Three years into his tenure, the chief digital officer of a mid-sized national government, Daniel Okonkwo, had a problem that looked like success. Every ministry had its own AI projects. The health ministry had a vendor. The tax authority had a different vendor. A national university ran a strong research lab that nobody in government talked to. A handful of civic-tech nonprofits were building tools to watchdog the very systems the ministries were deploying. He had a great deal of AI activity and almost no ecosystem: no shared rails, no flow of talent or standards between the parts, and four separate groups quietly working at cross-purposes. When he asked whether the customs agency could reuse the tax authority's fraud model, the answer was no. Nothing connected.

An AI ecosystem is what you have when those parts reinforce each other instead of duplicating each other. It is not built top-down and it cannot be, because most of the actors do not report to you. It emerges from the interaction of agencies with mission problems, vendors with capacity, researchers with ideas, and civil society with oversight, and your role as a leader is to shape the conditions under which that interaction is healthy rather than wasteful. This lesson is for the leader whose job is no longer to run one program but to shape the environment dozens of programs live in, and it uses Daniel's fragmented landscape as the raw material.

What an ecosystem actually is

An ecosystem is not an org chart. It is a set of independent players, agencies, vendors, universities, and civil society, connected by shared infrastructure, shared standards, and the free movement of talent and ideas. The test is reuse and reinforcement. In a real ecosystem a model built by one agency can be adopted by another, a graduate trained in a university lab can flow into government and back out, a vendor's tool meets a standard every agency recognizes, and a critic's finding makes the next system better rather than merely embarrassing the last one.

That gives you a diagnostic question that is easier to answer honestly than any maturity score: when one part of the system gets better, does anything else get better automatically? If the answer is no, you have activity rather than a system, and the activity may still be growing. Daniel's landscape was expanding steadily, which is exactly what made the problem hard to name. More projects, more vendors, and more spend all read as momentum on a slide. None of them produced a single instance of one agency benefiting from another's work.

The four players and what each needs

Each part of the ecosystem plays a distinct role, and each fails in a distinct and predictable way when the connective tissue is missing. Knowing the specific failure mode of each player is what lets you tell whether a gap is a capacity problem, which money can fix, or a connection problem, which money usually makes worse by funding more disconnected activity.

• Agencies are the demand side and the proving ground. They hold the mission problems and the data. In isolation they reinvent the same wheel repeatedly and lock themselves into incompatible vendor stacks, each of which was a reasonable local decision.
• Vendors and industry bring capacity and speed. Without shared standards they fragment the government into proprietary islands and capture agencies through switching costs that nobody priced at the time of purchase.
• Academia brings rigor, independent evaluation, and a talent pipeline. Without connective tissue to government its research never meets a real public problem, and its graduates never consider public service because no path into it was visible to them.
• Civil society brings legitimacy and accountability. Excluded, it becomes a pure adversary in court and in the press. Included, it functions as an early-warning system, though it remains free to oppose you and should be.

Build shared rails, not shared mandates

You will not get four ministries onto one platform by decree, and you should not spend your authority trying. Build the connective infrastructure that makes cooperation the easy choice instead: a common data-sharing framework with clear privacy rules, a shared model registry where agencies publish what they have built and what it is for, a set of reference architectures, and a recognized standard for what tested means. Rails work because they lower the cost of the behavior you want, which is a more durable mechanism than an instruction that expires with the person who issued it.

Anchor the standard in something credible so that a vendor's claim and an agency's requirement can meet. The AI Risk Management Framework published by the National Institute of Standards and Technology is widely used for this, because it gives public and private actors a common vocabulary for risk. Be precise about what that buys. The framework is voluntary and non-binding unless a contract or a rule makes it a requirement, and a vendor asserting alignment with it has made a claim rather than passed a test. Shared standards make claims comparable. Verification is a separate job that someone in your ecosystem still has to do.

Use procurement as ecosystem policy

How a government buys AI shapes the entire vendor landscape, and it does so whether or not anyone intends it. If every contract demands open standards, data portability, and the right to take your models and data elsewhere, the market has a strong reason to organize around interoperability. If contracts tolerate lock-in, the market will deliver lock-in efficiently and at scale. Daniel's most powerful lever turned out not to be a new platform but a rewrite of standard procurement clauses, so that every vendor entering the government had to meet shared interface and portability requirements.

Treat the clause as the beginning rather than the end. A portability requirement in a contract is a promise about a future event, and organizations discover the difference between a promised export and a usable one at the worst possible moment. Exercise the exit before you need it: require a test extraction during the contract, in a format your next system can actually ingest, and check that model artifacts, training data rights, and operational logs come with it. Procurement is the quiet constitution of an ecosystem, and like any constitution it means what is practiced rather than what is written.

