Getting started
AI design partner agreements: what data are you giving away?
By SourceX Editorial · Reviewed by Noah Loul ·
Short answer
An AI design partner agreement usually gives a vendor access to your records in exchange for early access, discounts or product influence, and its data terms often reach well beyond the pilot. Before signing, check data scope, training rights, ownership of model improvements, exclusivity, deletion at the end and what you receive if the vendor trains general models.
Key takeaways
- Design partner agreements are drafted by the vendor, and their data terms often amount to an unpaid data license.
- Integrations reach whole workspaces, so scope should name systems, record types and date ranges.
- Staff corrections to a vendor's outputs are human-generated, domain-specific signal with real value.
- Training a vendor's general model on your records should require a separate license with its own terms.
- Deletion at the end should cover derived data and backups and come with a signed certificate.
What is an AI design partner agreement?#
An AI design partner agreement is a contract in which an early-stage AI vendor gives a company early access, discounted pricing or influence over the product in exchange for feedback, a reference and, very often, access to the company's records. The vendor usually drafts it, and the data terms are written for the vendor's roadmap rather than yours.
The pilot itself is often worth doing. The issue is that a document framed as a product trial can contain a data license broader than one the company would sign if it were asked directly to license its support history or proposals for model training.
What data are you actually giving away?#
A design partner usually receives more data than the pilot needs, because integrations connect to whole workspaces rather than to a chosen set of records. A connector to Zendesk, Salesforce, Slack or SharePoint can reach every ticket, account note or channel the connecting user can see.
Staff corrections deserve particular attention. When experienced estimators or support engineers fix the vendor's drafts, they produce exactly the human-generated, domain-specific signal that the SourceX Enterprise Data Value Framework treats as a value driver, and a broad feedback clause can hand all of it to the vendor.
- Historical records pulled through the integration, often going back years.
- Live records created during the pilot.
- Derived data: embeddings, labels, summaries and search indexes built from your records.
- Your staff's corrections, ratings and edits to the vendor's outputs.
- Usage data showing how your team works inside the product.
- Prompts and documents your staff upload directly.
Clause checklist for design partner agreements#
The clause checklist compares the version a vendor usually drafts with the version a supplier of records should ask for. Most early-stage vendors expect to negotiate these points with design partners they value, because a reference customer matters to them.
It helps to know what the vendor genuinely needs. Product feedback, a named reference and permission to describe the pilot in sales conversations usually matter more to an early-stage vendor than open-ended training rights, so offering those freely can make the data terms easier to narrow.
| Clause | Vendor-friendly draft | Supplier-friendly version |
|---|---|---|
| Data scope | All customer data available through the integration | Named systems, record types and date ranges |
| Training rights | Use customer data to train and improve models | Use only to deliver the pilot to your company |
| Model improvements | All improvements belong to the vendor | No general model trained on your records without a separate license |
| Aggregated or de-identified data | Undefined, perpetual right | Defined standard, no re-identification, ends with the agreement |
| Feedback | Vendor owns all feedback | Vendor may use product feedback, not your underlying records |
| Exclusivity | Silent, or you commit to the vendor's category | No restriction on licensing your records elsewhere |
| Deletion at end | Return or delete on request | Delete all copies, derived data and backups, with a certificate |
| What you receive | Early access | Defined pricing, credits or fees if training rights are granted |
How design partner terms compare with a data license#
A design partner agreement and a data license can cover the same records but serve different purposes, and the difference shows up in compensation, scope and documentation. Treating training rights as a license, rather than as a side term of a pilot, is the main protection.
The comparison also explains why the same records can be worth very different amounts in the two settings. In a pilot, the vendor's access is a cost of getting a product; in a license, access is the product, and it is prepared, scoped and documented accordingly.
| Question | Design partner pilot | Data license |
|---|---|---|
| Purpose | Test and shape a product | Train or evaluate models |
| Compensation | Early access or discounts | License fees agreed with the buyer |
| Scope | Often whatever the integration reaches | A defined package of records |
| Preparation | Usually none; live records flow in | Personal and confidential details removed first |
| Documentation | Pilot terms and a security questionnaire | A record of provenance, rights, permitted use and release approval |
Red flags in a vendor's draft#
Red flags in a design partner draft usually hide in definitions and survival clauses rather than in the headline terms. Read these sections before the commercial ones.
