Two constitutions, one venture
Every joint venture signs two constitutions. Only one of them is written down.
The first is the shareholder agreement: equity split, board seats, reserved matters, deadlock clauses. Lawyers negotiate it for months. The second is the operating system: who makes which decisions, whose systems run the work, whose data becomes the record, whose people fill the roles that matter, and how a disagreement travels from the floor to a decision. Nobody negotiates that. It gets left to "implementation", and it is decided by whichever partner arrives first with a platform, a team and a plan.
That is the structural thesis of this paper. A joint venture is an operating system, not a shareholding structure. Multi-owner structures do not create new governance failures. They amplify every failure a single-owner company can hide: shared authority with no one to break the tie, parents whose incentives compete with the venture they own, and a structure complex enough that real control drifts away from formal control without anyone voting on it. AI makes this sharper, because it moves decisions out of meetings and into the configuration of systems one partner usually controls.
This paper sets out the mechanisms behind that drift, using public cases, takes on the best argument for leaving things loose, and ends with the structural moves that make the second constitution as explicit as the first.
Equity buys a vote. Operations buy the venture.
The deal team negotiates what it knows how to negotiate. Ownership percentages, board composition, the list of reserved matters, the deadlock mechanism. This work is real and it is hard. But it produces a legal relationship, not an operating one.
The research on how companies build joint ventures shows the imbalance plainly. When McKinsey interviewed senior JV practitioners at 30 S&P 500 companies, with experience of more than 300 ventures between them, they estimated that 40 to 60 per cent of their completed JVs had underperformed their potential. The same study found that companies spend more time on the steps where less value is at risk, that operating-model design comes late, and that top executives are involved in deal terms at a handful of points and then step back. The launch team often pulls out halfway through the launch.
Follow the mechanism. The board is reactive by design: it reviews what has already happened. The operating system is proactive: it defines what happens before there is anything to review. A partner that installs its finance system, its customer process and its people in the first months is not asking permission. It is setting the default. Defaults in a joint venture are extraordinarily hard to reverse, because reversing them means disrupting live operations that the other partner now depends on.
None of this needs bad faith. It needs only that one partner moves faster, cares more about the operating detail, or simply has people available while the other is still finishing the legal drafting.
The clearest public illustration is the joint venture between Danone and the Wahaha Group in China. The venture was set up in 1996, and by 1998 Danone held a 51 per cent stake. On paper, Danone controlled it. In practice, day-to-day operations were delegated entirely to Wahaha's founder, Zong Qinghou, who staffed the venture with his own family and Wahaha Group people. From 2000, companies outside the venture began selling the same products under the same brand, partly owned by Wahaha and by an offshore entity controlled by members of Zong's family. Chinese media reported that the same sales people sold both, with profits allocated between them by Zong.
Danone found out in 2005. It began arbitration in Stockholm in 2007, under the venture agreement. In September 2009 it settled by selling its 51 per cent stake to its Chinese partners.
Read the sequence structurally. The majority owner held the equity, the board seats and the contract. The minority partner held the operating system: the people, the sales force, the distribution and the daily decisions. When the two diverged, the operating system won, and the majority owner's best remaining option was to sell.
That is the first rule of multi-owner governance. Equity buys a vote. Operations buy the venture.
There is a timing pattern underneath it. Year one of a joint venture is almost always easy. Both partners want the launch to succeed, so awkward questions about authority are deferred rather than resolved, because resolving them would slow down a launch everyone is rewarded for hitting. The governance document reflects that moment. It is an alignment instrument, built for a time when both parties want the same thing.
Year three is different. Markets move, one parent's strategy changes, a new executive arrives on one side with different priorities, and resource allocation becomes contested. This is the first real test of the governance document, and it has never been tested. The escalation committee has never escalated anything. The deadlock clause has never been used. What decides the dispute is not the procedure in the agreement. It is whoever already held operational control when the disagreement surfaced, because operational control decides what actually happens on Monday morning while the committee is still scheduling its meeting.
The partner that shaped the operating system in year one does not need to win the argument in year three. Structurally, it already has.
