Henrik Solvberg is SVP of AI at Nordvik Industrials, a Nordic maker of factory automation equipment that sells into fourteen countries. His board wants an AI-driven predictive-maintenance service rolled out across all of them within eighteen months. Henrik quickly discovered that "roll it out globally" is three different problems wearing one sentence. The model his team built in Oslo needs machine sensor data that, in the EU, is governed one way; in a planned China deployment, cannot legally leave the country at all; and in a US pilot, sits with a partner who wants to co-own the resulting improvements. The technology was the easy part. The strategy, the partnerships, and the data flows were where the plan lived or died.

What Changes at the Border

International AI strategy is not domestic strategy with translation added. Three things change once you cross a border. First, your data stops being freely movable: localization rules and transfer restrictions dictate where a model can even be trained or run. Second, your regulatory surface multiplies, since the EU AI Act, sector rules, and national requirements can all apply to the same product at once. Third, you rarely operate alone, because entering a new market usually means a partner, a joint venture, or a reseller, each of which reshapes who owns the model, the data, and the customer relationship.

This chapter is about making those three forces work for you instead of discovering them after you have committed. By the end of it you should be able to choose a market-entry mode using criteria specific to AI products rather than generic ones, map data-residency and cross-border transfer constraints before they force an expensive re-architecture, structure partnership terms so that model ownership, data rights, and IP in derived improvements are settled up front, sequence entry so early wins fund later and harder markets rather than spreading yourself thin everywhere at once, and track whether an expansion is actually creating value or quietly draining it.

Where This Fits in Your Global Playbook

This chapter assumes the two that come before it. The chapter on cultural and regulatory differences told you how governance expectations shift across regions; the chapter on global talent told you whether you can staff the operations you commit to. Here you decide where to go, with whom, and on what terms. The chapter that follows, on research and knowledge creation, depends on the partnership structures you set here, because who owns the data and the model improvements determines who gets to learn from them. Treat international strategy as the load-bearing decision: get the entry mode and the data architecture right, and the rest of your global programme has somewhere solid to stand.

Market-Entry Modes and Data Constraints

The first real decision is how you enter a market. Each mode trades control against speed and local fit, and the trade is different for an AI product than for a physical one, because what you are exposing to a partner is not only a customer relationship but a data flow and a model. Henrik used the table below to reject a wholly direct entry into two markets where he had neither the regulatory standing nor the local relationships to succeed alone.

Running underneath every entry mode is the data question. Before committing to a market, classify it on two axes: can the data leave the country, and can the model be served from outside it. A market where data cannot leave, under a strict localization regime, forces you toward in-country training or inference and often toward a local partner who can host it. A market that permits transfer under a mechanism such as EU standard contractual clauses is far cheaper to serve centrally. Getting this classification wrong is the most expensive mistake in international AI, because it usually surfaces only after you have built an architecture that assumes data can move.

Henrik learned this concretely with the China deployment. Sensor readings from a customer's factory floor could not legally leave the country, which meant the central model his team ran from Oslo was simply not a legal option there. The choice narrowed to two paths: stand up an in-country inference environment with a local hosting partner, or decline the market. There was no cheap middle option, and no amount of contract drafting could move data that the law said had to stay put. Recognizing that constraint before signing a customer, rather than after, was the difference between a planned cost and a broken commitment.

A Sequenced Entry and Partnership Playbook

Work the expansion as a sequence, not a simultaneous launch. Consider Henrik's economics (all figures hypothetical). Standing up the predictive-maintenance service in a data-transfer-friendly EU market cost his team about 400,000 dollars and reached break-even in roughly nine months. A strict-localization market requiring an in-country deployment and a local hosting partner was projected at about 1.2 million dollars with a two-year payback. Launching all fourteen markets at once would have spread his small team across every hard problem simultaneously. Instead he sequenced: two friendly markets first to generate reference customers and cash, then the harder localized markets funded by those wins.

The playbook, in order:

• Classify every target market on regulatory burden, data mobility, and revenue potential before you rank them, and enter high-mobility, high-revenue markets first.
• Choose the entry mode per market using the table above. Do not apply one mode everywhere out of habit, which is the most common way a company ends up with a joint venture it did not need.
• Settle the four ownership questions in writing before any partnership goes live: who owns the base model, who owns the customer data, who owns improvements derived from that data, and what happens to all three if the partnership ends.
• Design the data architecture to the strictest market you will serve, so that a later localized deployment is a configuration change rather than a rebuild.
• Stage-gate expansion on reference proof, funding the next and harder market only once an earlier one has produced a live reference customer and hit its payback assumptions.

