Why business agents make a strong case for small language models

Preparing a customer briefing involves several tasks. The system needs to identify the customer, find recent activity, retrieve relevant agreements and assemble a useful summary.

Each step has a different purpose. Some involve a small set of possible decisions. Others require interpreting several sources together.

This makes model selection an architectural question: what capability does each step need, and how should the system handle exceptions?

The case for specialised models

In their position paper Small Language Models are the Future of Agentic AI, Peter Belcak and colleagues argue that many agent tasks are repetitive and sufficiently narrow for small language models. They propose systems that combine specialised models with larger models where broader capabilities are needed.

The paper presents an architectural and economic argument. For an individual business workflow, the gains still need to be measured.

Start with a task you can define

For Traize, useful candidates could include turning a question into a search query, choosing a search tool or extracting structured fields from a request.

Consider “What changed in the Northwind account this month?” Before writing the briefing, the system needs to resolve the account and date range. A specialised component could produce a structured search request with clearly defined fields.

That output is relatively straightforward to check. Does the account exist? Is the date range correct? Are the required fields present?

A well-defined task gives us a practical target for training and evaluation.

The surrounding system determines reliability

A model needs tools, context, validation and a way to recover when something goes wrong. Those components form the operating framework around an agent.

For a customer briefing, that framework should enforce access permissions, validate tool arguments and retain references to retrieved sources. It should also recognise when a request needs a more capable model or clarification from the user.

These decisions affect both quality and cost. A cheap model call becomes expensive if it causes repeated searches, invalid tool requests or an unusable answer.

The useful metric is the cost of a successfully completed task, including retries and escalations.

More control over deployment

Specialisation also raises an opportunity to evaluate models that can run within a chosen deployment environment. For Traize’s interest in sovereign European AI, that means examining model provenance, licences, hosting and operational control alongside quality.

A smaller model can make a deployment more feasible. Sovereignty depends on the actual conditions under which it is trained, supplied and operated.

Our direction is to build around measurable business tasks: define the work, choose suitable models and evaluate the complete result. That creates a practical basis for efficient AI that companies can use repeatedly throughout the day.