At a quarterly interagency meeting, the national AI lead for a federal department, Dr. Helen Asante, listened to four agencies present their new AI research priorities. By the third presentation she felt a cold recognition. Three of them were building, more or less, the same thing: a large language model fine-tuned to answer citizen questions about benefits. Three teams, three contracts, three vendors, roughly $4 million each, all solving the same problem in parallel without knowing about each other. None was talking to the agency that had already deployed exactly this two years earlier and learned why it was harder than it looked. This is the quiet, expensive failure mode of government research: not too little effort, but uncoordinated effort.

For a leader operating at the ecosystem level, the job is not to do the research. It is to make the collective research investment of many agencies add up to more than the sum of its parts. That means setting shared agendas, coordinating funding, killing duplication, and getting findings out of reports and into operations. This lesson is a practical playbook for that coordination role, built around the two instruments that do most of the work and the charter that keeps them alive after the leader who created them moves on.

The four failure modes you are fighting

Uncoordinated government research and development fails in four predictable ways. Name them so you can target each one, because the fixes are different and a leader who treats "better coordination" as one problem will build one mechanism and solve at most a quarter of it. Each failure mode also has a different constituency that quietly benefits from the status quo, which is why none of them corrects itself.

• Duplication. Multiple agencies independently fund the same work, like Helen's three benefits chatbots. This is the most visible failure and the most embarrassing when someone external finds it.
• Gaps. Because everyone chases the same visible problems, hard shared problems such as bias auditing methods and procurement standards go unfunded. No single agency owns them, so no single agency pays for them.
• Fragmentation. Findings stay locked inside the agency that produced them, so lessons are relearned at full cost by the next agency to encounter the same problem.
• The research-to-practice gap. Good findings sit in reports and never change how an agency actually operates, which means the money bought knowledge that nobody uses.

Government rarely fails for lack of effort. It fails because the effort is uncoordinated, so the same dollar gets spent three times to learn the same lesson. Notice what the four modes have in common: every one of them is a consequence of agencies being able to see their own portfolio clearly and everyone else's not at all. That is the underlying condition, and most of the instruments in this lesson are attempts to fix visibility rather than to fix intent. Agencies are not duplicating work on purpose.

It is worth being precise about the cost of each mode, because they are not equal. Duplication wastes money that has already been appropriated, which is painful but bounded. Gaps are worse, because the unfunded shared problems tend to be exactly the safety, evaluation, and procurement problems that every deployment depends on. Fragmentation and the research-to-practice gap are the most expensive of all, because in those cases the work was done, the money was spent, the answer exists, and it changes nothing.

Reading the table across rather than down is the useful move. Notice that no single instrument covers more than one mode, which is why a coordination program built around one mechanism, usually a registry, closes one failure and leaves three open. Notice too that the detection column is mostly work somebody has to do rather than a report a system produces. Coordination programs that assume the tooling will surface problems on its own tend to discover, a year in, that everything is being recorded and nothing is being noticed.

Instrument one: the shared research agenda

Coordination starts with a common map of what matters. A shared research agenda is a jointly owned, published list of priority questions that government AI needs answered, with rough ownership attached to each. It works on duplication and gaps at the same time, which is unusual: the same document that shows two agencies they are chasing the same question also shows everyone which questions nobody has claimed. That second effect is the one leaders underestimate and the one that produces the most durable value.

Build it by convening the agencies, not by dictating to them. An agenda that arrives from above gets nominal agreement and no behavior change, because the people who control the research budgets were not in the room when the priorities were set. An agenda that agencies co-authored gets defended by the people who wrote it, and they will defend it in the budget conversations you are not present for, which is where it actually has to survive. Convening is slower and it is the only version that lasts.

A workable agenda has three tiers, and the tiering is what makes it usable rather than a wish list. Without tiers, every question on the page looks like it deserves the same treatment, and the group defaults to arguing about which is most important instead of deciding who pays for what. The tiers answer the funding question directly: one tier is funded once and shared, one is owned by whichever agency the domain belongs to, and one is deliberately spread across several agencies. That is a division of labor, not a ranking.

1. Shared foundational problems that every agency faces, such as bias testing, evaluation methods, and procurement language. Fund these once, centrally, and share the results. These are the questions that are nobody's mission and everybody's dependency.
2. Domain-specific problems that belong to a sector such as health, defense, or benefits. Owned by the lead agency in that domain and visible to all, so that an agency entering a new domain knows who has already worked the ground.
3. Exploratory bets on emerging capabilities. Deliberately distributed so different agencies try different approaches, with results pooled. Here duplication is not the enemy; unshared results are.

