Data Sharing Agreements With Partners: Drafting and Reviewing With AI
Nonprofits now share client and program data with more partners than ever: collective impact tables, referral networks, school districts, health systems, government agencies running homeless services databases, university evaluators, and technology vendors. Every one of those relationships should rest on a written data sharing agreement that says exactly what data moves, why, who can touch it, how it is protected, and when it is destroyed. This guide explains what these agreements are, walks through the standard clauses (including a new one every agreement now needs about AI model training), maps the US laws that shape them, and shows how AI tools can help your team draft first versions, compare a partner's paper against your standard positions, flag missing terms, and track renewals, without replacing the legal judgment that regulated data requires.

Picture a typical week at a mid-sized human services nonprofit. The program director agrees to send attendance data to the school district so the after-school program can show its effect on chronic absenteeism. The housing team enters client records into the regional Homeless Management Information System. A university evaluator asks for three years of case data for an outcomes study the foundation is paying for. A new referral platform wants to connect your intake team with food banks and legal aid clinics across the county. And the case management software vendor has just updated its terms to mention "improving our AI features." Each of these is a data sharing relationship, and each one exposes the people you serve if it goes wrong.
Most nonprofits handle these relationships unevenly. Some are governed by carefully negotiated agreements reviewed by counsel. Others rest on an email thread, a two-paragraph memorandum of understanding signed five years ago by a former executive director, or a partner's template that nobody on your side read closely. The gaps rarely matter until something happens: a breach at the partner, a client who asks where their information went, a funder audit, or a discovery that a partner has been using your clients' records for a purpose nobody agreed to. At that point, the agreement (or the lack of one) determines what you can do about it.
Artificial intelligence has raised the stakes in two directions at once. On the risk side, partners and vendors increasingly want to use shared data to build or improve AI models, and data that has been absorbed into a model is very hard to get back. Older agreements almost never address this. On the opportunity side, AI tools are genuinely good at the tedious parts of agreement work: turning a term sheet into a first draft, comparing a forty-page partner template against your standard positions, spotting a missing breach notification clause, and summarizing obligations in plain language for the program staff who actually have to follow them.
This article covers what a data sharing agreement is and how it differs from related documents, the standard clauses and what good language looks like for each, the US legal frameworks that most often apply (HIPAA, FERPA, 42 CFR Part 2, HUD's HMIS standards, and state privacy laws), how to handle AI training and secondary use, a practical AI-assisted workflow for drafting and review, how to keep an inventory of active agreements, and the limits you need to respect. It is written for executive directors, operations leads, program managers, and data staff, not lawyers, and it is not legal advice. It should, however, make your conversations with counsel shorter, cheaper, and more productive.
What a Data Sharing Agreement Is, and What It Is Not
A data sharing agreement (DSA) is a written contract between two or more organizations that governs the transfer and use of specific data for a specific purpose. You will also see the terms data use agreement (DUA), data use license, and information sharing agreement. The labels vary by sector, but the core function is the same: the organization that holds the data sets the conditions under which another party may receive and use it, and the receiving party makes enforceable promises about how it will behave. A good DSA turns vague goodwill ("we'll keep it confidential") into specific, checkable obligations.
It helps to distinguish a DSA from documents it is often confused with. A memorandum of understanding (MOU) usually describes a partnership's goals, roles, and governance in broad terms, and many MOUs are deliberately non-binding. A vendor services contract governs a company you pay to perform work, and it may include data protection terms, but its center of gravity is the service. A DSA is narrower and more technical than an MOU and focused on the data rather than the service. In practice, larger partnerships often use more than one layer. The Actionable Intelligence for Social Policy (AISP) guidance at the University of Pennsylvania on legal frameworks for data integration recommends a tiered structure for integrated data systems: an MOU among the participating agencies, data sharing agreements governing what each contributor provides, and a data use license for outside researchers who receive data. That structure is a useful mental model even for a nonprofit with just a handful of partners.