Make talent flow in both directions

The scarcest resource in government AI is people, and ecosystems generate them where isolated agencies merely compete for them. Build the channels deliberately: fellowship programs that bring university researchers into agencies for a defined term, secondments that send civil servants into labs and companies, and visible pathways for graduates to begin a career in public service rather than discovering it a decade later. A government that is only ever a buyer of talent will always be short of it. One that helps grow and circulate talent has a renewable supply and, just as importantly, alumni in every other part of the ecosystem.

The return flow matters as much as the inbound one, and leaders often resist it. A civil servant who spends a year in a research lab or a company is not a loss to the agency; the knowledge comes back, and so usually does the person. Design the terms so that circulation is normal rather than exceptional: clear return rights, roles that are worth returning to, and honest handling of the conflict-of-interest questions that arise when people move between buyers and sellers. Those questions are answerable, but only if you answer them in advance instead of case by case under scrutiny.

Bring civil society inside the tent early

The cheapest accountability is the kind that happens before launch. Standing advisory relationships with civic-tech groups, academic ethicists, and community organizations surface design problems while they are still cheap to fix, which is the entire argument for the practice. Daniel's largest critic became a standing reviewer of the model registry, and the first problems that reviewer raised were ones his own teams had seen and deprioritized. External attention changes what an organization considers urgent, which is a real effect and not a soft one.

Be clear-eyed about what engagement does not do. It does not convert a critic into an ally, and it must not be designed to. A group that reviewed your design retains every right to litigate, campaign, or testify against the system you eventually ship, and an engagement program that quietly aims to neutralize opposition will be recognized as such and will cost you more than silence would have. Engagement gives you earlier information and gives the public an independent voice with standing. It does not give you cover, and treating a review as endorsement is how agencies end up surprised by a partner.

The conditions a government supplies

Ecosystems grow where governments supply things the market cannot supply for itself. Estonia's digital government is the frequently cited case: it built capability by creating the conditions for a domestic technology sector, legal certainty about what was permitted, real procurement opportunity for smaller firms, a talent base, and shared infrastructure. Startups built solutions and government adopted them. The order matters. The conditions came first and the market response followed, which is the opposite of the sequence most governments attempt when they announce a strategy and wait for an industry to appear.

Patient capital is the other thing only a public body reliably provides. The Defense Advanced Research Projects Agency in the United States is the standard example, funding high-risk research on long horizons where a near-term application is not required to justify the work. A meaningful share of that portfolio fails, which is the design rather than a defect, and the spillovers have repeatedly landed well outside the original purpose. The transferable lesson is not the agency's structure but its risk posture: government money can fund the work whose payoff is too uncertain, too distant, or too diffuse for a market to finance, and that is a genuine comparative advantage.

A third pattern is worth naming even though the connective work is harder. Governments can create a structured route by which civil-society organizations propose public-benefit AI projects, with public funding and access to shared infrastructure attached. It widens the set of people who get to decide what AI gets built for, which otherwise defaults to whoever already has capital. It is not automatically democratizing: if the application process demands professional grant-writing capacity, it will select for well-resourced organizations and reproduce the existing distribution with public money. Design the intake for the applicants you say you want.

Making it last past your tenure

The graveyard of government innovation is full of ecosystems that depended on a single champion and dissolved when that person left. Most governments also underestimate how long this work takes, and quick fixes rarely survive contact with the second budget cycle. Sustainability is therefore a design requirement rather than an afterthought, and the parts of your design that will still exist in a decade are the parts that were made someone's explicit job and someone's budget line.

• Institutionalize the connective tissue. Put the shared registry, the standards, and the procurement clauses into law or durable policy rather than into a memo a successor can quietly stop enforcing.
• Fund the commons. Shared infrastructure needs an owner and a budget. Shared things with no owner decay, and the decay is invisible until the moment someone needs them.
• Create reasons to participate. Agencies that contribute reusable assets should get recognition, priority support, or funding advantages. Make joining the rational choice for a manager whose own targets have nothing to do with your ecosystem.
• Protect the effort across leadership change. Write down why the arrangement exists, not only what it requires, so that the next leader inherits reasoning rather than rules whose purpose has been forgotten.

Measuring ecosystem health, not project counts

The metric most governments report is the number of AI projects underway, which is precisely the number that grows fastest when the ecosystem is failing. Measure the connections instead: reuse rates, meaning how often an asset built by one agency is adopted by another; talent flows in both directions; the share of vendors meeting shared interface and portability requirements; and the proportion of contracts where an exit has actually been tested. These are harder to collect and they are the ones that tell you whether the parts are reinforcing each other.

