The research agreement between Isabel Fuentes's state Medicaid agency and the state's flagship public university took eleven months to negotiate. Isabel is the Chief Data Officer at an agency serving 2.4 million beneficiaries, and the university's health informatics department had a machine learning model that could predict which beneficiaries were at high risk of emergency department overutilization, the kind of insight that, acted on by care coordinators, could reduce both avoidable hospitalizations and program costs. The science was solid. The partnership almost failed on a clause about data ownership.

The university wanted to retain rights to the model and to publish findings using the agency's data. The agency's legal team was not certain the Medicaid data could be used for academic publication under federal privacy rules, and was deeply uncertain about a university retaining intellectual property built on state government data. Eleven months later the agreement was signed. The partnership has since produced a model that is in production, has been presented at two national conferences, and has informed care management protocols that the Centers for Medicare and Medicaid Services has highlighted as a model for other states. This lesson is about building the structures that make that outcome reachable without eleven months of attrition.

Why Academic Partnerships Are Different from Vendor Relationships

Government agencies are accustomed to vendor relationships: a vendor builds something to agency specifications, the agency pays for it and owns it. Academic partnerships work differently, and the difference is not a matter of degree. Universities bring research capacity, graduate student labor, access to methods that are years ahead of commercial practice, and publication venues that can legitimize an agency's AI approach with peer-reviewed evidence rather than a vendor's case study. That combination is not purchasable through a normal acquisition, which is exactly why the agreement in front of you looks unfamiliar.

Universities also have incentives that sometimes conflict with agency needs. Publication, intellectual property development, and student training are how an academic department measures itself and how its faculty advance, and none of those align automatically with an agency's need for data control, operational confidentiality, and clear ownership of tools running in production. Managing that tension productively requires understanding both sides of the equation and designing agreements that satisfy the legitimate interests of each without forcing either party to compromise a core obligation.

The mistake that costs the most time is reading the conflict as bad faith. A university counsel who will not sign away rights in a generalizable method is not trying to take something from you; they are protecting the thing their institution exists to produce. An agency lawyer who cannot approve publication using beneficiary data is not obstructing science; they are enforcing a rule that binds them. The conflict is structural, it is predictable, and predictable conflicts can be designed for in advance instead of discovered one clause at a time.

What Consumed Eleven Months

It is worth being specific about where Isabel's year went, because the pattern repeats almost everywhere. Three questions took nearly all of it. Could Medicaid data lawfully support academic publication under the federal privacy rules that govern it? Who would own a model built from state government data using university methods? And what could the university publish, in what detail, about an operational program that had not agreed to be described in public?

None of those were novel questions. All three have standard resolutions that this lesson describes, and the agency arrived at each of them eventually. What made them expensive was that they were raised sequentially, by different offices, at different stages, with each answer requiring a fresh round of review. The privacy office saw the arrangement after the research design was fixed. The federal program office saw it after the state privacy office had already approved. Every reordering of that sequence would have saved months.

The resolutions themselves were unremarkable, which is the point. Publication was handled by a review window rather than a prohibition. Ownership was handled by a licensing split rather than by one side conceding. The privacy question was answered by the terms that already governed the data rather than by anything the two parties invented. Every one of those answers is described below, and every one of them was available on day one to an agency that had written down what it concluded last time.

The lesson to take is not that agreements are slow. It is that the slowness is concentrated in a small number of recurring questions whose answers your agency can determine once and reuse. Everything below is organized around that idea: know which instrument you are in, know which of its terms are genuinely negotiable, and know which are already settled by someone above you before you spend a month negotiating them.

University Collaboration Structures

Three collaboration structures cover most government and university AI partnerships, and knowing which one you are in tells you which terms are yours to negotiate.

Data use agreements. A data use agreement governs how university researchers may access and use agency data for specific research purposes. It specifies the data elements that can be accessed, the security controls the university must maintain, the purposes for which the data can be used, restrictions on publication that might identify individuals, and the disposition of the data at the end of the project. Each of those five is a place where a partnership can stall, and each has a defensible default your privacy office can settle in advance rather than under deadline.