The AISP report frames every data sharing decision around three questions that are worth adopting as your own: is it legal, is it ethical, and is it a good idea? The legal question is about whether you have authority to share. The ethical question is about whether the people in the data would reasonably expect and accept the use. The "good idea" question is about whether the benefit justifies the risk and effort. An agreement documents the answers, but it cannot substitute for asking them.
Where nonprofits share data
Common partnerships that need a written agreement
- Collective impact initiatives and shared measurement systems
- Closed-loop referral networks across community providers
- School districts and charter networks (student records)
- Hospitals, health plans, and clinics
- Government agencies and HMIS lead organizations
- University researchers and external evaluators
- Technology and AI vendors that process data on your behalf
Telling the documents apart
Similar names, different jobs
- MOU: partnership goals, roles, and governance, often non-binding
- DSA / DUA: binding terms for a specific data transfer and use
- Data use license: terms for an outside researcher receiving data
- Business associate agreement: HIPAA-required terms for a vendor handling PHI for a covered entity
- Vendor contract / DPA: a paid service with data protection terms attached
This article focuses on agreements between your nonprofit and partners who receive or exchange data for a shared mission purpose. Agreements with paid technology vendors overlap heavily, and many of the clauses below apply to them too, but vendor contracts raise additional issues around pricing, service levels, and lock-in. For those, see our guides on what executive directors should look for in AI vendor contracts and managing AI vendor contracts.
The Clauses Every Data Sharing Agreement Needs
Agreements vary in length from three pages to fifty, but the strong ones cover the same core ground. The list below is the backbone of a standard positions document, which is the single most useful thing you can build before you start using AI to draft or review. For each clause, think about what your organization's preferred position is, what you would accept as a fallback, and what is a dealbreaker.
Purpose, data elements, and legal authority
What is shared, why, and on what basis
- Purpose limitation. A specific statement of why the data is being shared ("to evaluate the effect of the tutoring program on reading proficiency for students enrolled in the 2026-27 school year"), not a general one ("to support program improvement"). Every other clause hangs off this one.
- Data elements and minimum necessary. An attached data dictionary listing each field, ideally with a one-line justification tying it to the purpose. If the evaluator does not need dates of birth, send age bands. If they do not need names, send a study ID.
- Legal authority and consent. The specific basis for the disclosure: client consent and where it is documented, a statutory exception, a HIPAA permission, or a FERPA exception. If consent is the basis, the agreement should say what happens when a client revokes it.
Permitted uses, AI training, and re-disclosure
What the recipient may and may not do with the data
- Permitted uses. A closed list. Anything not listed is prohibited. Avoid catch-all phrases such as "and related purposes" or "to improve services generally."
- AI and model training. An explicit statement on whether the data, or anything derived from it, may be used to train, fine-tune, evaluate, or improve machine learning or AI models, including the recipient's vendors' models. Silence is no longer acceptable. This is covered in detail below.
- Re-disclosure limits. The recipient may not pass data to anyone else, including subcontractors and affiliates, without written permission, and any approved subcontractor must be bound by terms at least as protective.
- Linking and re-identification. A prohibition on linking the data to other datasets or attempting to re-identify individuals unless the agreement specifically authorizes it.
- Publication and small cells. For research and evaluation, rules on what can be published, prior review rights, and suppression of small counts that could identify individuals in a small community.
Security safeguards and access controls
How the data is protected once it leaves your hands
- Security safeguards. Encryption in transit and at rest, secure transfer methods (no email attachments of identifiable files), and a reference to a recognized framework or the recipient's written security program.
- Access controls. Access limited to named roles or individuals with a need to know, multi-factor authentication, prompt removal of access when staff leave, and confidentiality acknowledgments signed by each user.
- Location and tools. Where data may be stored and processed, and whether it may be entered into third-party AI tools at all. A recipient who pastes your client records into a consumer chatbot has made a disclosure you never authorized.
Breach notification, audit, retention, and term
What happens when things go wrong, and when the relationship ends
- Breach notification. A short, specific deadline for the recipient to notify you of a suspected or confirmed incident (commonly a few business days, sometimes shorter), what the notice must contain, cooperation duties, and who pays for required notifications to individuals and regulators.