If the agreement is a click-through, save a copy of the version accepted, with the date and the name of the person who accepted it. Vendors update online terms, and you will need to know which version applied when your records were shared. FTC staff warned in February 2024 that quietly adopting more permissive data practices, such as AI training, through a retroactive change to terms may be unfair or deceptive, but that guidance protects consumers and does not replace reading your own contract.
- Customer Data defined to exclude usage data, outputs or feedback, which then fall under broader vendor rights.
- A right to use data to improve the vendor's products and services generally, with no limit to your account.
- An aggregated data carve-out with no definition of aggregation or de-identification.
- License grants that survive termination with no end date.
- No deletion certificate, or deletion only on written request.
- No list of the model providers and subprocessors that will receive your records.
Illustrative: a consulting firm narrows a proposal-writing pilot#
Illustrative: a fictional management consulting firm agrees to be a design partner for a startup building an AI proposal assistant. The draft would connect the startup to the firm's SharePoint, which holds proposals, staffing plans, internal playbooks and client deliverables, and grants the startup the right to train its models on all of it.
The firm's general counsel narrows the scope to firm-owned proposals and playbooks, excludes every client folder because client contracts restrict reuse, limits use to the firm's own instance and requires deletion with a certificate at the end. Any training of the startup's general model would need a separate paid license. The startup accepts, the pilot runs, and the firm keeps the option to license its proposal history later on its own terms.
How SourceX treats training rights in a pilot#
SourceX treats any request to train general models on a company's records as a data license, whatever the document is called. That means running it through the SourceX five-step transaction, Supply, Rights, Preparation, Approval and Delivery, with the company approving each step.
The SourceX Enterprise Data Value Framework helps weigh what a pilot asks for: uniqueness, domain expertise and human-generated signal raise value, exclusivity raises price, and preparation cost and privacy burden reduce net value. This is general information, not legal advice; review specific agreements with counsel.
Frequently asked questions
Is being an AI design partner a bad idea?
No. Design partnerships can give a company useful tools early and real influence over a product. The risk sits in data terms that reach beyond the pilot. Separating product feedback, which the vendor needs, from training rights over your records, which deserve their own terms, keeps the benefits without the hidden license.
What does model improvements mean in these agreements?
It usually means any change to the vendor's models made using your data or feedback, including retraining, fine-tuning and evaluation results. Drafts often give the vendor ownership of all of it. Ask whether improvements stay within your instance or flow into models sold to other customers, including your competitors.
We already signed a broad design partner agreement. What now?
Read the termination, deletion and survival clauses first. You may be able to end training rights going forward, request deletion of datasets and derived data, and ask for a certificate. Training that already happened usually cannot be undone, so focus on future scope and on stopping new records from flowing in.
Should a vendor pay for training rights in a pilot?
If the vendor wants to train models it will sell to others on your records, that is a data license, and compensation is a reasonable request. It can take the form of fees, credits or pricing commitments. If the vendor only needs your data to run your own pilot, early access may be a fair exchange.
Who should sign off on a design partner agreement?
Whoever signs commercial contracts should sign, but only after counsel reviews the data terms and the system owner confirms what the integration can reach. Pilots are often started by a department head with a click-through, which is how broad data rights get granted without anyone reading them.
Sources
- On February 13, 2024, FTC staff warned that adopting more permissive data practices, such as using consumers' data for AI training, through a surreptitious, retroactive change to terms of service or a privacy policy may be unfair or deceptive. Source
Related resources
- InsightOpt-in vs opt-out for AI training in B2B SaaS contracts
- InsightDoes AIA B101 let an architect license project records for AI?
- InsightCan you license data from customers whose contracts say nothing about AI?
- DataSales call transcripts
- QuestionShould companies sell or license their data?
- QuestionDo AI labs buy legal documents?
See if your company qualifies
A short company assessment. No data uploads are needed.