vietnamese crab exporter
Commercial Content

Amidst rising supervisory pressure on AI Agents, Alinia AI bets Seny can supply the runtime control layer

Matthew Kayser

Supervisors on both sides of the Atlantic are tightening expectations around AI systems that can plan and act with limited human involvement. Transparency obligations under Article 50 of the EU AI Act have applied since August 2, 2026, high-risk obligations for uses such as creditworthiness and insurance have been pushed to December 2027 under the Digital Omnibus, and U.S. banking guidance has carved generative and agentic AI out of traditional model-risk frameworks while still expecting firms to demonstrate control. Against that backdrop, Alinia AI is betting its newly released Seny model can give regulated institutions a practical way to check agent behavior against specific legal requirements in real time.

Regulatory pressure is shifting toward continuous oversight

What these rules and guidance documents share is an emphasis on evidence of control rather than one-time model validation. Under the EU AI Act, systems that interact with people or generate content must already meet transparency duties, and high-risk uses such as credit assessment will eventually require risk management, logging, and human oversight even if the full obligations arrive later. In the United States, SR 26-2 excludes generative and agentic AI from its scope but states that banks’ broader risk management and governance practices should guide appropriate controls for these systems. It pushes banks and other institutions to rely on broader governance practices, third-party risk management, and sector rules such as fair lending and consumer protection. FINRA’s 2026 Annual Regulatory Oversight Report reinforces that existing rules continue to apply when firms use generative AI and highlights monitoring, documentation, and human oversight as considerations for managing AI agents.

The practical result is a demand for continuous, auditable checks on agent behavior in live workflows. Manual review alone may struggle to keep pace with the volume of decisions made by AI agents, and generic safety filters may need additional, specialized checks to assess issues such as investment suitability under MiFID II or explanations for loan denials under Regulation B.

Alinia AI, founded in late 2023 by Ariadna Font Llitjós and Carlos Muñoz Ferrandis, has positioned Seny as the foundation for that control layer. Font Llitjós previously led Twitter’s machine-learning platform and its Responsible AI efforts and held senior roles at IBM Watson; Muñoz Ferrandis previously served as AI Counsel at Hugging Face and worked on AI governance issues including the EU AI Act.

“The world is asking how to keep AI safe, and part of the answer is in the law,” Font Llitjós said in the announcement. “The other part is an independent control layer, extremely scalable and domain-specialized, that does not add noticeable latency. Seny is the foundation that enables this at Alinia.”

Seny is a 4-billion-parameter base model trained on legislative corpora curated by legal experts and optimized for risk classification and legal reasoning. Alinia post-trains policy-specific “judges” on top of it for particular rules. Those judges can operate as runtime auditors, guardrails, or offline evaluators across banking, insurance, healthcare, and life-sciences workflows.

“Verifiable trust and compliance at scale is what lets a bank or an insurer put an agent in front of customers,” Muñoz Ferrandis said. “Before Seny, training a judge could take our research team months. Now, our legal engineers can do it in hours.”

Latency and accuracy claims

According to Alinia’s Seny launch announcement, a median latency of 151 milliseconds for Seny-based judges is claimed to be five to 17 times faster than the frontier and guard models it tested. Inside a customer VPC, the company says latency can often drop toward 50 milliseconds. That speed, it argues, makes it feasible to evaluate every interaction rather than a sample.

On accuracy, Alinia reports that its Seny-based judges achieved the highest or near-highest F1 scores in several specialized domains, including legal advice and creditworthiness assessment, while recording lower median latency than the comparison models. The company also reports performance improvements across nine languages and has released public benchmark datasets for legal and tax advice. These results reflect Alinia’s own testing.

What the timing implies

The regulatory calendar still leaves institutions in a transitional period: some obligations are already enforceable, high-risk technical requirements are delayed, and traditional model-risk frameworks may need to be adapted to address the characteristics of agentic systems. Alinia AI is betting that a small, purpose-built model focused on specific legal provisions can fill part of the resulting control gap for firms that want to put agents in front of customers without waiting for the next wave of detailed rules. Whether specialized legal-alignment models become a standard component of that infrastructure will depend on how well the performance claims hold up under real supervisory scrutiny and independent testing.

The information provided in this article is for general informational and educational purposes only. It is not intended as legal advice. Readers should not rely solely on the content of this article and are encouraged to seek professional advice tailored to their specific circumstances. We disclaim any liability for any loss or damage arising directly or indirectly from the use of, or reliance on, the information presented.