- De-identification standards. If the recipient receives or produces "de-identified" data, the specific standard that applies (for health data, HIPAA's safe harbor or expert determination methods) rather than an undefined promise.
- Audit and monitoring rights. Your right to request evidence of compliance, such as security attestations, access logs, or a completed questionnaire, and in higher-risk cases to conduct or commission an audit.
- Retention and destruction. A date or trigger for returning or securely destroying the data, including backups and derived files, plus a written certification of destruction.
- Term and termination. A fixed term with deliberate renewal rather than indefinite auto-renewal, the right to terminate for breach, and obligations that survive termination (confidentiality, destruction, breach cooperation).
Two practical points make these clauses work in real life. First, attach a data dictionary or data specification as an exhibit rather than describing data elements in prose. When the evaluator asks for "a few more fields" in year two, you amend the exhibit, and everyone can see exactly what changed. Second, name the people responsible on each side, by role, for day-to-day compliance and incident contact. Agreements fail quietly when the person who negotiated them leaves and nobody else knows they exist.
The US Laws That Shape Your Agreements
Much nonprofit data is not governed by any single federal privacy law, which is precisely why contracts matter so much: in the absence of a statute, the agreement is the main protection your clients have. But when regulated data is involved, the law dictates specific contract terms, and a generic DSA template will not be enough. Here are the frameworks nonprofits most often encounter. Each deserves a conversation with counsel when it applies.
HIPAA: BAAs versus DUAs
If you are a HIPAA covered entity, or you receive protected health information from one, two different agreements may apply. A business associate agreement is required when an organization performs a function on behalf of a covered entity that involves PHI. A HIPAA data use agreement governs disclosure of a limited data set, which removes direct identifiers, for research, public health, or health care operations.
A DUA is not a substitute for a BAA, although HHS guidance notes the two can be combined into one document that meets both sets of requirements. Many community partnerships with health systems fall into neither category, and those need a standard DSA with HIPAA-informed terms.
FERPA: written agreements with schools
Schools generally cannot disclose personally identifiable information from education records without parental consent. When a district shares data with your organization under the studies exception or the audit and evaluation exception, FERPA requires a written agreement with specific elements, including purpose, scope, and duration, and a time limit for destroying the data.
The US Department of Education's Student Privacy Policy Office publishes a Written Agreement Checklist listing mandatory elements and recommended best practices. It is an excellent input for an AI-assisted clause review.
42 CFR Part 2: substance use records
Records from federally assisted substance use disorder programs carry protections stricter than HIPAA, including limits on re-disclosure and use in proceedings against the patient. The 2024 final rule aligned Part 2 more closely with HIPAA, and HHS states that compliance was required by February 16, 2026. If your partners send or receive Part 2 records, check whether agreements signed before that date still reflect the current rule, including qualified service organization terms.
HMIS: homeless services data
Projects funded through HUD's Continuum of Care and Emergency Solutions Grants programs generally must participate in their local Homeless Management Information System under HUD's privacy and security standards. In practice, your obligations are set by the local CoC's HMIS policies, partner agency agreements, end-user agreements, and client release-of-information forms. Read them before you build any AI workflow that touches HMIS exports.
State law fills in much of the rest. Every state has a data breach notification statute, and many have specific confidentiality rules for mental health, HIV status, domestic violence services, child welfare, or public benefits records. A growing number of comprehensive state consumer privacy laws also exist, and while many of them exempt nonprofits, several do not, including Colorado and Oregon. Health-specific state laws can reach organizations that HIPAA does not. Your agreements should require both parties to comply with applicable state law, and your standard positions should reflect the strictest regime that applies to your data. Our guide to state privacy laws and AI covers the state landscape in more depth, and our sector guides on HIPAA and AI and student privacy under FERPA go further into those regimes.