The third tier deserves a word, because leaders new to coordination often try to eliminate all parallel effort. That is a mistake at the frontier. When nobody knows which approach will work, several agencies trying different ones is how the field learns, and the coordination job is not to pick a winner early but to make sure every attempt is visible and every result comes back to the group. Duplication is waste when the answer is known and investment when it is not.

The agenda is only real if it is published and revisited. A document that sits in a drawer coordinates nothing, and an agenda that is never revised quietly becomes a description of what mattered eighteen months ago while the actual work moves elsewhere. Put a review cadence in writing, publish each version, and keep the old ones, because the change between versions is itself useful evidence about where the field is moving and which questions have been answered.

Instrument two: funding coordination

Agendas align intentions; funding aligns reality. An agency whose budget points one way and whose signature sits on an agenda pointing another way will follow the budget every time, and no amount of convening changes that. You coordinate funding not by controlling everyone's budget, which you almost certainly cannot do and should not try, but by adding two lightweight mechanisms that change what agencies see and what they can share the cost of.

• A visibility requirement. Before an agency funds AI research above a threshold, say $1 million, it checks the shared registry and posts its own planned work to it. Helen's three chatbots would have appeared on the same registry before a single contract was signed.
• A pooled fund for shared problems. Agencies contribute to a common pot that funds the foundational tier, so no single agency pays alone for work everyone needs. This is how the hardest, least glamorous problems finally get funded.

The visibility requirement is the higher-leverage of the two. It is cheap, it asks nobody to give up control of their money, and it is therefore politically easy in a way that pooled funding never is. It also produces its benefit early, at the point where a duplicate effort can still be stopped without anyone having to cancel a contract or explain a termination to a vendor. If you can only get one mechanism adopted in your first year, take this one.

Be clear-eyed about what a registry does. It reveals duplication; it does not prevent it. A registry only sees work that gets posted, only helps when someone actually reads it before contracting, and only surfaces a collision when two entries are recognizable as the same problem, which they frequently are not when each agency describes its work in its own program language. Treat the registry as a detection mechanism with known blind spots rather than as a control, and staff someone to read it across entries, because the pattern Helen spotted in a meeting is exactly the pattern a registry will not spot by itself.

Pooled funding is harder for reasons that are not about willingness. Whether and how agencies may combine appropriated funds for a joint effort is a question for your budget office and your counsel, and the answer varies by agency and by the authority under which the money was appropriated. Ask that question early, before you have socialized a mechanism you may not be able to build. The failure mode here is a coordination body that spends a year designing a pooled fund and then discovers the pooling is the part it cannot do.

Running the convening so it produces something

Convening rather than dictating is the right instinct, and it is also the point where most coordination efforts quietly fail, because a convening that produces agreement in the room and no commitment outside it looks identical to success for a surprisingly long time. The difference comes down to who is in the room. Invite the officials who control research budgets and can commit their agency, not only the technical staff who will do the work. A meeting of researchers produces an excellent agenda that nobody can fund, and a meeting of budget officials produces a fundable agenda about the wrong questions. You need both, and you need them in the same session.

Put something concrete on the table at the first meeting. An empty invitation to co-create priorities produces a long conversation about scope; a draft agenda that is visibly wrong produces corrections, and corrections are commitments in disguise. The person who tells you that your foundational tier has missed evaluation methods has just taken partial ownership of the foundational tier. Draft badly on purpose if you have to, and be genuinely willing to have the draft rewritten, because a group that changes your document believes it is theirs afterward.

Then close every session with named owners and dates rather than with themes. Coordination bodies drift toward discussion because discussion is comfortable and commitment is not, and the drift is invisible until someone asks what the group has decided in the last year. If a meeting does not end with somebody's name against something, it was a briefing. Briefings are fine occasionally. A coordination council that has held several of them in a row has become a standing meeting rather than a mechanism, and it will be dissolved by the first person who audits its output.

The coordination charter: a usable artifact

Helen turned her playbook into a one-page interagency coordination charter, signed by participating agency leads. It is the artifact that makes coordination durable past any one meeting, and past any one person. Every coordination arrangement that lives only in the relationships between the current occupants of a few senior jobs will end when its current members move on. A signed charter is what a successor inherits. Adapt these clauses.

