October 13, 2026
Written for the Romanian market, adapted here for readers outside Romania. EU rules apply everywhere; national details are given for Romania as an example.
On October 6, 2026, Siena AI announced a $17 million round led by York IE, bringing its total funding to about $29.7 million. The company, founded by Andrei Negrău and Lisa Popovici and based in the US, builds agents for consumer brands and works with hundreds of brands in beauty, health, nutrition, beverages, and fashion. The round confirms what had been visible for some time: the category has left its experimental phase.
For an online store, the practical question isn't whether the technology works, but what changes from a legal standpoint the moment the agent answers your customer.
A chatbot that provides product information stays within the register of communication. An agent that recommends, answers a complaint, confirms a return condition, and completes the order within the same conversation enters the register of commercial acts. What the agent says binds the store toward the customer, in the same way a statement from a customer service colleague would.
How directly this applies became clear in a Canadian case decided in 2024. An Air Canada customer was told by the company's chatbot that he could apply retroactively for a bereavement fare reduction, which the actual policy did not allow. The company refused to honor the information, and before the British Columbia Civil Resolution Tribunal argued that it was not responsible for what the chatbot said, that is, in the tribunal's reading, that the conversational agent is a separate entity, responsible for its own statements. The argument was rejected, and the company was ordered to cover the difference. The decision has no effect in the European Union, but the reasoning is logical: an agent installed on your channels speaks on your behalf.
The second lesson comes from the opposite direction, that of efficiency. In 2024, Klarna announced an assistant doing the equivalent work of 700 full-time employees, with 2.3 million conversations in its first month and an average resolution time under two minutes. A year later, the company's CEO publicly acknowledged that the focus on cost had lowered quality, and the company began hiring again for support. Total headcount had meanwhile fallen from 5,527 employees at the end of 2022 to 3,422 at the end of 2024, according to the IPO prospectus, amid a hiring freeze. The measurement showed speed, and what it lacked was real resolution in ambiguous or emotionally charged cases, exactly where an unhappy customer needs a human.
For an online store in the EU, the two examples point to a list of actions to take before the technical demo.
Informing the customer. The European AI Regulation requires that a person interacting with an AI system knows it, and the obligation (Article 50, applicable from August 2, 2026) falls directly on the provider of the system: a store that builds its own agent is also a provider, while one that uses someone else's agent must require the disclosure through the contract concluded with the provider. In Romania, for example, ANCOM was proposed through a government memorandum adopted on March 12, 2026 as the national market surveillance authority and single point of contact for this regulation, while the national law setting its powers and penalties is still being drafted. Each Member State designates its own authority, so it is worth checking the one in your country. Displaying the notice on the site remains, in practice, the store's job.
The mandate. An agent that can confirm a return condition or grant a discount exercises a power of representation, and its limits are written in the configuration at the infrastructure level, not in the text instructions given to the model. The maximum discount it can offer, which deadlines can be extended, and the exact conditions under which the conversation stops and is taken over by a human operator are rules that must be locked in code before the first order is processed.
Decision traceability. If a commercial dispute arises months after a purchase, the store must be able to fully reconstruct the exchange of messages: what catalog information the agent read at that moment, what configuration version was running, and whether there was any human approval along the way. This tamper-proof traceability is negotiated directly in the contract with the provider, alongside the data processing agreement for customers' personal data under GDPR rules.
The quality of the data source. The agent has no judgment of its own and amplifies the information it receives from the store's database. If an outdated price, a wrong warranty term, or an incorrect description from the catalog ends up in the response sent to the customer, the legal responsibility for the erroneous offer remains entirely with the store, not with the technology provider.
Delegating the customer conversation to an autonomous agent is not a simple software efficiency step, but a decision of commercial representation. Clarifying the limits of the mandate, the traceability logs, and the escalation rules is done before signing the contract, for the protection of the business.
MAAS Technology - design of governance architecture and control registers that allow companies to adopt autonomous agents without taking on uncontrolled representation risks.