One more federal point deserves attention, even though it is aimed at companies rather than charities. In February 2024, the Federal Trade Commission warned that a business which quietly changes its terms of service or privacy policy to permit AI training on data it has already collected may be engaging in unfair or deceptive practices. The FTC's jurisdiction over most charitable nonprofits is limited, but the principle is a sound ethical benchmark for your own conduct and a strong argument when a vendor or partner tries to expand its rights through a unilateral terms update.
The New Clause: AI Training and Secondary Use
If you reviewed your active agreements today, most of them would probably say nothing about artificial intelligence. That made sense when they were written. It is now a real gap. Partners and vendors increasingly want to use the data flowing through their systems to build predictive tools, improve AI features, or train models they can offer to other customers. Even a well-intentioned research partner may want to use your program data to develop a risk model they plan to publish or license. None of this is automatically wrong, but it is a different use from the one your clients agreed to, and it needs to be addressed explicitly.
Model training creates a particular problem for the retention and destruction clause. When a partner deletes a spreadsheet, the data is gone. When data has been used to train a model, the patterns learned from it persist in the model's parameters, and current techniques for making a model "forget" specific records are limited. The FTC has, in several enforcement actions against companies, required the deletion of models and algorithms built on improperly collected data, a remedy sometimes called algorithmic disgorgement. That shows how seriously regulators take the issue, and also how blunt the fix is once training has happened. Prevention through contract language is far easier.
What a strong AI use clause covers
Discuss specific wording with counsel; these are the issues to address
- A default prohibition. The recipient may not use shared data, or data derived from it, to train, fine-tune, or improve any AI or machine learning model, unless a specific use is approved in writing.
- Coverage of the recipient's vendors. The prohibition extends to any AI tool or platform the recipient uses, so data entered into a third-party service that trains on inputs counts as a violation.
- De-identified data addressed directly. State whether the prohibition also applies to de-identified or aggregated data. Many vendors will ask to carve these out; decide in advance whether you accept that and under what de-identification standard.
- Conditions for approved AI uses. If you do allow model development, require a description of the model and its intended use, your review before deployment, restrictions on commercial licensing, and limits on use of the model to make decisions about the individuals in the data.
- No unilateral expansion. Changes to the recipient's terms of service or privacy policy cannot expand its rights under the agreement. Only a signed amendment can.
There are good reasons to say yes sometimes. A university partner building an early warning model for housing instability, with careful governance and community input, may produce something genuinely valuable for the population you serve. The point of the clause is not to block every AI use but to make it a deliberate decision, with the people affected in mind, rather than something that happens because the agreement was silent. If your organization has an AI data governance policy, the clause should reflect it, and your client consent language should match what your agreements actually permit. Our article on informed consent in the age of AI covers that side of the equation.
How AI Speeds Up Drafting and Review
Agreement work is a good fit for AI because so much of it is structured comparison: does this document contain the things it should, and do its terms match what we want? Large language models are strong at reading long documents, extracting provisions, comparing them to a reference, and explaining the differences in plain language. They are weaker at knowing what the law requires in your state today, judging how much risk is acceptable, and noticing what a clever drafter has deliberately left ambiguous. A good workflow plays to the first set of strengths and assigns the second set to people.
The foundation for all of this is a standard positions document: a few pages, approved by leadership and ideally reviewed by counsel, that lists each clause from the section above with your preferred language, acceptable fallbacks, and dealbreakers. Without it, AI can only compare a partner's draft to generic norms. With it, AI can compare the draft to your norms, which is far more useful. Building this document is itself a task AI can help with: give a model your three best existing agreements, the SPPO checklist, and the AISP report's sample terms, and ask it to propose a first version organized by clause. Then have your team and counsel edit it.
1. First drafts from a template
Give the model your approved template, your standard positions, and a short term sheet for the new partnership: parties, purpose, data elements, legal basis, term. Ask it to produce a draft that fills the template, flags every place where it had to make an assumption, and lists questions for the program lead. You get a working draft in minutes instead of days, and the assumptions list becomes the agenda for your internal review.
