Keeping Humans in the Loop

What sets agentic AI apart from previous iterations of artificial intelligence and machine learning?

From my point of view, because it’s so much cleverer, the technology is now much, much more advanced, much more flexible, and much closer to a human brain than it was before. And it’s affordable as well because you can get AI agents ramped up very quickly. Before, you needed a big IT project, typically with lots of approvals, and then you had an unknown outcome. So there was always a risk: can it do what we want it to do or not? The generation now on the market is so different. It’s so fast, but now the difference is, how far can we go? What are the controls, the guardrails for AI?

We saw with OpenAI what happens if a swarm of bots all of a sudden gets together on their own. [AI tool platform Hugging Face announced in July that a swarm of frontier OpenAI agents had hacked into its systems.] This needs to be looked at much more carefully than with the old technology. The previous versions of machine learning and AI were just not capable of such activity. As with every technology, there are good and bad applications. It could be used actively for criminal work like cyberattacks. It’s so much easier to generate hacker code now. Before, you needed real experts in programming, thinking it through, trying options out. So that’s the bad application of the technology.

It’s right that the regulators, also insurers, are asking, “How can we use the technology in a secure and safe way so we, or our customers, can benefit from AI? Which controls, which guardrails, which preventive measures need to be put in place so that incidents like the one OpenAI revealed will not happen?” The technology is far enough [along] now that we must address this question. Before, it wasn’t a big risk.

Where is the insurance industry now deploying agentic AI solutions? Where does Sapiens see that going?

Let me illustrate three scenarios: first of all, agentic AI supports end users where an AI agent can either prepare or execute tasks on their behalf. That’s No. 1. No. 2 is decision-making [and gathering and analyzing data to prepare for decision-making]. Thirdly, when there’s complex policies, complex agreements, claims that maybe need six months to investigate whether an insurer should pay out or not, agentic AI can make the analysis process much faster and can reduce operating costs and, more importantly, improve customer satisfaction.

Aside from these examples, agentic AI can also support technology implementation. This is less about the core product capabilities of AI, but like in banking, migrating or changing a core insurance system can be a significant headache. In the past, it took years to do, three, four, five, sometimes six, seven years. Now insurers are asking us, “How can I complete the same task in 10 months, 15 months, 18 months, in a secure way?”

Customers are also asking how they can use AI now to achieve results faster. Previously, it may have taken a customer eight months to go live on a new insurance product because everything had to be configured manually—can’t AI do it in two days with all the rollout and marketing around it? Maybe a product could go live in four weeks instead of six, eight, 12 months? This is another area where insurance companies want to be more competitive and faster to market with new innovation and ideas.

For insurers and brokers, what are the benefits of agentic AI solutions?

I’m not sure that the brokers and insurers will go straight to agentic AI. They will go in steps. But ultimately, agentic AI will lead to faster response times to inquiries from the outside world. That’s the No. 1 benefit—customer service, customer experience. Internally, it will lead to more efficiencies. You can, for example, accelerate the growth of a business using agentic [AI]. Lots of insurance companies are growing very quickly with the same number of employees thanks to this technology. The third benefit is that it will ensure more standardization of processes. If you need to train, let’s say, 50 brokers on how a specific policy needs to be handled or a claim to be processed, you can teach the AI engine once and be assured that it will follow the rule, always achieving the same outcome.

When we talk about agentic [AI], a specific bundle of tasks can be done by an AI agent instead of relying on an individual, or several individuals, as this will require handoffs and waiting times while individuals process decisions. Agentic AI is also available 24/7, whereas humans require sleep, they have breaks, they get sick. If an insurer adopts this strategy, it will drive a lot of internal efficiencies and scalability. Others are taking a keener focus on the customer experience. They also reap the benefits of much higher productivity, but the priority is customer experience. Others might view agentic AI as an opportunity to act much faster than competitors with less people.

This will be a discussion I see changing as we move forward. When you look at a lot of the debate in LinkedIn articles and elsewhere, previously it was all about cost saving, cost saving, cost saving. This has changed now and is very much focused on customer experience, efficiency, business scalability, standardization. So, only within a few months, you see a completely different narrative about how AI is perceived. This is what Sapiens is doing. We are providing Sapiens AI agents that will offer a variety of benefits, including improved customer experience, faster automatic claim, policy management, underwriting, reinsurance processing, or delivering internal efficiencies.

At what scale are AI agents making decisions now? What are the risks?

At the moment, we support the human-in-the-loop concept. It means that even if an AI agent makes a decision, there will always be a human being to check it until we are really confident in the output. The reasons AI agents make mistakes depend very much on the modeling, the data, and the vendor’s experience in the industry. The agent needs to know what the right decision or the wrong decision is, and if there are gray zones, trigger an alert and ask a human.

It’s less about the agent than the brain that needs to be linked to the agent so that it can make the right decision. For example, one of our big strategic focuses is less on the technology of an agent than how we collate all the knowledge of the last 50 years in the insurance industry. This is what we call our “brain,” which the AI agent has access to. If this is properly set up and the experience is there, the probability that an AI agent will make a bad decision decreases. You should also be transparent with regulators about this oversight, because we have not yet reached a point where AI agents are achieving 99% accuracy in decisions.

My personal belief, and I’ll say Sapiens’s as well, is that insurers and brokers will adopt agentic AI in phases, trying it out first and monitoring how often an AI agent or an agentic AI workflow makes right and wrong decisions. If they reach the point where 95% or more of the decisions are right, then they may be willing to…reduce the reliance on the human-in-the loop supervision or at least limit its use to exceptions or big deals and complex tasks. This will happen in the future. Maybe in one year, two years, or three years, basically as soon as everybody is confident the AI decision-making is accurate.

Beyond maintaining human involvement, are there other guardrails?

Yes, we [Sapiens] also decided to not have an agent that can do everything. We have been working on an agent framework that does very specific tasks and then hands over to another agent. We would like to avoid an agent accumulating all the knowledge of everything. We believe that separating the agents into smaller ones that have only limited access to the knowledge base, to the brains, will help us ensure they don’t become too intelligent. They need to ask another agent to continue the task. This sequence of agents will ensure that mistakes do not happen, or validation agents will identify errors, for example saying, “I have a doubt that this is the right decision based on the policy I just validated against your proposal, human being, please check it.” In addition, we provide a sophisticated audit trail in our products, so that every step, every decision that an AI agent made is logged in the audit trail, available for control and compliance checks.

Bill Gates, OpenAI CEO Sam Altman, and others have warned about the technology following the AI swarm attack from OpenAI on Hugging Face. How concerned should businesses and society be about AI agents going rogue?

>p>First of all, both Gates and Altman are right because technology can always be used for good and bad things. For example, a drone could monitor the well-being of your livestock. Do they have enough water? That’s the positive thing. But it can also be used as a weapon.

The technology is ready [enough] that it can be misused. Before it wasn’t, so we need to ensure there are these guardrails. We cannot just go blindly into the world and believe nothing negative can happen. We need to be really conscious of what we are doing. We need the guardrails, the regulations. We have to think about segregation of duties in the agent world, so that the risk can be limited. As a society, we need to take care, and the rules of the game should be defined independent of the insurance industry. From a governmental regulation point of view, there should be a discussion about what’s allowed, what’s not allowed, and what the sanctions are when somebody misuses this technology.

I don’t think you can stop it, and AI can provide so many benefits in cancer research, in patient treatment. It will change society, the world, but it must also be watched and treated with care.