Take them one at a time, because each carries a practical consequence that only appears later. The data elements clause determines what research is possible at all; a minimized extract that satisfies the privacy review may also remove the variable the model needed, so the researchers have to be in that conversation rather than receiving its output. The security controls clause determines where the work can physically happen, and a university environment that cannot meet your controls means the analysis moves to yours, which is a resourcing decision disguised as a legal term. The purpose limitation determines whether a second study needs a new agreement or an amendment, which is the difference between weeks and months when a promising finding suggests a follow-up.

The publication restriction and the disposition clause are the two most often skimmed and the two most likely to generate an argument at the end. Restrictions protecting against identification of individuals belong here rather than in the operational review window, because they are a privacy obligation rather than an agency preference. And disposition, meaning what happens to every copy of the data when the project closes, is worth settling while both parties are enthusiastic; graduate students leave, laptops get reimaged, and a clause written in month two is far easier to enforce than a conversation started in month thirty.

For Isabel's agency the data use agreement was the foundational document. Nothing else could proceed until it existed, and it required approval from the state privacy office, the agency's legal team, and the federal Centers for Medicare and Medicaid Services regional office before a single record moved. That approval chain is the single strongest argument for building data use agreement templates that have already been reviewed and approved for standard research types, such as predictive modeling, program evaluation, and quality measurement. A template that three offices have already blessed converts the longest phase of a partnership into a review rather than a negotiation, and it is the highest-leverage investment available in this whole domain.

Sponsored research agreements. A sponsored research agreement governs a funded research project where the agency is the sponsor and the university is the researcher. It specifies the research scope, deliverables, timeline, budget, intellectual property rights, publication rights, and data handling. The intellectual property clause is typically the most contested provision, and the shape of the dispute is consistent across institutions: universities prefer to retain rights in tools developed with university resources, while agencies prefer ownership of tools built with government data and government funding.

The most common resolution is a licensing arrangement rather than a winner. The university retains rights in the research methods and any generalizable algorithms, and the agency retains a perpetual, royalty-free license to use the tool in its own operations. That split works because it gives each institution the thing it actually needs: the university keeps what supports future research and faculty careers, and the agency keeps unrestricted operational use of the specific tool it paid for. Publication rights are handled separately, through a pre-publication review window, typically 30 to 60 days, during which the agency can request redaction of sensitive operational details before a paper is submitted.

The other terms in a sponsored research agreement look routine and are not. Scope and deliverables are where the technology transfer requirement described later in this lesson either appears or does not, and adding it after signature is a renegotiation. Timeline has to accommodate an academic calendar in which the people doing the work graduate, take summers on other funding, and disappear for a semester of teaching, which is a normal feature of the institution rather than a performance problem. Budget determines whether anyone is paid to make the output usable. Data handling has to match the data use agreement exactly, because two documents describing the same obligation in different words is an argument waiting for a reason.

Cooperative agreements and grants. Where a federal agency is the funder rather than your own agency, the data use and intellectual property terms are often set by the federal funding terms rather than negotiated bilaterally. Medicaid research funded through Centers for Medicare and Medicaid Services Innovation Center grants, for example, operates under that agency's standard data use and intellectual property provisions. Understanding the applicable federal terms before negotiations begin saves months and prevents agreements that are later found to conflict with federal requirements, which is the worst discovery available in this work because it invalidates a negotiation that both parties considered finished.

Research Funding Pathways

Government agencies can reach research funding mechanisms that are not available to private organizations, and the most useful of them cost the agency very little cash. This is the part of academic partnership that agency staff most often overlook, because it requires thinking of the agency as a contributor to someone else's funding application rather than as a buyer.

Federal agency research grants. The National Science Foundation, the National Institutes of Health, the Defense Advanced Research Projects Agency, and the Advanced Research Projects Agency for Health fund research that may directly benefit government AI applications. A state health agency partnering with a university on AI-assisted disease surveillance may be able to support the university's application for federal research funding. The government partner's involvement often strengthens the application materially, because it supplies access to real-world data and a credible pathway to implementation, and those are exactly the two things most academic applications cannot demonstrate. The practical consequence is that your agency should be in the conversation while the proposal is being written rather than asked for a letter of support the week before it is due, since the parts you can strengthen are structural rather than decorative.

