How do you run AI agents across several countries?

Keep one platform and one set of floors, but treat each country as its own deployment: its own owner, its own sources and policies, its own limits, and its own certification on local cases. An agent that earned depth 3 in one market has proven nothing yet in another. Keep the record identical everywhere.

What changes from one country to the next?

More than the language. Each market brings different products, different policy wording, different regulators, different data protection law and often different systems of record. A collections agent in one country may need to observe contact-time rules that do not exist in another. A support agent may face a regulator that expects specific wording in complaint responses.

The same model and the same brief can produce very different results on another country's cases. Names follow different conventions, addresses have different structures, and identity documents differ. Those are exactly the details that break grounding.

What stays global and what goes local?

The left column is what makes the operation coherent. The right column is what makes it safe. A team that centralises everything ends up with agents that are compliant nowhere in particular.

What are the steps to add a country?

1. Local process owner: accepts the agent for that market and signs the scope. Evidence: a signed scope naming the country.
2. Local compliance: lists the rules that change the agent's bounds, such as contact limits, disclosure wording, and where personal data may be processed. Evidence: a requirement list, mapped to bounds.
3. Platform owner: connects local sources and checks each carries its provenance. Evidence: the source register for the country.
4. Agent, depth 1: shadows local work. Evidence: filed proposals on local cases.
5. Local reviewer: scores the shadow run. Evidence: scored cases, including local edge cases.
6. Owner: certifies and sets depth for that market only. Evidence: a certificate scoped to the country.

What does it look like across three markets?

Say a consumer lender runs a servicing agent in the UK and wants to add Ireland and Kenya. In the UK it holds depth 3, enough for payment-date changes.

For Ireland, the products are similar but the policy wording, complaint handling and data protection authority differ, so the brief takes Irish policies and certification runs on Irish cases. It earns depth 3 quickly. For Kenya, customers often pay through mobile money, contact happens in more than one language, and the systems of record are different. The agent starts at depth 1 and moves to depth 2 for drafting only, with local agents sending. Our guide on whether an AI agent can give support in several languages covers that part.

One record format across all three means a group risk team can look at any decision in any market and see the same fields: the agent, the depth, the bound that decided and the person who approved.

Which controls hold it together?

• One owner per market with authority to narrow or stop the agent there, without touching other markets.
• Country tags on every verdict, so reporting and audits can filter by market.
• Local data handling decided before go-live, including whether personal data may go to a model provider outside the country.
• Coordinated changes. A global brief change is recertified in each market before it applies there.
• Traceable hand-offs where one agent passes work across borders, which tracing a task across several AI agents explains, and clear limits where an orchestrator coordinates several agents.

The broader sequence is in our guide to deploying an AI agent.