The fourth of those deserves particular attention, because it is where sequencing and architecture meet. If the first market you build for is the most permissive one, every constraint you meet later arrives as a retrofit. If you design to the strictest regime you intend to serve, even while launching in the easiest, the hard markets become deployments rather than rewrites. That decision costs something up front and is difficult to justify on the first market's business case alone, which is exactly why it has to be made as a portfolio decision rather than a project one.

Metrics for International Expansion

Judge an expansion by whether it creates value net of its hidden costs, not by the number of flags on the map. Flags are the metric that boards reach for by default, and they are the one most likely to reward exactly the behaviour that destroys value. The scorecard Henrik reports quarterly is deliberately narrow.

The decision rule that ties them together: no market advances past pilot until its data classification is confirmed, its entry mode is chosen deliberately, and the four ownership questions are answered in a signed agreement. A pilot that skips these is not an early win; it is a liability that has not been billed yet. The partner dependency ratio deserves its own standing review, because concentration accumulates without anybody deciding to accumulate it. A partner who delivered your first reference customer is the natural choice for the second and third, and by the time the ratio is uncomfortable, the relationship has become expensive to renegotiate and hard to exit.

Applying This in Your Organization

Eighteen months in, Henrik had not launched all fourteen markets, and that was the point. He had five live, three of them profitable and funding the two hardest localized deployments, and he had walked away from two markets where the required joint-venture terms would have handed model ownership to a partner. His board initially read the narrower footprint as underdelivery, until he showed them the alternative: fourteen shallow, loss-making launches with contested IP. The disciplined sequence produced fewer flags but a durable, defensible business, and a much shorter list of agreements he would have to unwind later.

To apply this to your own expansion, work through these questions. Have you classified each target market on data mobility and regulatory burden, or are you assuming your home-market architecture will travel? Are you defaulting to one entry mode everywhere, and would a different mode fit some markets better? For every live or planned partnership, are the four ownership questions answered in writing, or only in a handshake? Is your expansion sequenced so that early wins fund later markets, or are you trying to launch everywhere at once? And what would you do in the next 30, 90, and 180 days if your largest partner relationship ended tomorrow?

Anti-Patterns to Avoid

Cross-border AI expansions fail in recognizable ways, and each has a warning sign that appears before the money is spent.

• Assuming the home-market architecture travels. Data mobility is an architectural constraint, not a legal footnote, and discovering it after the build means a rebuild rather than a filing.
• Applying one entry mode everywhere. Habit produces joint ventures where a reseller would have done, and direct entries into markets where you have no standing.
• Leaving ownership to the handshake. The four ownership questions are cheap to answer before a partnership goes live and contested afterwards, particularly the one about improvements derived from customer data.
• Launching everywhere at once. A small team spread across every hard problem simultaneously finishes none of them, and each unfinished market consumes attention that the profitable ones needed.
• Counting flags instead of paybacks. A market beyond its planned payback is not a presence, it is a subsidy, and it will keep running until somebody makes a stop-or-fix decision.
• Letting partner concentration accumulate. Nobody decides to depend on one partner; it happens by repeatedly choosing the partner who worked last time.

Practice Prompts

Run these against your actual expansion plan. Each should produce a document you can put in front of your board or your counsel.

• Classify your portfolio on the two axes. For every target market, record whether data can leave the country and whether the model can be served from outside it. Any market you cannot answer for is a market you are not ready to commit to.
• Audit the four ownership questions. List every live partnership and mark which of the four are settled in a signed agreement. The percentage you get is the IP-ownership clarity metric, measured honestly for the first time.
• Design to the strictest market. Take the toughest regime you intend to serve and ask your architects what would change if you built for it now. Price the difference against a later rebuild.
• Run the partner-exit scenario. Assume your largest partner terminates next quarter. Write down what happens to the customer relationships, the data, and the model improvements, and how long recovery takes.

