While corporations are faced with the challenges of using AI agents in critical applications, a brand new, more pragmatic model is created that brings people back into control as strategic protection against AI failure.
Such an example is mixtureA platform that uses a “colleague-in-the-loop” approach to make AI agents reliable for mission-critical work.
This approach is a response to the growing evidence that completely autonomous agents are a game with high use.
The high costs for unchecked AI
The problem of AI hallucinations It has grow to be a tangible risk because corporations explore AI applications. In a recently made incident, the AI-affected code editor Cursor saw his own support bot Find a fake guideline Restriction of subscriptions triggers a wave of public customer cancellations.
Similarly, the FinTech company Klarna Reverse course After replacing customer support employees by AI, the move led to lower quality. In a more alarming case, Chatbot from New York City advised to AI-affected business business chatbot to entrepreneurs operate illegal practicesEmphasis of the catastrophic conformity risks of non -monitored energetic ingredients.
These incidents are symptoms of a bigger capability gap. According to a Salesforce in May 2025 Research paperThe leading agents of today have only 58% of the time for individual level tasks and only in 35% of the time for multi-stage tasks and underline “a big gap between current LLM functions and the various requirements for real company scenarios”.
The collegine-in-the-loop model
In order to shut this gap, a brand new approach focuses on structured human supervision. “A AI agent should act of their direction and of their name,” Mixus co-founder Elliot Katz told Venturebeat. “However, without built -in organizational supervision, autonomous agents often often cause more problems than they solve them.”
This philosophy underpins Mixus' colleague-in-the-loop model, which embeds human review directly in automated workflows. For example, a big retailer may receive weekly reports from hundreds of outlets that contain critical operating data (e.g. sales volumes, working hours, productivity and compensation requests from the headquarters). Human analysts should spend hours checking the info manually and making decisions based on heuristics. With Mixus, the AI ​​agent automates the heavy lifting, the evaluation of complex patterns and the anomalies as unusually high salary requirements or productivity outlets.
For decisions with high missions resembling payment permits or violations of guidelines, that are defined by a human user as a “high risk” of the agent and requires the consent of man before they proceed. The division of labor between AI and humans was integrated into the agent's creation process.
“This approach means that individuals only become involved if their expertise actually increases added value-the critical 5-10% of the selections that would have significant effects-while the remaining 90-95% of the routine tasks robotically flow through,” said Katz. “You get the speed of complete automation for normal operations, but human supervision is precisely when the context, judgment and accountability are most vital.”
In a demo that showed the Mixus team Venturebeat, the creation of an agent is an intuitive process that may be carried out with instructions with easy text. In order to create a fact -checking agent for reporters, the co -founder Shai Magimof described, for instance, the multi -stage process within the natural language and instructed the platform to bed steps to confirm individuals with certain threshold values, e.g.
One of the core strengths of the platform are the integrations with tools resembling Google Drive, E-Mail and Slack, with which corporations bring users their very own data sources into workflows and interact directly with agents from their communication platform of selection without changing contexts or learning a brand new interface (e.g. the actual fact checked to send recognition to the editor's request to the editor's email address).
The integration functions of the platform proceed to fulfill certain corporate requirements. Mixus supports the model context protocol (MCP), with which corporations can mix agents with their tailor -made tools and APIs and avoid the necessity to reinvent the wheel for existing internal systems. In combination with integrations for other company software resembling Jira and Salesforce, agents can perform complex platform tasks, e.g.
Human oversight as a strategic multiplier
The Enterprise AI Space is currently going through a reality check because corporations switch from experiments to production. The consensus amongst many industry leaders is that individuals within the loop are a practical need for agents to work reliably.
The Mixus collaborative model changes the economy of the scaling of AI. Mixed predicts that the availability of agents can grow by 1000x and each human supervisor 50 times becomes more efficient if AI agents grow to be more reliable. But the general need for human beings will still grow.
“Every human supervisor manages more AI work over time, but they still need more overall monitoring, because the AI ​​provision explodes of their organization,” said Katz.
For company managers, which means human skills develop somewhat than disappear. Instead of being replaced by AI, experts are promoted to roles wherein they organize fleets of AI agents and address the high-stakes decisions marked for his or her review.
In this context, constructing a powerful human supervisory function becomes a competitive advantage, in order that corporations can use AI more aggressive and safer than their competitors.
“Companies that dominate this multiplication will dominate their industries, while those that pursue full automation will struggle with reliability, compliance and trust,” said Katz.
