In 2026, nearly every mid-sized company is being pitched an AI agent: a system that answers employee or customer questions around the clock, in any language, at a fraction of the cost. The pitch is new. The problem isn't.

Two decades ago, large companies solved the same problem with people. They built shared service centres: teams that answered the same kinds of questions for employees in many countries and many languages, with consistent quality. Miklós Róth ran one of those teams.

What did Miklós Róth do at IBM?

From March 2004 to May 2005, Róth was team leader of a multilingual AskHR team at IBM's International Shared Service Centre. The team, roughly 15 to 18 people, handled HR questions from IBM employees in several countries: pay, benefits, policies, procedures, the endless small questions that keep a large organization running.

On paper it was a management role between his finance degree and founding his own company in 2007. In practice it was a crash course in a discipline that has suddenly become central to AI: designing a service that gives correct, consistent answers at scale, across languages, without the person asking ever seeing the complexity behind it.

Why is a shared service centre a preview of AI agents?

Because the design questions are identical. Replace the people in the seats with software, and the same four problems decide whether the service works.

1. Is the knowledge base right? A service desk is only as good as the information behind it. Outdated policies, contradictory documents and tribal knowledge produce confident wrong answers, from people and language models alike. In S·I·C·T terms, that's Information without Structure.

2. When does a human step in? Every good service centre has escalation rules: which cases first-line staff can close, which need a specialist, which need a manager. AI agents need exactly the same thing, written down, tested and owned. Without it, the system either escalates everything and saves nothing, or escalates nothing and fails dangerously.

3. Does the tone survive translation? A multilingual team learns fast that an answer can be accurate in one language and confusing, offensive or legally risky in another. Translating a policy is not the same as communicating it. AI agents that "speak fifty languages" inherit this problem at scale.

4. Is the issue truly resolved? Service centres learned to measure resolution, not just tickets closed. A fast answer that brings someone back three days later is a failure, however good the dashboard looks. AI deployments that celebrate conversation volume are repeating an old mistake.

What does this add to his AI advisory work?

Two things that are hard to learn from AI tooling alone.

Respect for the operating layer. Strategy decks tend to stop at the model. Real service quality lives in knowledge management, escalation paths, quality reviews, cover for edge cases and the feedback loop between them. When Róth works as a fractional Chief AI Officer, this is often where he starts, because it is usually where AI agent projects quietly fail.

A multilingual instinct. Róth works in Hungarian, English, Spanish and German. For European companies serving several markets, the gap between a translated answer and a correct one isn't academic. It is customer trust, and in regulated areas like HR, finance or health it can be legal exposure.

What changes when the "agent" is software?

Three things get harder, not easier.

Scale hides errors. A human team of fifteen makes mistakes one conversation at a time, and a team leader reading samples will usually spot a pattern within days. An AI agent can repeat the same wrong answer thousands of times before anyone notices, because every individual answer looks fluent and confident.

Accountability gets blurry. In a service centre, every answer has a name behind it. With an AI agent, it's tempting to treat the system as the owner. It isn't. Someone has to be responsible for what it says, when it is updated and when it is switched off.

Regulation now applies. The EU AI Act requires that people be told when they are interacting with an AI system, and it classifies AI used for decisions about recruitment, promotion, termination or task allocation as high-risk. An HR agent that only answers policy questions is a very different legal object from one that influences decisions about people. Knowing which one you are building is part of the design, not an afterthought.

How should a company plan an AI agent rollout?

Borrow the shared service playbook:

None of this is glamorous, and that is the point. The companies that get real value from AI agents will be the ones that treat them as a service operation, not a software feature. Róth's broader argument, that AI amplifies whatever it touches, applies directly: a well-run service gets faster and cheaper, a messy one gets messier at scale. Here is that argument in full.