Not everything is an AI Agent

These days it seems like the only thing everyone on the internet is talking about is AI Agents automating everything. But at DataBraid, we're hearing a lot of confusion. What is an AI Agent, really? And more importantly, what can they actually do for your business today, especially in a regulated environment like the insurance industry where deterministic outcomes matter?¹ We're seeing a surge of "AI Agent" labels slapped on everything from chatbots to simple LLMs integrations, as the hype around agentic systems grows. Let's cut through the cacophony and clarify what these systems really are, and what the industry needs to build the agentic future we are being promised.

What is an AI Agent?   

Broadly speaking, agentic systems leverage LLMs to automate complex workflows and repetitive tasks reducing human error, scale seamlessly to handle increasing volumes of interactions, dynamically adapt to user specific requirements, integrate effortlessly across platforms and tools, and even operate autonomously within set boundaries. 

When discussing agentic systems it is useful to draw a distinction between an AI Agent and and a data orchestration that is AI enabled. Anthropic categorises agentic systems from an architectural perspective to include:

1. Workflows - systems where LLM calls and other traditional tools are orchestrated through predefined code paths. A simple example would be a data orchestration that retrieves unstructured data files from a third party source, parses the files to complete an online form or answer a set of questions using the data in the file without any intervention from a user.

2. AI Agents - systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish their tasks. Think systems involving multiple LLMs directed by an LLM. These systems are autonomous, flexible and are used to solve open-ended problems where it is difficult to predict the number of steps required to arrive at a solution, or a predictable path to get to the intended solution.²

The differences between these two categories of agentic systems require adopters to evaluate (i)  tradeoffs between speed of the solution (latency) and cost for better performance, (ii) the complexity of development and maintenance of the resources, (iii) the scalability of the solution in an enterprise context, and (iv) data privacy and security concerns. Therefore picking the right tool for the job, particularly in the enterprise context is critical. Using an AI Agent to solve a problem that can be solved by a simpler solution such as an Workflow or even a API call is not advisable in a regulated enterprise context where accuracy is critical and task outcome has real consequences (binding an insurance policy for instance). From our perspective, AI Agents can be powerful tools when the tasks they are performing are complicated, cannot be solved by simpler tools, and most importantly the cost of error is low. Regulated tasks in a space like insurance, where the cost of error can have very real customer and company consequences are certainly not appropriate for AI Agents as of the date of this writing. While the day that AI Agents become ubiquitous in the enterprise context might not be very far away, we believe that the date might be further in the future than the cacophony of “expert” voices on the internet might lead you to believe.

Vertical Agentic Systems

So far we have only discussed general purpose agentic systems. It is important to recognize that a general-purpose solution, while powerful, may not be the optimal solution for all enterprise needs. Just as vertical Software-as-a-Service (SaaS) solutions emerged to address the specific requirements of particular industries, there is a growing need for vertical agentic systems. These specialized systems can be tailored to the unique workflows, regulations, and data structures of industries such as insurance, finance, and healthcare. By focusing on a specific domain, vertical agentic systems can deliver more accurate, efficient, and reliable outcomes, addressing the critical need for deterministic results in regulated environments.

The development of vertical agentic systems allows for the integration of industry-specific knowledge and rules, which are essential for handling complex tasks with real-world consequences. By incorporating domain-specific data models and workflows, vertical agentic systems can minimize errors, ensure compliance, and optimize performance in a way that general-purpose agents cannot.

Infrastructure is Key 

One of the biggest misconceptions we hear when discussing agentic systems is that they are somehow going to be a silver bullet for industries with legacy systems and siloed data. It is important to remember that agentic systems operate in conjunction with existing data and operational infrastructure. Implementing vertical agentic systems in the insurance industry face significant challenges given the often outdated and fragmented data infrastructure present in many organizations. Legacy systems and siloed data sources make it difficult for AI agents to access and process information effectively. 

It is also important to remember that these systems operate on probabilistic models, and what users perceive as a “system getting smarter” when performing deterministic tasks is really an improved probability in the system’s ability to achieve an “expected outcome" particularly at scale. This is achieved by using inference and retrieval augmented generation systems (RAG) which require modern data orchestration infrastructure. A lack of this infrastructure results in increased complexity and risk of failure, making it harder to ensure the deterministic outcomes required in a regulated environment such as insurance.  

Why does this matter for the insurance industry?

Simply put, the state of data infrastructure in the insurance industry is abysmal. In an industry that provides its service through a network of service providers acting in concert the fundamental lack of modern data infrastructure is surprising. Software and industry tools built as walled gardens have stifled innovation to the detriment of industry operators and consumers. While there might be some merit to the argument that this is a feature and not a bug, we believe that insurance’s multi-modal data infrastructure needs³ are not a simple challenge if the solution is dependent on industry consensus i.e. standardization that offers no incentives to those with an incumbent advantage to participate. Agentic systems offer the industry a panacea in this regard.

However, if we want innovative companies to tackle vertical agentic solutions focussed on the insurance industry then they need to be able to focus entirely on their primary mission. If every new entrant has to start by figuring out how to get access to insurance data and building data infrastructure they become consultants to their first customers and never get the chance to build products at scale.

Where does DataBraid fit in?  

The ability to seamlessly connect, transfer and share siloed data across disparate disconnected systems and organisations has held the pace of innovation in the industry back, and is easily identifiable as the root cause of the process, service and operational inefficiencies that customers and operators within the industry experience every day.  Traditional approaches to these problems have been to automate the tedious process by building some “tool.” This approach is flawed and pretty soon results in an army of “tools” that now need to be managed and maintained — more tedious process and cost. The tedious process and operational inefficiency are symptoms. Stop treating the symptoms and eliminate the need for the process entirely with modern infrastructure. 

At DataBraid we are on a mission to build modern data orchestration infrastructure for insurance. We build connections through data orchestration, and our initial focus is the policy servicing where the burden of data collection, collation and re-keying falls on the broker. Get in touch if you would like to learn more about how our infrastructure-as-code solution that can revolutionize your connection with your customers and carriers.

Footnotes:

❶ We highly recommend "AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference" by Arvind Narayanan and Sayash Kapoor, two computer scientists and professors at Princeton University, which provides readers with a critical understanding of AI's capabilities and limitations; particularly in an enterprise context.

❷ See Anthropic’s blog post Building Effective Agents for a technical overview of this distinction.

❸ We refer to insurance data infrastructure as being multi-modal because it is distributed across multiple service providers and the customer. Service providers within the value chain try to create “single sources of truth" on their internal systems. If the data is subject to change, without real time data infrastructure these “copies” are fatally flawed and error prone.