Every parent is also a counterparty
A wholly owned subsidiary has one parent with one set of interests. A joint venture has two or more parents, and each of them is usually something else to the venture as well: a supplier, a customer, a competitor, a source of seconded staff, or the employer of the directors who sit on its board.
That makes conflicting incentives the normal state, not an exception. Each parent wants the venture to succeed, but on terms that also serve the rest of the parent. Transfer prices, the allocation of the best people, the choice of which market the venture may enter, and the decision to reinvest or pay out cash are all points where what is best for the venture and what is best for a parent can separate.
The directors carry that tension into the boardroom. BCG's 2026 review of JV boards describes boards made up of senior parent executives "volunteering" their time alongside their main jobs, often with little board experience, with loyalties shaped by the incentives, reporting lines and career paths of the parent that employs them. BCG's research found that as many as one in three failed JVs can trace at least part of their problems to boards that did not maintain clear decision rights, disciplined process and a commitment to the venture's interests over the parents'.
Each director is acting rationally. Their bonus, their next role and their reputation sit with the parent. The venture is a seat they hold, not a career they own. The result is a board in which every member represents a counterparty, and nobody is paid to represent the venture.
The same pressure runs below the board. Take a familiar pattern: a bank and an insurer set up a joint venture to sell insurance through the bank's branches. The bank seconds its distribution managers. The insurer seconds its product and claims people. Each group keeps its parent's grade, its parent's bonus scheme and its parent's career path, because nobody wants to take a job in a venture that might be unwound in five years. When the venture has to choose between a product that earns more for the bank's branches and one that carries better economics for the insurer, every person in the room knows which answer their next promotion depends on. The venture's plan says one thing. Its people are paid to do two others.
Seconded management is not neutral management. It is each parent's operating system, installed inside the venture, wearing the venture's badge.
TNK-BP showed how this plays out when the stakes are large. BP and the Alfa Access Renova (AAR) consortium were equal half-owners. In 2008 the partners split over a familiar question: invest or distribute. The Russian shareholders wanted to cut $900 million from a proposed $4.4 billion investment plan and pay it out as dividends. The chief executive, Robert Dudley, refused, warning that the cut would cost significant future production. The Russian partners accused him of putting BP's interests ahead of the venture's. In July 2008, after Russia declined to renew his visa, he left the country.
Both positions were rational for the parent that held them. One parent wanted cash. The other wanted reserves and long-term production. A chief executive drawn from one parent could not be seen as neutral by the other, because structurally he was not.
The shareholder agreement then became a weapon. In January 2011, BP announced a share swap and Arctic exploration partnership with Rosneft. BP's TNK-BP partners argued that this breached the TNK-BP shareholder agreement, obtained an interim order in London and took the matter to arbitration. In May 2011 BP announced that the Rosneft agreements had lapsed. In March 2013, Rosneft completed the purchase of BP's 50 per cent interest in TNK-BP, alongside BP's further investment in Rosneft.
The venture resolved its conflict the way many multi-owner structures eventually do. It stopped being multi-owner.
Shared intent is not a mechanism
When the operating system is missing, partners fill the gap with one of two substitutes. They govern by consensus, or they govern through a person. Both work for a while. Neither is a mechanism.
Consensus is the polite version. Haven, the healthcare venture formed by Amazon, Berkshire Hathaway and JPMorgan Chase, had about as much executive goodwill behind it as any venture could have. It was announced in January 2018 with the ambition of finding better ways to deliver healthcare to more than a million employees. It ceased operations at the end of February 2021. The public account of its closure is structural: it had proved harder than expected to come up with ideas that made sense for all three companies, which had very different businesses, internal systems and existing health arrangements. Its chief executive stepped down in May 2020 and was never replaced. The three companies said they would carry on less formally, each designing programmes for its own employees and local markets.
Look at where authority sat. Haven did not own the parents' health plans. It could design, propose and pilot, but the decision to change anything for an employee belonged to three separate organisations with three separate operating systems. Any idea had to clear all three. That is not governance. It is a standing request for unanimous consent.
