Matching agentic AI architecture to the right problem
The right AI agent architecture depends on whether a problem's ambiguity sits in the input data, where agents add real value, or in the system design, where it creates risk.
The right AI agent architecture depends on whether a problem's ambiguity sits in the input data, where agents add real value, or in the system design, where it creates risk.
Statistical models and business rules change on different schedules and should be built as separate components. Combining them means retraining the whole system every time a policy changes.
Typed data models replace free text at agent boundaries, and specialized validators review generated content before a human sees it. Quality problems drop sharply once ambiguity at interfaces is removed.
A dashboard showing memory accumulation under load is what separates a diagnosed failure from an unexplained restart. The same visibility later caught idle resource waste worth fixing through hardware tier changes.
Every dependency, including transitive ones, needs explicit versioning before production. An unpinned update once pulled in a dependency whose download was blocked by a corporate proxy, breaking a job that had run for months.
AI should accelerate the writing of deterministic code rather than run it, especially where output must be exact, auditable, and governed by explicit rules. This keeps AI-assisted speed without probabilistic risk in production.
Thomas Reichert, data scientist and AI solutions engineer at Helvetia Baloise, argues that AI amplifies whatever discipline or chaos already exists in a system rather than fixing what is broken. Drawing on two production systems built at the insurer, a classical dynamic pricing engine for motor insurance and a multi-agent financial planning tool for life insurance, Reichert shows that the choice of agentic AI architecture should follow the shape of the problem rather than the novelty of the technology.
The pricing engine uses no agents because its statistical and business-rule components change on different schedules and gain nothing from agentic reasoning. The financial planning tool uses multiple agents because synthesizing hundreds of data points into a coherent narrative is genuinely interpretive work.
Both systems failed in production for reasons that had nothing to do with model accuracy: database write locks, cleartext credentials, memory leaks, and unpinned dependencies. As Reichert frames it, the failure mode shifts once AI moves from notebook to production, and the fixes that matter most are ordinary software engineering discipline, not better models.