Daniel's turning point was modest and structural. He stood up a small ecosystem office with a budget, a legal mandate, and three deliverables: a public model registry, mandatory procurement interoperability clauses, and a fellowship rotating ten university researchers through agencies each year. Within eighteen months the customs agency was running an adapted version of the tax authority's fraud model, two fellows had taken permanent government roles, and the largest civic-tech critic had become a standing reviewer. Nothing he built was flashy. All of it connected things that had been separate, which is the only work that compounds.

Use the questions below as a periodic check on whether the connective tissue is real. Ask them about specific systems rather than in general, because the general answer is always more flattering than the specific one, and ask them of someone other than the person responsible for the answer being yes. Each question is written so that a truthful answer requires an instance rather than an intention, which is the only reliable defense against a review that confirms what everyone already believes. Where an instance does not exist yet, the honest entry is not yes, and recording it as no is what turns the checklist into a work plan instead of a certificate.

• Shared rails. Is there a model registry, a data-sharing framework, and a recognized standard that all four players can use, and has anything actually been reused through them?
• Procurement as policy. Do standard contracts require open standards and portability, and has an exit been tested rather than merely promised?
• Talent channels. Are fellowships, secondments, and entry pathways moving people between sectors in both directions, with return rights and conflict rules written down?
• Civil society inside. Are accountability partners engaged before launch, with their independence intact and their findings published?
• Sustainability. Is the connective tissue in durable policy, with a named owner, a budget line, and incentives that outlast any one leader?
• Health metrics. Are you tracking reuse, talent flow, and standards adoption rather than the count of projects underway?

Anti-Patterns to Avoid

• Counting activity as ecosystem. Reporting project counts, vendor numbers, and total spend as evidence of a healthy ecosystem. Those figures rise fastest during fragmentation, which is why they are reassuring exactly when they should not be.
• The mandated platform. Ordering every ministry onto one system. You will spend your political capital on compliance theater and produce shadow systems, because the ministries that resisted still have their own missions and budgets.
• The portability clause nobody exercised. Treating a contractual right to export data and models as an exit. A promised extraction is not a tested one, and the difference is discovered during a transition, under time pressure, with the incumbent vendor's cooperation now optional.
• The standard as certification. Reading a vendor's claim of alignment with a voluntary risk framework as a passed test. Frameworks make claims comparable. Someone in your ecosystem still has to check whether the claim is true.
• Engagement designed to neutralize. Building advisory relationships whose real purpose is to convert critics into cover. Participants recognize this quickly, the relationship ends publicly, and you are worse off than if you had never asked.
• One-way talent programs. Bringing researchers in while refusing to let civil servants out. Circulation is the mechanism; a program that only imports people is recruitment with a better name and it dries up when the novelty does.
• The unowned commons. Standing up a shared registry or platform without a named owner and a budget line. It works for a year, decays quietly, and is discovered to be stale by the first agency that trusts it.
• The champion dependency. Running the whole arrangement on one leader's relationships and personal insistence. Write down the reasoning, put the mechanisms into durable policy, and assume your successor will inherit no goodwill from your tenure.

Practice Prompts

• Run the reuse test. Pick the two best AI assets built anywhere in your government recently and trace whether any other body has adopted either. If neither, document exactly where the adoption stalled: discovery, licensing, technical incompatibility, or trust.
• Audit one contract for exit. Take a live AI contract and establish what you would actually receive if you terminated it next month, including model artifacts, data, rights, and logs. Compare that with what the clause promises.
• Map the four players in your domain. Name the specific agencies, vendors, research groups, and civil-society organizations in one policy area, and draw the connections that currently exist. The empty spaces in the drawing are your work plan.
• Design one rail. Choose a single piece of connective infrastructure your ecosystem lacks, and specify its owner, its budget line, the policy instrument that makes it durable, and the incentive that makes an agency use it voluntarily.
• Write the successor briefing. Draft the document explaining to whoever holds your job after you why each ecosystem mechanism exists and what breaks if it is removed. Anything you cannot justify in that document is unlikely to survive you.

Reflection

Think about the AI work happening across your government right now and ask which of it would still function if you personally stopped attending meetings. Consider where the genuine connections are, the ones that operate through policy, budget, and incentive rather than through your relationships, and be honest about how few of them there are. Then ask the harder question: of the connective mechanisms you have built or inherited, which have ever actually been used, and which exist mainly as evidence that someone once addressed the problem. That second list is where your next quarter's work is.