State innovation funds. Many states maintain economic development or technology innovation funds that can support government and university research partnerships, particularly where the research carries economic development implications, such as AI tools that improve workforce outcomes or reduce Medicaid costs and free state general fund dollars for other priorities. The eligibility framing matters here: the same project described as a research collaboration may not qualify, while described in terms of its economic effect it may.

In-kind contributions. Agency data is often the most valuable asset a government partner brings to a collaboration. Properly structured, data access functions as an in-kind contribution that offsets the agency's cash cost of the partnership and that universities genuinely value, because government administrative data at scale is rarely accessible to academic researchers through any other channel. Treat it as a contribution recognized in the agreement rather than as something you simply hand over, and the negotiation over everything else changes character.

Technology Transfer from University to Agency

Academic research produces findings. Agencies need tools. The gap between the two is the technology transfer challenge, and it is the most common point of partnership failure. A university team that produces a peer-reviewed paper describing a predictive model has not produced something a state IT team can deploy in a production case management system. Production deployment requires cleaned and documented code, integration specifications, testing and validation protocols, and ongoing support arrangements, and none of those are standard outputs of an academic research project or standard lines in a research budget.

Isabel's agency addressed this by building a technology transfer requirement into the sponsored research agreement itself. The university team was required to produce not only a research paper but a deployment-ready codebase with documentation meeting state IT standards, a technical specification for integration with the agency's case management system, and three months of post-deployment technical support. The university pushed back initially, and its objection was reasonable on its face: these are software engineering tasks, not research tasks, and no faculty member is evaluated on them.

The resolution was a budget line rather than an argument. The agency funded a graduate student research engineer whose specific role was production readiness, at a cost of $42,000 for the year, and the result was a model in production within four months of the research being completed. That is the whole lesson in one figure. The work does not become unnecessary because nobody budgeted for it; it simply lands on whoever is least equipped to refuse it, usually months later, usually after the researchers have moved on.

Measure the partnership accordingly. A model in production serving 2.4 million beneficiaries is the goal, and publications are evidence of rigor rather than the end result. Track the elapsed time from research completion to production deployment and hold both parties accountable to it, because that single number is where an unfunded transfer gap becomes visible while there is still time to close it.

Building the Student Pipeline

Academic partnerships are also talent pipelines, and this is the return that never appears in the project budget. Graduate students and recent graduates who work on government AI research projects are among the most effective recruiting sources for agencies trying to build internal AI capacity. They already understand the agency's data, the policy context, and the specific problems the agency is trying to solve, and many have developed a genuine commitment to public service through the research itself rather than through anything a recruiter said to them.

There is a second advantage that agencies rarely name out loud. A multi-year research partnership is the longest and most honest evaluation you will ever get of a prospective hire. You have watched them work on your actual data, against your actual constraints, through the parts of a project that go badly, and you know how they behave when a result does not hold up. No interview loop produces that information, and no reference check substitutes for it. The candidate has equally had a long look at your agency, which is why the ones who accept an offer tend to stay.

Isabel's agency has hired three graduate student researchers from the Medicaid partnership into full-time positions over the past four years, and all three now hold senior data science roles. The agency's own assessment is that developing equivalent talent through training programs and open recruiting would have cost considerably more than the partnership that produced it. Design for that outcome deliberately rather than hoping for it: structure internships and fellowships with conversion in mind, keep students close to operational staff instead of isolated in a research track, and start the hiring conversation well before the dissertation is finished rather than after another employer has made an offer.

