Simple co -pilots are the news of yesterday. Competitive differentiation is to launch a network of specialized agents who collaborate, self -criticism and the call of the good model for each step. The latest episode of the IA Impact of Venturebeat, presented by SAP in San Francisco, addressed the problem of the deployment and governance of multi-agent AI systems.
Yaad Oren, CEO SAP Labs US and global manager of research and innovation at SAP, and RAJ JAMPA, SVP and CIO with Agilent, an analytical and clinical laboratory technology company, have discussed how to deploy these systems in real environments while remaining within the cost, latency and compliance. SAP’s objective is to ensure that customers can evolve their AI agents, but safely, Oren.
“You can be almost entirely autonomous if you wish, but we make sure that there are a lot of control and surveillance points to help improve and repair,” he said. “This technology must be monitored on a large scale. It is not yet perfect. It is the tip of the iceberg around what we do to make sure that agents can evolve, and also minimize all vulnerabilities. ”
Deploy active AI pilots throughout the organization
Currently, Agilent actively integrates AI through the organization, said Jampa. The results are promising, but they are still tackling these problems of vulnerability and scaling.
“We are in a stage where we see results,” he said. “We now have to deal with problems such as, how to improve AI monitoring?” How do we optimize costs for AI? We are definitely in the second step, where we no longer explore. We examine new challenges and how we deal with these costs and these monitoring tools. ”
In Agilent, AI is deployed in three strategic pillars, said Jampa. First, on the product side, they explore how to accelerate innovation by integrating AI into the instruments they develop. Second, on the side intended for the customer, they identify which AI capacities will offer the greatest value to their customers. Third, they apply AI to internal operations, creating solutions such as self-healing networks to increase efficiency and capacity.
“While we implement these use cases, something on which we focused on the governance framework is the framework of governance,” said Jampa. This includes the definition of the limits based on policies and the guarantee of railing for each solution removes unnecessary restrictions while maintaining compliance and security.
The importance of this was recently underlined when one of their agents update, but they did not have any control in place to ensure that its limits were solid. The upgrade immediately caused problems, said Jampa – but the network was quickly detected, because the second piece of the pillar is audit, or ensuring that each input and each output is recorded and can be traced.
The addition of a human layer is the last piece.
“Small cases of use in tiny are quite simple, but when you talk about natural language, large translations, these are scenarios where we have complex models involved,” he said. “For these more important decisions, we add the element where the agent says, I need a human to intervene and approve my next step.”
And the question of speed in relation to precision comes into play early during the decision-making process, he added, because costs can be added quickly. Complex models for low latency tasks push these considerably higher costs. A layer of governance helps monitor the speed, latency and precision of the results of the agents, so that they can identify the opportunities to rely on their existing deployments and continue to extend their AI strategy.
Resolve agents integration challenges
Integration between AI agents and existing business solutions remains a major point of pain. Although systems inherited on site can connect via data APIs or an architecture focused on events, the best practice is to ensure that all solutions work in a cloud frame.
“As long as you have the cloud solution, it is easier to have all the connections, all the delivery cycles,” said Oren. “Many companies have on -site facilities. We help, using AI and agents, to migrate them into the cloud solution. ”
With the integrated SAP tool chain, complexities such as personalization of inherited software is also easily maintained in the cloud. Once everything is in the cloud infrastructure, the data layer enters, which is just as or more important.
At SAP, the Data Cloud company serves as a unified data platform which brings together information from SAP and non SAP sources. Like Google Index Web content, commercial data cloud can index commercial data and add a semantic context.
Oren added: “The agents then have the possibility of connecting and creating business processes from start to finish.”
Make gaps in the company’s agent activations
While many elements take into account the equation, three are critical: the data layer, the orchestration layer and the confidentiality and safety layer. The clean and well structured data are, of course, crucial and successful agent deployments depend on a layer of unified data. The orchestration layer manages agent connections, allowing powerful agent automation through the system.
“The way you orchestrate [agents] is a science, but also an art, “says Oren.” Otherwise, you can not only have failures, but also audits and other challenges. “”
Finally, investment in security and confidentiality is non -negotiable – in particular when a swarm of agents works in your databases and your corporate architecture, where the authorization and management of identity are essential. For example, a member of the HR team may need to access a personally identifiable salary or information, but no one else should be able to view them.
We are heading to a future in which human company teams are joined by members of the agent and the robotic team, and it is at this time that identity management becomes even more vital, said Oren.
“We are starting to look at agents more and more as if they are humans, but they need additional surveillance,” he added. “This implies integration and authorization. He also needs change management.