Glossary

• AI ecosystem. A set of independent actors, agencies, vendors, academia, and civil society, connected by shared infrastructure, standards, and talent flows such that improvement in one part benefits others.
• Shared rails. Connective infrastructure that lowers the cost of cooperation, such as a model registry, a data-sharing framework, or reference architectures, as distinct from a mandate to use one platform.
• Model registry. A published inventory of the models an organization or government has built, describing what each does, who owns it, and on what terms others may adopt it.
• Vendor lock-in. The condition in which switching suppliers is prohibitively costly because data, models, or interfaces cannot practically be moved elsewhere.
• Portability. The practical ability to move data, models, and operating artifacts to another provider, which is established by testing an extraction rather than by a contractual clause.
• Secondment. A temporary placement of a person from one organization into another, with defined terms and a right of return, used to circulate skills between sectors.
• Patient capital. Funding committed on long horizons without requiring near-term application or return, which public bodies can supply where markets will not.
• Reuse rate. The share of assets built by one body that are subsequently adopted by another, the most direct available measure of whether an ecosystem exists.

Related Lessons

• Building Innovation Ecosystems extends this material into the wider innovation environment surrounding government technology work.
• Public-Private Partnerships for Government AI covers the structures through which government and industry share risk, cost, and control on specific undertakings.
• Public-Private Innovation at Scale takes up what changes when those partnerships move beyond pilots.
• Academic and Research Collaboration develops the university side of the ecosystem, including joint research and evaluation arrangements.
• AI Talent Pipeline: Education System Alignment addresses where the people in your talent channels come from in the first place.
• Shared Services and Infrastructure Models goes deeper on funding and governing the commons that ecosystems depend on.
• AI Interoperability Across Agencies covers the technical and standards work that makes reuse possible in practice.
• Technology Transfer and Commercialization examines how publicly funded work reaches the wider economy.

Closing

Nothing in Daniel's second year was innovative. A registry, a set of contract clauses, a small office with a budget, a fellowship, and a standing reviewer who did not like him very much. What made it an ecosystem rather than another program was that each piece raised the value of the others: the registry only mattered because the procurement clauses made models portable, the clauses only held because someone owned them, and the fellows carried both into agencies that had never heard of either. Ecosystem work rarely produces a moment worth announcing. It produces the condition in which everything else your government builds is worth more than it cost.

Key Takeaways

• Activity is not an ecosystem. The test is whether one part improving makes others improve automatically, and project counts rise fastest when fragmentation is worst.
• Shape conditions rather than issuing commands. Ecosystems emerge from interaction between actors who mostly do not report to you, so lower the cost of the behavior you want instead of mandating it.
• Build shared rails. A registry, a data-sharing framework, and a recognized standard make cooperation the easy choice without forcing ministries onto one platform.
• A shared standard makes claims comparable, not true. Voluntary risk frameworks give every player one vocabulary; verifying a vendor's claim against one remains someone's job.
• Procurement is the quiet constitution. Open standards and portability requirements organize the vendor market, but only an exit you have actually tested is an exit you have.
• Talent must circulate in both directions. Fellowships and secondments turn the scarcest resource into a renewable one; a program that only imports people is recruitment with a better name.
• Bring critics in early without expecting cover. Engagement buys earlier information and public legitimacy; it does not and should not neutralize anyone's right to oppose you.
• Supply what markets will not. Legal certainty, procurement opportunity, shared infrastructure, and patient capital for work whose payoff is too distant to finance privately.
• Design for life after you. Durable policy, a named owner, a budget line, and written reasoning are what survive a change of leadership.

Frequently Asked Questions

Can a single agency build an ecosystem, or does this require central authority? A single agency can build a great deal of it, because most of the mechanisms in this lesson are things one organization controls: its own contract clauses, its own registry entries, its own fellowship and secondment terms, and its own engagement practices. What central authority adds is the ability to make those things standard rather than local. Waiting for the mandate is usually a way of postponing work that would have created the demand for the mandate.

How do we avoid a shared registry that nobody uses? Give it an owner with a budget, make contribution a condition of something people already want, such as funding, platform access, or approval support, and reduce the cost of publishing to something a team can do in an afternoon. Then measure adoption rather than entries. A registry with many entries and no reuse is a catalog of work nobody could figure out how to adopt, which is a discovery and integration problem rather than a publishing one.

Does requiring open standards drive good vendors away? Some will decline, and those are disproportionately the ones whose commercial model depends on switching costs. The larger risk runs the other way: portability and interface requirements written by people who do not understand the technology can be onerous without being useful, which excludes capable smaller firms while sophisticated incumbents comply on paper. Write the requirements with practitioners, keep them outcome-based, and test compliance during the contract rather than at the end.

How long before an ecosystem shows results? Longer than a political cycle for the compounding effects, which is why the mechanisms have to be made durable early. Some signals arrive quickly, though: a first reuse, a first tested exit, a first fellow who stays. Track those deliberately and report them, because the early wins are what fund the survival of the arrangement until the structural effects arrive. Governments that promise transformation and report nothing for years on end tend not to get another term.