Anti-Patterns

• Running the partnership like a procurement. Specifying a deliverable, paying for it, and expecting to own everything is how vendor relationships work and it is not how universities work. An agency that opens with a demand for full intellectual property assignment has usually spent its first two months on a clause that a licensing structure would have settled in a week.
• Raising the recurring questions sequentially. Isabel's eleven months went largely to three predictable questions surfacing one at a time, from different offices, each answer triggering a fresh review cycle. Bring the privacy office, the legal team, and the federal program office to the same early conversation, because the alternative is paying for the same negotiation three times.
• Negotiating terms that are already settled above you. Where a federal grant or program funds the research, its data use and intellectual property provisions govern. Bilateral negotiation of a term you do not control produces an agreement that is later found to conflict with federal requirements, which is worse than no agreement because both parties believed it was finished.
• Treating the paper as the deliverable. A peer-reviewed publication describing a model is not a deployable tool, and no amount of goodwill converts one into the other. Documented code, integration specifications, validation protocols, and support arrangements are separate deliverables that have to be written into the agreement and funded.
• Using the pre-publication review as a veto. The window exists so an agency can request redaction of sensitive operational details. An agency that uses it to suppress unflattering findings will not get a second partnership, and word will reach every other department in the university before the paper does.
• Building the templates during the negotiation. A pre-approved data use agreement for a standard research type is worth more than any single clause you could win, and it can only be built when nothing is urgent. Agencies that never build one negotiate the same document from scratch every time and mistake the repetition for complexity.

Practice Prompts

• Identify the instrument. For a partnership you are considering, decide whether you are in a data use agreement, a sponsored research agreement, or a federally funded cooperative agreement or grant. Then list which terms you actually control and which are set above you.
• Draft the licensing split. Write the intellectual property clause you would propose: what the university retains in generalizable methods, and what operational license the agency takes in the specific tool. Test it by asking whether each institution gets the thing it genuinely needs.
• Map the approval chain. Name every office whose approval is required before data can move, in order, with the realistic review time for each. Then determine which of those reviews could run in parallel and which genuinely have to be sequential.
• Budget the transfer. For a partnership you are planning, write the technology transfer deliverables as contract line items: documented codebase, integration specification, validation protocol, and a support period. Then identify who is funded to produce each one.
• Find the funding you are not using. With a university partner, identify one federal research grant program and one state fund your collaboration could pursue, and determine what your agency's data access contributes to each application.

Reflection

• Of the research partnerships your agency has run, how many produced something running in production? What distinguished those from the ones that produced only findings?
• Where does your agency currently keep the answers to the recurring questions: privacy basis, intellectual property position, publication terms? If the answer is in a lawyer's memory, what happens when that lawyer leaves?
• What does your university partner get from the arrangement, in their own terms rather than yours, and would they describe the exchange as fair enough to renew?
• If a model arrived from a partner tomorrow, could your agency deploy and maintain it? If not, that gap is a technology transfer requirement you have not yet written.
• Which of your agreements were negotiated from a reviewed template, and which were negotiated from scratch because nobody had built the template yet?

Glossary

• Data use agreement. The instrument governing how researchers may access and use agency data for specific research purposes, covering the accessible data elements, required security controls, permitted purposes, publication restrictions, and disposition of the data at project end.
• Sponsored research agreement. The instrument governing a funded research project where the agency sponsors and the university performs, covering scope, deliverables, timeline, budget, intellectual property rights, publication rights, and data handling.
• Pre-publication review window. An agreed period, typically 30 to 60 days, during which the sponsoring agency may request redaction of sensitive operational details before a paper is submitted. A review of operational sensitivity, not an approval right over findings.
• Perpetual, royalty-free license. The usual resolution of the intellectual property dispute: the university retains rights in generalizable methods and algorithms, while the agency retains an unlimited, ongoing right to use the resulting tool in its own operations.
• In-kind contribution. Non-cash value a party brings to a partnership, most often agency data access, structured so that it offsets cash cost and is recognized in the agreement rather than given away informally.
• Technology transfer. The work of converting research output into something deployable and maintainable in a production environment, including documented code, integration specifications, validation protocols, and post-deployment support.

Related Lessons

• Data Sharing Agreements for AI treats the data use agreement in depth, including the security and disposition terms this lesson summarizes.
• Technology Transfer and Commercialization extends the research-to-production problem beyond a single agency partnership.
• Public-Private Partnerships for Government AI covers the parallel structures for industry collaboration, where the incentive conflicts take a different shape.
• AI Talent Pipeline: Education System Alignment develops the student pipeline theme into a deliberate workforce strategy.
• Leading Government-Wide AI Initiatives addresses partnership portfolios and cross-agency coordination at scale.
• Shared Services and Infrastructure Models covers the production environment a transferred model has to land in.

