Aissist.io starts self-evolve AI pilot and flags five open problems

Aissist.io has launched a customer pilot of self-evolve AI with several support and sales customers and published five unresolved problems in continuous AI improvement. The move highlights how agentic systems may learn from live traffic, but still face hard questions around rewards, observability, human oversight and experiment design.

Why it matters: - Aissist.io is testing whether AI systems can improve themselves in production instead of waiting for humans to manually rewrite prompts, rules or workflows. - The company says the approach could help AI keep up with changing customer policies, incidents and live business conditions. - The pilot also puts a spotlight on five problems the field has not solved, including how to define success and how much human oversight the loop still needs.

What happened: - Aissist.io began a customer pilot of self-evolve AI with several customers across customer support and sales operations. - The company also published five unresolved problems facing continuous AI improvement. - The pilot is built to identify performance gaps, propose changes, test them against live traffic and keep the changes that verify.

The details: - Aissist.io says the pilot uses a closed-loop architecture built from three existing components. - AgentMesh handles live interactions. - Pulse scores conversations handled by both AI and human agents. - Evolve turns evaluation signals into testable changes, runs them against live traffic and deploys the changes that verify. - Release uses risk-based grading. - Low-impact changes deploy automatically. - Changes tied to policy, compliance or brand require human approval with supporting evidence. - Lifan Xu, co-founder at Aissist.io, said the effort started because customers kept hitting the same wall: the AI is good, they want it better, and the only reliable path from there is a human reviewing transcripts. - Aissist.io says the field has shifted from humans authoring systems to systems proposing changes to themselves under human supervision and against human-defined goals. - The company says the pilot cohort is intentionally small and autonomy is graded by risk. - Aissist.io provides agentic AI automation for customer support and sales operations. - The platform integrates with existing service and CRM tools rather than replacing them. - More information is available at the company's announcement.

Between the lines: - Aissist.io is arguing that traditional tuning methods no longer fit agentic systems, where behavior emerges from instructions, retrieved context, tool results and conversation history. - The company says optimization can become non-local, so a fix in one area can create regressions elsewhere. - Aissist.io also points to slower knowledge updates than product changes, which can leave the system reasoning from stale information. - The company says AI systems can stay static during incidents even as customer conditions change within hours. - The five unresolved problems suggest the biggest challenge is not automation itself but building a safe learning loop that can separate signal from noise. - The unresolved problems are defining a positive outcome, separating knowledge gaps from behavior gaps, dealing with partial observability, incorporating human guidance and designing experiments. - Aissist.io says reward signals differ by business, so a universal optimization target can push systems in the wrong direction. - The company says failures can look the same even when the remedy should differ. - Aissist.io also says outcomes often arrive without the full context that caused them, which can produce unreliable conclusions.

What's next: - Aissist.io will use the pilot to test how far self-evolve AI can move from theory into live operations. - The company’s risk-based release model suggests more autonomous changes may follow if early results hold up. - The field still needs answers on when to use human judgment, when to run in simulation and when to learn from real traffic.

The bottom line: - Aissist.io is trying to turn AI improvement into a managed feedback loop, but the company’s own list shows the hardest problems are still unresolved.