AI Agents in Trade: Costs, Risks and Effort of Switching

What changes

The use of AI agents in trade means that traditional, menu‑driven applications are replaced by interactive, conversational services. Instead of a pure product‑search function, agents such as Boardy, Accio or Nomi offer personalized recommendations, order and returns processing, and integrated payment and logistics interfaces. These agents can be provided both internally for employees and externally for end customers, shifting user interaction from click‑ to dialogue‑based workflows. For companies, this means that existing front‑end systems, such as shop front‑ends or ERP portals, must be extended with AI layers to support the new dialogue flows.

Another difference lies in the type of business models behind the agents. While classic software licenses are usually offered as a one‑time payment or annual subscription, AI agents increasingly rely on freemium models, transaction‑based fees or service packages geared to usage and outcome. This leads to an altered cost structure where variable components carry more weight and billing often occurs via API calls or user sessions. For trade, this means that budget planning must be made more flexible, yet at the same time more transparent.

What it costs

Provider Boardy mentions in its public material a freemium model where basic functions are free of charge and advanced features are paid. Concrete pricing figures are not published, however. The costs for operating an AI agent typically consist of several components: license or usage fees for the underlying AI model, cloud‑infrastructure costs for compute power and storage, integration effort for interfaces to ERP and merchandise management systems, as well as ongoing maintenance and support fees. Since no concrete numbers are available from the source, we can only describe the cost structure, not name absolute amounts.

A typical cost plan could look as follows:

• Basic license or API usage: dependent on the provider, often tiered by number of requests per month.
• Cloud hosting: variable, depending on compute power (e.g. GPU instances) and data volume.
• Integration services: one‑time project costs for connecting to existing systems, usually billed in time units.
• Support and maintenance: monthly flat rates or based on effort, often as a percentage of the license fee.

Companies should factor these components into their financial planning and especially budget for variable costs at high transaction volumes.

What can break

The introduction of AI agents can heavily strain existing processes and systems. First, data from merchandise management, CRM and logistics must be converted into a format that the AI model understands. This means that interfaces to ERP systems (e.g. SAP, Microsoft Dynamics) need to be newly developed or adapted. During this migration, temporary outages or inconsistencies can occur if data are not synchronized. Likewise, existing front‑end applications that previously relied on classic search and order workflows must be extended with dialogue logic, which can lead to short‑term usability issues.

Another risk factor is staff training. Sales and customer‑service teams need to learn how to interact with the agent, review requests and intervene in case of misinterpretations. Without adequate training measures, the risk of incorrect orders or wrong recommendations increases, which can impair customer satisfaction. Moreover, existing third‑party integrations, such as for payment or shipping providers, can be destabilised by new API calls if the interfaces are not precisely aligned with the agent’s requirements.

What a switch demands

A successful switch to AI agents requires structured project management. First, an interdisciplinary team must be assembled comprising IT architects, department heads (e‑commerce, logistics, procurement) and change‑management experts. The planning phase includes analysing the existing system landscape, defining use cases for the agent and selecting a suitable provider. The actual implementation usually takes between three and six months, depending on the complexity of the interfaces and the amount of data to be migrated.

During implementation, clear milestones should be defined, such as completion of the data‑mapping phase, the test and pilot phase with selected customer groups, and the gradual activation in live operation. Parallel to this, a training plan must be created for the affected staff, comprising both technical training (API usage, error diagnosis) and customer‑oriented training (dialogue guidance, escalation processes). Finally, a monitoring framework is needed to measure the agent’s performance, control costs and make swift adjustments if required.