• Purpose: coordinate AI research and development to reduce duplication, fill gaps, and accelerate adoption.
• Shared agenda: the group maintains and publishes a tiered research agenda, reviewed twice a year.
• Visibility rule: members post planned AI research above the threshold to the shared registry before contracting.
• Pooled fund: members contribute an agreed amount yearly to fund foundational problems.
• Knowledge sharing: members publish findings and after-action lessons to a common repository within 90 days of completion.
• Translation duty: each funded project names an operational owner responsible for piloting findings, not just publishing them.
• Governance: a standing coordination council, rotating chair, quarterly meetings, decisions by consensus with an escalation path.

Two clauses carry more weight than the rest. The knowledge-sharing deadline works because it converts a good intention into a date somebody can be asked about; without it, publication happens when the team has time, which is never. The rotating chair matters more than it looks, because a coordination body permanently chaired by one agency is read by the others as that agency's initiative, and participation decays accordingly. Rotation is cheap and it buys shared ownership of a structure that has no authority to compel anything.

Keep it to one page. A charter that runs to dozens of pages will be negotiated for a year and signed by nobody, and the length signals that the group is trying to legislate the behavior it should be building trust for. The point of the document is not to bind agencies, which it largely cannot do. It is to make the commitments explicit enough that a member who ignores them is visibly ignoring something they signed, in front of peers who also signed it.

Closing the research-to-practice gap

The most expensive waste is research that works and is never used. The money was spent, the answer exists, and operations carry on exactly as before. The fix is structural rather than exhortative, because the gap is not caused by researchers failing to communicate or operators failing to read. It is caused by nobody being accountable for the transition, and by an incentive system that rewards producing findings and is indifferent to using them. Three moves address it.

• Fund the pilot, not just the study. Tie a portion of every research dollar to a real-world pilot in a partner agency. A finding that has never touched an operation is unproven, however well the study was designed.
• Name an operational owner up front. Every project gets someone in operations, not research, who is accountable for trying the result. This is the translation duty in the charter, and it must be a named person before the work starts, not a role assigned at the end.
• Reward adoption, not publication. Recognize the agency that adopted another's finding as highly as the one that produced it. Coordination dies when only originality is celebrated.

The third move is the one that changes behavior over years, and it is the one most likely to be dropped because it costs nothing and therefore appears to be worth nothing. Consider what your current recognition actually rewards. If every award, briefing slot, and leadership mention goes to the agency that built something new, then adopting a peer's finding is a career-neutral act at best, and the rational move for an ambitious program lead is to build their own version. Recognition is one of the few levers a coordinator holds that requires no budget authority at all.

The first move needs one clarification, because "fund the pilot" is easy to say and awkward to implement. A pilot in a partner agency is not the same as a demonstration run by the research team, and only the first tests what you actually need to know, which is whether the finding survives contact with someone who does not care about the study. Build that partner relationship at the design stage, when the study can still be shaped around what a pilot would require, rather than arriving at an operational unit at the end with a completed report and a request for a favor.

What coordination actually bought

Helen did not cancel anyone's project by decree, which she had no authority to do and which would have ended her standing with the group in a single meeting. She stood up the registry and the charter and let the mechanisms surface what the meeting had surfaced by accident. Two of the three benefits-chatbot teams discovered each other on the registry, merged into one effort, and redirected roughly $4 million toward the unfunded bias-auditing work that nobody had owned. The agency that had built the chatbot two years earlier was named operational owner and ran the pilot.

Look at what changed in the ledger, because the arithmetic is the argument. Before coordination, roughly $12 million across three contracts was buying one distinct result, delivered three times. After coordination, the same appropriated money was buying a merged chatbot effort informed by an agency that had already made the mistakes, a bias-auditing capability that had previously gone unfunded because it was nobody's mission, and a real pilot rather than another report. Same money, three distinct results instead of one. That is the entire job in one sentence, and it was accomplished with a registry, a signed page, and no new authority whatsoever.