2. Comparing a partner's paper
Large partners, especially hospitals, universities, and agencies, usually insist on their own template. Ask the model to map each of their provisions to your standard positions in a table: their language, your position, whether it matches, is weaker, or is missing, and a suggested redline. This turns a long read into a focused list of issues to negotiate.
3. Flagging missing clauses
Missing terms are easier to overlook than bad ones. Run every incoming agreement against a clause checklist and ask specifically for what is absent: no AI training language, no destruction certification, no breach deadline, no subcontractor flow-down. For school agreements, include the FERPA required elements from the SPPO checklist so gaps are caught before signature.
4. Plain-language summaries
The staff who must follow an agreement rarely read it. Have AI produce a one-page "what this means for you" summary for program staff (what we may send, how, who to call if something goes wrong) and a separate summary of risks and obligations for leadership or the board. Have the agreement's owner check both against the source before distributing.
Prompts matter. Ask the model to quote the exact language it relies on and cite the section number for every finding, so a reviewer can verify each point in seconds. Ask it to separate "missing" from "present but weaker than our position" from "present and acceptable." And ask it to list anything it is uncertain about rather than resolving uncertainty on its own. A model that is instructed to show its work is much easier to check, and it is less likely to confidently describe a clause that is not there. For more on using AI to review legal documents generally, see our article on AI contract review for nonprofits.
Consider how this works with a university evaluator. The research office sends a twenty-five-page data use agreement. Your operations manager uploads it, along with your standard positions, to an approved AI workspace and asks for a comparison table. The output shows that the university's draft is strong on security, allows data retention "for as long as necessary for research purposes," says nothing about AI model training, and permits sharing with "research collaborators" without your approval. That is three clear negotiation points, each with the source language quoted. The manager forwards them to counsel with a short note, and the review that might have taken a week of back-and-forth takes a focused hour.
Keeping an Inventory of Active Agreements
An agreement nobody can find is almost as bad as no agreement at all. Many nonprofits cannot quickly answer basic questions: how many active data sharing agreements do we have, which partners hold client data right now, which agreements expire this year, and which allow AI use? An inventory answers those questions, and the 2025 edition of the AISP report includes a checklist for conducting a data sharing agreement inventory, which is a useful starting structure.
AI makes the initial build much easier. Collect every agreement you can find, from shared drives, email attachments, and program staff's files, and use a model to extract a standard set of fields from each one into a spreadsheet. A person should spot-check every row against the source, especially dates and data elements, because extraction errors in an inventory compound over time. Once it exists, the inventory can drive simple automations: calendar reminders ninety days before each renewal, a quarterly report to leadership, and a trigger to request destruction certificates when projects end.
Fields to track for each agreement
One row per agreement, owned by a named staff member
- Partner, agreement type (DSA, DUA, BAA, MOU), and internal owner
- Direction of data flow (we send, we receive, or both) and the purpose in one sentence
- Data categories and whether any regulated data is involved (PHI, education records, Part 2, HMIS)
- Legal basis and where client consent is documented, if applicable
- AI training permitted, prohibited, or silent
- Effective date, expiration, renewal terms, and notice deadline
- Breach notification deadline and partner incident contact
- Destruction requirement and date the certificate was received
The "silent" value in the AI training column is where an inventory pays for itself. Sort by it, prioritize partners who hold the most sensitive data, and send amendments or side letters adding AI use language at renewal. For coalitions and shared measurement efforts, a common inventory across members also clarifies who holds what, which matters when a partner leaves. Our article on AI in collective impact initiatives discusses the governance side of shared data, and for homeless services, our guide to improving HMIS data quality with AI covers the operational side of working with HMIS data.
Where AI Stops and Counsel Starts
Using AI on data sharing agreements creates its own data sharing question. The agreement itself is often confidential, and a draft may include descriptions of client populations, data elements, and security arrangements that you would not want exposed. Some partner agreements explicitly prohibit sharing their terms with third parties. Pasting a confidential agreement into a consumer AI tool that may retain or train on inputs can breach the very obligations you are trying to review. Use only tools your organization has approved, on plans whose terms prohibit training on your content and limit retention. Our article on shadow AI in nonprofits explains how to keep staff from reaching for unapproved tools under deadline pressure.