Closing

Isabel's partnership succeeded, and it cost eleven months it did not need to cost. Almost everything that consumed those months was knowable in advance: the federal terms that governed the data, the privacy approvals required before a record could move, the intellectual property resolution that both parties eventually accepted because it is the resolution nearly everyone accepts, and the technology transfer work that no research budget includes by default. None of it was a surprise in retrospect. All of it was a surprise at the time, because nobody had written any of it down after the last negotiation.

The durable investment is not any single agreement. It is the infrastructure between agreements: reviewed templates for standard research types, an intellectual property position you have already settled, a technology transfer line item you now budget automatically, and a written record of what was decided and why. Build that when nothing is urgent and nobody is waiting, and the next partnership starts where this one finished rather than where it began.

Key Takeaways

• Build data use agreement templates before you need them. A pre-approved template for standard research types is intended to bring time to first data access down from the eleven months Isabel's first agreement consumed to 6 to 8 weeks. Invest in that infrastructure between partnerships, not during them.
• Know which instrument you are in. Data use agreement, sponsored research agreement, or federally funded cooperative agreement or grant. The instrument determines which terms are yours to negotiate and which were settled before you sat down.
• Resolve intellectual property with a licensing structure, not a winner. The university retains rights in generalizable methods; the agency retains a perpetual, royalty-free license for operational use. A pre-publication review window of 30 to 60 days protects operational sensitivities without blocking academic output.
• Federal funding terms govern; negotiate within them, not around them. Where federal programs or grants apply, their data use and intellectual property provisions apply to the research agreement. Learn them first so you do not sign something later found to conflict with them.
• Budget a research engineer in every partnership. The gap between academic findings and production-ready tools is real and expensive to cross without dedicated resources. A graduate student research engineer funded by the agency is the most cost-effective bridge available.
• Agency data is a funding asset, not only a risk to manage. Structured properly, data access strengthens federal grant applications and counts as an in-kind contribution. Identify funding opportunities with your university partner before committing agency cash.
• Treat partnership graduates as a talent pipeline. Researchers who join the agency bring domain knowledge, existing relationships, and policy commitment that external recruits cannot match. Design internships and fellowships with conversion in mind.
• Measure partnership outcomes in production, not publication. A model serving 2.4 million beneficiaries is the goal; publications are evidence of rigor. Track elapsed time from research completion to production deployment and hold both parties accountable to it.

Frequently Asked Questions

Why do these agreements take so long?

Because they touch privacy, security, intellectual property, publication, and federal program rules at the same time, and each of those lives in a different office with its own review cycle. Isabel's eleven months went mostly to three recurring questions surfacing sequentially rather than together. The two levers that shorten the timeline are pre-reviewed templates for standard research types and early engagement with every office whose approval is required. Neither is available in the middle of an urgent negotiation, which is precisely why both have to be built between partnerships.

The university insists on retaining intellectual property. Is that a deal breaker?

Usually not, because the dispute is narrower than it sounds. Universities are generally protecting generalizable methods and algorithms that support future research and faculty careers. Agencies generally need unrestricted operational use of the specific tool. A license structure gives each party what it actually needs: the university keeps rights in the methods, the agency takes a perpetual, royalty-free license for operational use. Where federal funding terms apply, check them first, because they may already settle the question and make the negotiation unnecessary.

We got a published paper and nothing we can deploy. What went wrong?

The technology transfer work was never scoped, budgeted, or made a deliverable. Documented code, integration specifications, validation protocols, and post-deployment support are software engineering tasks that no research budget covers by default and that no university team will absorb without agreement. Write them into the agreement as deliverables and fund the person who does them. Isabel's agency funded a graduate student research engineer for the year at $42,000 and had a model in production within four months of the research finishing.

What does the university actually want from us?

Access to data they cannot get anywhere else, a research question worth a publication, funding where you can provide it, and a pathway to real-world implementation that strengthens their next grant application. Notice that only one of those costs you money. Understanding this list is what lets you negotiate from something other than the budget, and it is why agency data access, structured as a recognized in-kind contribution, changes the character of every other conversation in the agreement.