Anti-Patterns to Avoid

• Trusting the registry to prevent what it can only reveal. A registry sees only posted work, helps only when someone reads it before contracting, and misses collisions where two agencies describe the same problem in different program language. Staff someone to read across entries. The mechanism detects; people prevent.
• Publishing an agenda and calling it coordination. An agenda aligns intentions and a budget aligns reality. If the shared agenda does not show up in what agencies actually fund, you have a document, not a program, and the next duplicate contract will be signed regardless.
• Dictating from above. Priorities set without the people who control research budgets get nominal agreement and no behavior change, because those people were not in the room and will not defend the agenda in the budget conversations you never see.
• Eliminating all parallel effort. Duplication is waste when the answer is known and investment when it is not. Stamping out every parallel attempt at the frontier destroys the only mechanism the field has for finding out which approach works.
• Designing the pooled fund before checking whether you can build it. Whether agencies may combine appropriated funds is a question for counsel and the budget office. Spending a year socializing a mechanism that turns out to be unavailable costs credibility you will need later.
• Ending the project at publication. A finding with no named operational owner and no funded pilot has not been proven and will not be adopted. The translation duty has to be assigned before the work starts, not offered as a next step at the end.
• Celebrating only originality. When every award and briefing slot goes to whoever built something new, adopting a peer's result becomes career-neutral and every program lead rationally builds their own version of the same thing.

Practice Prompts

• List every AI research or development effort your organization has funded in recent years. For each, name the problem in one sentence using words a peer agency would recognize. Then look for two entries that are the same problem described differently.
• Draft the three tiers of a shared agenda for your domain: the foundational problems everyone depends on and nobody owns, the domain problems with a natural lead agency, and the exploratory questions worth trying several ways at once.
• Write the visibility rule you could actually get signed. Pick the threshold, decide what has to be posted and when, and identify who will read the registry across entries. If nobody is assigned to read it, the rule is decoration.
• Take one completed study your organization funded. Name the operational owner it should have had, the pilot it should have been tied to, and what would have had to be different at the start for either to exist.
• Put the pooled-funding question to your budget office and counsel this quarter, before you propose the mechanism. Write down the answer, including which authorities matter, so the next coordinator does not have to ask again.
• Audit your recognition practices. List who was publicly credited in the last year and for what. Count how many of those were for adopting somebody else's finding rather than producing a new one.

Reflection

Think about the last time your organization discovered it had duplicated another agency's work. How was it found? In almost every case the answer is accident: a conference conversation, a shared vendor, a chance meeting like Helen's. That should be unsettling, because it means the discoveries you know about are the ones lucky enough to happen, and the ones you do not know about are still running. The question is not whether duplication exists in your portfolio. It is how it would ever be found.

Then ask the harder question about the other end of the pipeline. Of the research your organization has funded in recent years, how much has changed an operational practice? Not how much was published, briefed, or well received, but how much altered what someone does at a desk. If that number is small, the constraint is not research quality and more coordination of funding will not fix it. What is missing is somebody in operations whose job depended on trying the result.

Glossary

• Shared research agenda. A jointly owned, published list of priority questions government AI needs answered, with rough ownership attached, tiered into foundational, domain-specific, and exploratory work.
• Visibility requirement. A rule that members check and post planned research above an agreed threshold to a shared registry before contracting, so duplication surfaces while it can still be stopped.
• Pooled fund. A common pot to which agencies contribute so that foundational problems everyone depends on can be funded without any single agency bearing the whole cost.
• Translation duty. The charter obligation that every funded project names an operational owner accountable for piloting the finding rather than only publishing it.
• Research-to-practice gap. The distance between a finding that works and an operation that changes because of it. The most expensive failure mode, because the money has already been spent.
• Coordination charter. A short signed document setting out purpose, agenda, visibility rule, pooled fund, knowledge sharing, translation duty, and governance, so that coordination survives changes in personnel.

Related Lessons

• Cross-Agency AI Coordination covers the operational coordination mechanics this lesson applies specifically to research investment.
• Academic and Research Partnerships extends the same agenda-setting question beyond government to university partners.
• Building Innovation Ecosystems places research coordination inside the wider ecosystem a government is trying to build.
• Public-Private Innovation at Scale addresses the funding and partnership routes that sit alongside interagency pooling.
• Technology Transfer and Commercialization deals with the other direction of the research-to-practice problem.
• AI Pilot Program Design is the practical companion to the fund-the-pilot rule described here.
• Shared Services and Infrastructure Models covers the shared-cost pattern applied to platforms rather than to research.

Closing

Nothing Helen did required new authority, new appropriations, or anyone's permission to reorganize anything. She built a registry so agencies could see each other, an agenda so they could argue about priorities in one place instead of five, and a one-page charter so the arrangement would outlive her tenure. Each of those is unglamorous. Together they turned an accidental discovery in a quarterly meeting into a standing mechanism that finds the same class of problem on purpose.