Never put actual client data into an AI tool to "test" what an agreement permits, and do not upload data exports or sample records along with the agreement. The AI only needs the contract text, your standard positions, and the term sheet.
What AI cannot do for you
- Give legal advice or tell you whether a disclosure is lawful in your state
- Reliably know recent regulatory changes, such as the updated Part 2 rule
- Decide how much risk your organization should accept
- Catch every deliberately vague or cross-referenced provision
- Weigh community trust and the expectations of the people in the data
When to bring in counsel
- Any agreement involving PHI, education records, Part 2 records, or HMIS data
- Any request to permit AI model training or commercial use of derived data
- Indemnification, liability caps, or insurance requirements
- Data about minors, immigration status, or other high-sensitivity populations
- Your standard positions document and templates, at least once a year
The best way to think about AI here is as preparation for counsel, not a replacement for counsel. When your attorney receives a partner's draft along with a clause-by-clause comparison, quoted language, and a short list of open questions, their time goes to judgment rather than reading. Many nonprofits work with pro bono attorneys or bar association programs whose time is limited, and arriving prepared is how you make the most of it. The same discipline applies in reverse: if counsel's redlines come back, AI can help you explain them to the partner's program staff in plain language, so negotiations move faster.
A 60-Day Plan to Get Your Agreements in Order
You do not need to renegotiate every agreement at once. The goal is to know what you have, set clear standards for new agreements, and close the most important gaps on a sensible schedule. A small team can make real progress in two months.
Sequenced steps
Adapt the timing to your size and the number of partners
- Weeks 1-2: Find everything. Ask every program lead which outside organizations receive or send client data. Collect agreements, and list relationships that have none.
- Weeks 3-4: Build the inventory. Use an approved AI tool to extract key fields into a spreadsheet, then verify every row. Flag expired agreements, missing agreements, and those silent on AI.
- Weeks 5-6: Write standard positions. Draft your clause-by-clause positions with AI assistance, drawing on the SPPO checklist and AISP guidance, and have counsel review them along with a base template.
- Week 7: Set the review workflow. Document who drafts, which prompts and tools are used, what triggers counsel review, and who signs. Train the handful of staff who will run it.
- Week 8: Prioritize remediation. Rank gaps by data sensitivity. Put formal agreements in place where there are none, and schedule AI use amendments for the highest-risk partners at their next renewal.
Brief your board along the way. Data sharing agreements are a governance matter as much as an operational one, because they determine your organization's exposure if a partner fails. A short annual report showing the number of active agreements, the share that address AI use, and any incidents gives the board visibility without drowning them in contract detail. It is also the kind of evidence funders and partners increasingly ask for when they assess whether a nonprofit can be trusted with shared data.
Conclusion
Data sharing is how modern nonprofit work gets done. Referral networks, school partnerships, health collaborations, coordinated entry, and outcome evaluations all depend on information moving between organizations. The people whose information moves rarely see the agreements that govern it, which makes your organization the steward of their trust. A clear, specific, enforceable agreement is how you keep that trust when the data is no longer in your hands.
The fundamentals have not changed: a precise purpose, minimum necessary data, a documented legal basis, strong security, prompt breach notice, firm limits on re-disclosure, and a real plan for destruction. What has changed is that every agreement now needs to say whether data may be used to train AI models, because silence will increasingly be read as permission. And what has improved is your ability to do this work efficiently, with AI tools that can draft from templates, compare partner terms against your positions, find missing clauses, summarize obligations, and keep your inventory current.
Use those tools for the heavy reading, keep them away from confidential material unless they are approved for it, and keep the judgment calls with people, including counsel when regulated data is involved. Start with an inventory. You will likely find that a few hours of organized effort closes gaps that have been sitting open for years.
Get Your Data Partnerships on Solid Ground
We help nonprofits inventory data sharing relationships, build standard positions and templates, and set up AI-assisted review workflows that make counsel's time count.
