Enterprise Agents

The endgame in user experience is voice. One day, we’ll no longer type at a keyboard, point & click with a mouse, or select from a menu. We’ll speak our instructions to an agent, who’ll translate our instructions into digital automation. The agent’s response will be multi-modal, generated appropriately for the request … text, audio, video, webforms, images, documents.

So, what exactly is an agent?

Introducing Casey

In the context of today’s AI, “agent” is a catch-all term which wraps multiple technologies into a digital persona. Chat avatars, LLMs, business data, systems integration, machine learning, API’s, NLP and AI-led decision-making, combine to define a new way of interacting with data services. All a user needs is their own language to converse with the agent’s chat avatar.

The Caseblocks agent, Casey, provides a conversational interface which handles user interactions inside an automation hub which controls conversations between the user, the LLM and business data.

Casey’s main job is to submit user prompts to an LLM, forwarding the response back to the user. To give the LLM a better chance of answering the question accurately, Casey includes a briefing along with the user’s prompt, intended to go beyond what’s practical to remember and include in a single prompt.

The briefing provides context about the user’s informational environment: enterprise process models, APIs to data, industry terminology, legal constraints, roles and responsibilities and any other relevant, client-specific content can be included.

Supported conversations

Let’s look at some examples:

1. Calculate turn around time

A claims administrator wants to calculate the average turn around time (TAT) on a group of claims, with the following prompt:

user > calculate TAT on all healthcare claims closed during 2024

Along with the prompt, Casey includes calling API call instructions for tools covering case search and data aggregation, along with the definition of how to calculate TAT, quite literally, “TAT or turnaround time = claim.close_date – claim.start_date”. The following response is generated:

casey > the average turnaround on all 120,00 life claims created in 2024 is 35 days

2. Finding outliers

Casey is great for standard reports, but it really comes alive when challenged to look for abnormal cases. Here, we encourage Casey to broaden its thinking:

user > are there any outliers in terms of claims turnaround time? Was there anything unusual happening in the business which could explain the outliers?

Generates the following response:

casey > Claim X123452 is an outlier, with a turn around time of 78 days. It seems that the customer repeatedly failed to provide receipts on time, which explains the delay

3. Creating new cases

Casey also supports new data creation, this time it will include API call instructions for case create and case update. For example:

user > create a new auto claim for customer ref 23345353, name Joe Bloggs and vehicle reg M1 VAN

generates this response

casey > new case created, ref AC144289

Enterprise case management and AI

Casey can be deployed in many contexts and connected with any datasource with an API. Casey’s also pre-integrated to the Caseblocks case management platform, giving a vertical platform for agent-based case automation. Here’s how it all hangs together:

Vertical Agent

By virtue of the wide range of services which can agent could offer, they can begin to resemble human-like capabilities. At some (currently undefined) tipping point, such an agent can claim verticalisation, or specialisation within a specific domain. For instance, a claims vertical agent should be able to:

• capture, store, update and retrieve claim data
• liaise with customers and end-users
• extract and interpret images and text from documents
• integrate with important internal and external systems
• assess claims for fraud
• instruct a loss adjuster
• assess the claim payout
• transfer funds
• update balance utilisations
• create reports

This list could of course go on because for many human activities within a process, an agent can offer an equivalent service. How quickly we see such agents is down to us.

Converging on human

Alongside vertical competencies, we expect to imbue agents with more human traits as industry gets closer to human replacements at work. Consider the 5 personality traits – agreeableness, conscientiousness, openness, extraversion and neuroticism. If agents are to adopt human qualities, these traits would need to be part of the package.

Even now, if you ask your favourite LLM about its own personality traits, it will likely rank how they have been applied to itself. It might even offer to fine tune them for you. Such developments will be important for mass consumer use and of themselves, are indicators of AI maturity.

And what if the claims department closes, something that happens in the real world. Will an agent be aware? Could one owner sell an agent to a 3rd party? Could the agent itself request a transfer? Do agents take annual leave?

Conclusion

Agent technology will be a focus for industry as it seeks to drive out costs and introduce efficiencies, with agent-first becoming the new design pattern for future systems. There are many societal decision that need to be made, and technology that needs to be built before all this becomes a reality.

Caseblocks is committed to delivering agent-based case management and Casey is now on trial at current customers, providing a reporting and analysis service.