Reflection

Think about the last market your organization entered. Was the entry mode chosen deliberately, with alternatives considered, or inherited from how you entered the market before it? Consider your current data architecture: if the strictest regime you plan to serve became a firm commitment next quarter, would that be a configuration change or a rebuild, and who would be the first to know? And when your board measures international progress, does it count markets entered or markets that have hit their payback, and what behaviour does that choice reward?

Glossary

• Entry mode. The structural choice between direct or wholly owned operation, partnership or reseller, joint venture, and acquisition. Each trades control against speed and local fit, with a distinct AI-specific risk.
• Data localization. A requirement that certain data be stored or processed inside a jurisdiction. Where it applies, in-country training or inference becomes the only lawful option regardless of contract terms.
• Data mobility classification. The two-axis test applied before committing to a market: can the data leave, and can the model be served from outside. It determines cost and architecture more than any other single factor.
• Standard contractual clauses. An EU transfer mechanism permitting cross-border data movement under defined terms, which makes a market substantially cheaper to serve from a central deployment.
• The four ownership questions. Who owns the base model, who owns the customer data, who owns improvements derived from that data, and what happens to all three if the partnership ends.
• Stage-gating. Funding the next and harder market only after an earlier one has produced a live reference customer and met its payback assumptions.
• Partner dependency ratio. The share of revenue riding on a single partner relationship, and a concentration risk that accumulates without any explicit decision.

Related Lessons

This chapter is the hinge of the global arc. Cultural & Regulatory Differences supplies the regime-by-regime reading that your market classification depends on, and Global Talent & Brain Drain immediately precedes it, answering whether you can staff what you commit to. Research & Knowledge Creation follows and inherits the partnership terms set here, since who owns the data and the derived improvements decides who is allowed to learn from them.

Closing

Henrik's expansion looked slower than the one his board originally asked for, and it was. What it also was, at the eighteen-month mark, was solvent, defensible, and free of agreements he would later need to unwind. The discipline that produced that is not complicated: classify the constraint before committing, choose each entry mode on its own merits, settle ownership in writing before anything goes live, and let each market earn the funding for the next. None of it requires predicting how regulation will evolve. All of it requires refusing to treat a border as a translation problem.

Key Takeaways

• Three things change at the border. Data stops being freely movable, the regulatory surface multiplies, and you rarely operate alone. Each reshapes the product, not just the go-to-market plan.
• Classify data mobility before anything else. Whether data can leave and whether the model can be served from outside determines architecture and cost, and getting it wrong surfaces only after you have built.
• Choose the entry mode per market. Direct, partnership, joint venture, and acquisition each carry a distinct AI-specific risk, from residency burden to contested ownership of derived improvements.
• Settle the four ownership questions in writing. Base model, customer data, derived improvements, and what happens on termination, agreed before the partnership goes live rather than during its breakdown.
• Design to the strictest market you intend to serve. Built that way, a localized deployment becomes a configuration change; built the other way, it becomes a rebuild funded by a business case nobody prepared.
• Sequence, and stage-gate on reference proof. Early wins in permissive markets fund the hard localized ones. Launching everywhere at once spreads a small team across every difficult problem simultaneously.
• Measure paybacks, not flags. Time-to-first-reference, payback against plan, residency incidents, partner concentration, and ownership clarity tell you whether the expansion creates value or quietly drains it.

Frequently Asked Questions

Should we design for the strictest market even if we may never enter it? Only for markets you actually intend to serve. The value of the discipline comes from the portfolio view: the cost is paid once, early, and it converts later localized deployments into configuration work. Designing for a regime you have no plan to enter is speculative cost with no payback attached, which is a different mistake with the same shape.

What if a market requires a joint venture we are not comfortable with? Then declining is a legitimate outcome and should be recorded as a decision rather than a failure. Henrik walked away from two markets whose required joint-venture terms would have handed model ownership to a partner. The board read a narrower footprint as underdelivery until the alternative was laid out beside it: shallow, loss-making launches with contested IP are not presence, they are exposure.

How do we keep partner relationships from quietly becoming dependencies? Measure the partner dependency ratio and review it on a schedule, because concentration is never chosen deliberately. The partner who delivered your first reference is the obvious choice for the next, and the pattern compounds until renegotiation is expensive and exit is slow. Reviewing the ratio regularly forces the question while you still have alternatives.