That is the shape of the coordination role at the ecosystem level. You will rarely be the person who commissions the important study or writes the influential paper. You will be the person who made sure two agencies knew about each other before the contracts were signed, that the problem nobody owned finally got funded, and that the finding somebody produced actually reached somebody who could use it. It is a job measured almost entirely in things that did not happen, which is why it has to be built into structures rather than carried in one person's head.

Key Takeaways

• The enemy is uncoordinated effort, not insufficient effort. Duplication, gaps, fragmentation, and the research-to-practice gap are the four failure modes to target, and they need different fixes.
• The underlying condition is visibility. Agencies see their own portfolios clearly and everyone else's not at all. Most coordination instruments are attempts to fix that rather than to fix intent.
• Start with a shared, published agenda. A tiered map of foundational, domain, and exploratory problems prevents duplication and reveals gaps at the same time.
• Convene, do not dictate. Coordination that agencies co-own gets defended in budget conversations you will never attend; coordination imposed from above does not.
• A visibility rule is your highest-leverage tool. It is cheap, asks nobody to surrender budget control, and surfaces duplication while a contract can still be stopped.
• A registry detects, it does not prevent. It sees only what is posted, only helps when read before contracting, and misses collisions described in different program language. Staff someone to read across entries.
• Check the pooled fund is legally buildable before you socialize it. Whether appropriated funds can be combined is a question for counsel and the budget office, and the answer varies.
• Parallel effort at the frontier is investment, not waste. Distribute exploratory bets deliberately and pool the results rather than picking a winner early.
• Make it durable with a one-page charter. Agenda, visibility, pooled fund, knowledge sharing within a stated window, translation duty, and a rotating chair outlast any single meeting or occupant.
• Fund pilots and reward adoption. Tie research dollars to a real pilot with a named operational owner, and credit the agency that adopted a finding as highly as the one that produced it.

Frequently Asked Questions

We have no authority over other agencies' budgets. Can we coordinate at all?

Yes, and the instruments in this lesson are built for exactly that position. A registry, a co-authored agenda, a signed charter, and a recognition practice all work without budget authority, because they change what agencies can see and what their peers will notice rather than what they are compelled to do. Helen had no power to cancel a contract and the coordination still redirected roughly $4 million. Authority makes coordination faster; visibility is what makes it possible.

What threshold should the visibility rule use?

The right threshold is the one your members will actually sign, and the illustrative figure of $1 million in this lesson is a starting point for negotiation rather than a recommendation. Set it too low and you bury the registry in small procurements nobody will read; set it too high and the duplication you most want to catch slips underneath it. Pick a number, publish it, and revisit it once you can see what the registry is and is not catching.

Is duplication always bad?

No, and treating it as always bad is a real error at the frontier. When nobody knows which approach will work, several agencies trying different ones is how the field learns, which is why the exploratory tier exists and why its results are pooled. Duplication is waste when the answer is already known somewhere in government, and it is investment when it is not. The coordinator's job is to tell the two situations apart, not to eliminate parallel work.

How do we get agencies to publish findings they would rather keep quiet?

Put a date on it in the charter and make the obligation mutual. The knowledge-sharing clause works because it converts an intention into something a peer can ask about at the next quarterly meeting, and because every member is equally exposed. Findings that reflect badly on a program are the hardest and the most valuable, so make after-action lessons an explicit part of what gets published rather than leaving the definition to the agency deciding whether to publish.

Who should chair the coordination council?

Rotate it. A council permanently chaired by one agency reads to the others as that agency's initiative, and participation decays as a result. Rotation costs almost nothing and buys shared ownership of a structure that has no power to compel anyone. Pair it with decisions by consensus and a defined escalation path, so that the one case per year that genuinely cannot be resolved in the room has somewhere to go.

What is the first thing to build if we can only build one?

The visibility requirement and the registry behind it. It is the cheapest to stand up, the easiest to get agreement on because it takes nothing away from anyone, and it acts at the moment when duplication can still be stopped without cancelling a contract. Add the shared agenda next, since the registry will quickly show you which problems are crowded and which are empty, and that evidence makes the agenda conversation much easier.

How do we know whether coordination is working?

Count the things that did not happen, which is harder than it sounds and is the honest measure. Track duplicate efforts identified before contracting, previously unowned foundational problems that got funded, findings that reached a pilot, and adoptions of another agency's result. Meeting counts and registry entry counts measure activity rather than effect, and a coordination body that reports those numbers has usually stopped asking whether it is changing anything.