The second substitute is a person. Renault and Nissan built a cross-border alliance that lasted two decades, held together by cross-shareholdings that were never balanced. Renault held about 43 per cent of Nissan with voting rights. Nissan, which made more cars, held 15 per cent of Renault with no voting rights, an imbalance that had long bothered executives in Japan. For years the gap between the formal structure and the operating reality was bridged by one executive, Carlos Ghosn, who chaired Nissan, ran Renault and led the alliance. An analyst at the time called him the glue that held the two companies together.
When he was arrested in Japan in November 2018, the glue went with him, and the structure underneath was exposed. The repair took more than four years. Under the new alliance agreement, which came into force in November 2023, Renault and Nissan hold cross-shareholdings of 15 per cent each, voting rights are capped at 15 per cent, and Renault transferred 28.4 per cent of Nissan's shares into a French trust where they are voted neutrally.
The lesson is not about one executive. It is about what the alliance had been using in place of an operating system. A strong individual can carry decision rights that the structure never allocated. That works until the individual leaves, falls out with a parent or is removed. Then the venture discovers that its real governance was a person, and that the written governance was never designed to carry the load.
AI moves the constitution into the platform
Every pattern so far predates AI. Systems choice has always mattered more than board composition. Workflow design has always outlasted the equity negotiation. What AI changes is where the decisions are made and how fast the default sets.
In a process-era venture, the system recorded decisions that people made. In an AI-enabled venture, the system makes a growing share of them: what to flag, what to approve, what to price, what to escalate. The decision rights are no longer only in the delegation-of-authority matrix. They are in the model's decision boundary, its confidence thresholds, the data it was trained on and the rules for when it hands a case to a human. Whoever controls that configuration controls those decisions, whatever the reserved-matters list says.
That control usually sits with one partner. The model runs on someone's platform. The data pipeline feeding it is someone's. The team that retrains it reports to someone. In a single-owner company that is an architecture choice. In a joint venture it is an allocation of power that looks like an architecture choice.
Consider a pattern that is now common. A logistics joint venture between a port operator and a freight forwarder deploys a model to allocate berth slots and prioritise cargo. The forwarder's data team builds it, on the forwarder's cloud tenancy, trained largely on the forwarder's shipment history. Within a year the model is making thousands of allocation decisions a week. Nobody voted on its objective function. When the port operator notices that the forwarder's own cargo is rarely the one that waits, it has a board seat, an audit right it has never used, and no practical way to see inside the model. The dispute that follows is described as a technical one. It is a governance one.
AI also compresses the time in which a board can catch up. A process-era system rollout ran for many months, with several checkpoints where a partner could object. A model can move from pilot to operational dependency far faster, often without a single checkpoint that anyone labels a governance event. By the time the board receives a paper on "the AI programme", the model is already shaping the decisions the board believes it oversees.
The board's instruments do not fit the thing it now has to govern. A reserved-matters list was written to approve budgets, appoint officers and sign off major contracts. It asks yes-or-no questions about discrete events. A model making thousands of small decisions a day does not produce discrete events. It produces a distribution of outcomes that drifts as the data changes. A board cannot approve each decision. What it can govern is the boundary: the classes of decision the system may take alone, the confidence level below which a human must review, the audit trail and who can read it, and the approval route for retraining, which is the AI equivalent of amending the operating agreement. Few JV agreements have that vocabulary. Most route model changes through the IT change process of whichever partner runs the platform, which is the one process least likely to recognise an allocation of power when it sees one.
Regulation is starting to force the allocation into the open. The EU's AI Act attaches obligations to roles, including the "provider" that develops an AI system or places it on the market under its own name, and the "deployer" that uses a system under its authority. In a joint venture those roles have to land on a specific legal entity: the venture, one parent or both. A governance document that is silent on which partner is the provider of the venture's models has left a regulatory question unanswered, as well as a commercial one.
The decisions that now need to be written into the second constitution are specific:
- whose platform hosts the venture's data and models, and what audit and query rights the other partner holds;
- who may change a model's objective, thresholds or training data, and which changes need joint approval;
- who owns the training data, the resulting model and its outputs if the venture ends;
- which classes of decision the system may take without human review, and who sets that boundary;
- which entity carries the provider and deployer obligations.
Leaving these to the integration workstream is how one partner ends up running the venture's intelligence while both partners share its liabilities.
The strongest objection: specify less, trust more
The best argument against all this deserves a fair hearing, because experienced dealmakers make it.
