extended to name the AI-specific artifacts.
What does not exist. No published Netbase case study states its contract's clause wording, and Netbase does not publish its contract templates. This page makes no claim about dispute history, states no evaluation-set size or retention window as a Netbase figure, and is not evidence of any particular clause having been negotiated on a named engagement.
Alternatives: who negotiates these clauses for you
| Route | Strength | Fits when |
|---|---|---|
| Your own legal counsel, briefed on AI-specific terms | Counsel who already knows your jurisdiction and risk tolerance | You have in-house or retained counsel and only need the three AI-specific clauses added to a template they already use |
| The vendor's standard contract, reviewed clause by clause | Fast to sign if the vendor's default terms already name all three artifacts | The engagement is small and the vendor's own template already covers evaluation-set ownership and a no-training default, so little needs to change |
| A delivery partner that writes these terms into its standard engagement from the start | The evaluation set, the data-use terms and the model-ownership clause come from the same team that then builds the system, instead of a late addendum negotiated after a generic contract is already signed | You want the AI-specific terms handled as part of how the engagement is set up, not as a separate legal exercise bolted onto a generic outsourcing contract |
In Netbase custom development, the client owns the IP created for the work, assignment is written in from the start rather than negotiated as an afterthought, and most projects run on fixed-price contracts agreed after discovery, with milestone-based, monthly team retainer and KPI-linked terms also offered where they fit the scope better. The offer sits in dedicated development teams and the offshore development center model; the AI knowledge assistant solution and the professional services industry page cover an adjacent delivery shape for the kind of buyer who most often needs these clauses reviewed by counsel.
Limits of this guide
- It is general information, not legal advice; contract law, data-protection rules and what counts as a trade secret differ by country, so take local legal advice before signing.
- It does not repeat the general assignment, background-IP, open-source and GDPR processor-contract material already covered in protecting IP when outsourcing, or the pricing and acceptance-criteria framework in the fixed-price AI projects guide; read those alongside this one.
- The sample clause language described here is illustrative of what each clause must accomplish, not wording to copy into a contract without review by qualified counsel for your jurisdiction.
- This guide does not cover what happens when a dispute over these clauses actually arises; that is a question for counsel once a disagreement is confirmed, not an engineering or governance framework question.
Plan the next step with a Netbase consultant
Frequently asked questions
Only if the contract lets it. The buyer-favorable default is the opposite: the vendor processes your data only to deliver your contracted service and needs your prior written approval before using it, or anything derived from it, to train or improve any model.
You should, together with the scoring logic used to grade it, for the same reason you own the rest of the custom work you pay for: without it, nobody — including a future vendor — can re-test whether a later change to the model still works.
Not automatically. A code-assignment clause is usually written around source code and documentation; a trained or fine-tuned model is a separate, statistical artifact that needs its own ownership sentence, or it can be left in a grey zone the vendor calls its own know-how.
No. Confidentiality stops disclosure to third parties; it says nothing about the vendor's own systems using your data internally to improve a model, which needs its own explicit restriction.
Next step
Share your draft AI vendor contract or statement of work and we will book a solution review to check whether the evaluation-set, data-use and model-ownership clauses are in it before you sign. You can also read how to hire an offshore AI engineering team in Vietnam or scope a fixed-price AI project, or browse more Netbase insights.
Work behind this guide
Netbase delivered an MLOps pipeline remotely, combining AI consulting, engineering and integration.
Keep ReadingRelated services and solutions
An AI-augmented dedicated development team in Vietnam that answers for results
Netbase provides dedicated development teams from Vietnam for product companies and enterprises to add engineering capacity without losing control. Teams of 3 to 30 people work only on your product, use AI-assisted engineering under review, report to a named account manager and project manager, typically start within one to two weeks after discovery, and the IP they create is yours.
Learn More
An AI-enabled offshore development center in Hanoi, when a long roadmap needs one
An offshore development center (ODC) in Vietnam is a long-running engineering team that a partner abroad staffs, manages and documents for your product roadmap. Netbase runs ODC teams from its only office, in Hanoi, with AI-assisted engineering, weekly governance and knowledge transfer built in. When the work has a defined outcome, a fixed-price project usually fits better.
Learn More
AI knowledge assistant: answers from company knowledge, with sources and access control
An AI knowledge assistant is an internal tool that answers employees' questions from your company's own documents and systems, cites the source behind every answer and shows each person only what they may see. Netbase offers it as a growth capability, designed with retrieval, evaluation and human review from the first pilot, and has delivered one for an unnamed client.
Learn More
Work with Netbase remotely
Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.
Discuss a project
Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.project@netbasejsc.com
+84 937 869 689
91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam
Get in touch
Tell us what you want to build, modernize, or operate.