It goes like this. Joint ventures run on relationships, not contracts. Partners cannot foresee how the market, the technology or their own strategies will change, so a venture needs room to adapt. Trying to specify the operating system in detail at signing slows the deal, signals distrust and produces a rigid structure that breaks the first time reality differs from the plan. Worse, allocating operating control to one partner turns the other into a passenger. Better to sign the equity, build trust in the first year and let the operating model emerge.
Parts of this are right. Trust matters. Over-specified agreements do break. A venture designed to the last detail at signing will need redesign within two years.
But the argument confuses two things: specifying outcomes and specifying mechanisms. Nobody can contract for every outcome. Every venture can decide, at formation, who decides what, on whose systems, with what visibility for the other side, and how a deadlock gets broken. Leaving those mechanisms open does not create flexibility. It creates a vacuum, and a vacuum is filled by the partner that moves first. That is the opposite of a relationship of equals.
The case that is often cited for the relational view actually supports the structural one. NUMMI, the plant in Fremont, California, jointly owned by General Motors and Toyota, opened in 1984 and is remembered as a joint venture that worked. It worked because the operating system was explicit. Toyota ran the plant on its production methods; GM handled distribution of the car and came to learn how the system worked. Nobody pretended that operating control was shared. The allocation matched what each partner wanted from the venture, so it held. When GM went into bankruptcy in 2009 and pulled out of the deal, the plant closed the following year. The venture ended because one parent's situation changed, not because the operating system failed.
Trust is not a substitute for a mechanism. It is what a working mechanism produces over time. Haven had the trust. It did not have the mechanism.
The objection does carry one answer this paper accepts. If the operating system cannot be agreed, the parties may not need a joint venture at all. A joint venture is a governance-heavy vehicle. It is justified when two parties need long-term shared ownership of something neither can build or buy alone. When one party simply needs a capability the other has, a contract with clear deliverables may deliver it faster, with none of the shared-authority overhead. Choosing a joint venture for its strategic signal, when a contract would do the job, is a governance failure that starts before the first board meeting.
What to change, and in what order
The moves below are not a launch checklist. They are the structural decisions that keep the written constitution and the operating constitution from separating.
First, decide whether this should be a joint venture at all. Before any equity term is drafted, compare the venture against the alternatives: a supply contract, a licence, an acquisition, a minority investment. Ask what the venture must own jointly that a contract cannot deliver. If the answer is thin, the overhead of shared authority will cost more than it returns.
Second, negotiate the operating system with the equity, not after it. The people who decide the cap table should decide, or be bound to sign off, the decision-rights map, the systems and data architecture, the leadership appointments and the escalation path. These are one negotiation. Splitting them guarantees that whoever controls implementation controls the venture.
Third, write architecture and AI authority into the agreement with the same precision as board composition. Name the authoritative platform, the other partner's audit and query rights, the changes to models and data that need joint approval, ownership of data and models on exit, and the allocation of regulatory roles. Silence is not neutrality. It is a grant to the partner that acts first.
Fourth, give the venture its own management, with its own incentives. Appoint a chief executive accountable to the venture, not on loan from a parent. Pay the leadership on the venture's results. Give directors a clear duty to the venture and recruit some who have no parent career at stake. A board of counterparties will behave like one.
Fifth, test the dispute mechanism before you need it. Before launch, run the board through a simulated conflict: a contested investment, a disputed transfer price, a personnel decision one parent opposes. If the escalation path cannot resolve the simulation, it will not resolve the real thing in year three.
Sixth, design the end at the beginning. Most of the cases in this paper ended with one partner buying the other out, under pressure, after years of dispute. Agree the exit mechanics, valuation basis and data and model separation at formation, while both parties still want the venture to happen.
Leaders should also stop three things. Stop treating systems selection as an IT workstream. Stop measuring partnership health by the absence of disputes on the board agenda, because disputes that never reach the board are being settled by operational control. Stop assuming a strong individual can carry decision rights the structure has never allocated.
A joint venture is two or more organisations agreeing to share something none of them fully controls. That can be done well. It cannot be done implicitly.
Ownership is negotiated once. Control is exercised every day.