As document-intensive enterprises search methods to scale AI with out disrupting core operations, international threat and claims administration firm Sedgwick is integrating new AI instruments into its legacy programs to course of large volumes of case documentation. 

The Memphis, Tenn.-based firm, which employs roughly 30,000 folks globally, not too long ago expanded its “Sidekick+” platform — launched in 2023 to assist claims examiners summarize and analyze paperwork — with new agentic capabilities designed to help claims workflows and discover related info extra rapidly. 

Sedgwick developed the proprietary Sidekick instruments utilizing OpenAI GPT-4 know-how as a part of a broader technique to modernize and scale AI capabilities over time, whereas persevering with to depend on present claims infrastructure. 

“Consider Sidekick because the wrapper that sits round our giant language fashions that provides info to … our purchasers, our examiners, adjusters,” stated Sean Safieh, the corporate’s CIO of world platforms and digital options. 

Associated:Why AI scaling is so laborious — and what CIOs say works

Safieh spoke with InformationWeek about how Sedgwick built-in generative AI and agentic AI capabilities into present claims workflows whereas avoiding disruptions to its core programs and for purchasers.

The constraint: The claims workflow problem 

Claims examiners and adjusters at Sedgwick usually should sift by means of 1000’s of pages of medical data and different documentation whereas working below time-sensitive circumstances. 

For Safieh and his staff, the problem was discovering methods to assist staff course of giant volumes of knowledge extra effectively whereas sustaining the velocity and consistency purchasers anticipate from claims operations.

“How will we help the examiner of their on a regular basis job with instruments to provide them info at their fingertips, as a substitute of them looking out and searching [and] studying documentation?” Safieh defined.

As a result of client-facing applied sciences additionally fall below Safieh’s obligations at Sedgwick, he and his staff needed to take into account how the AI instruments would have an effect on the claimant expertise.

The choice level: Discovering the appropriate AI use case

Given Sedgwick’s scale, the corporate makes use of a variety of applied sciences, together with optical character recognition (OCR), robotic course of automation (RPA) and a claims platform with automation assets to help the claims administration course of, Safieh stated. OCR allows RPA bots to see and scan paperwork. 

With the rise of AI, the corporate sought particular use circumstances for the know-how: 

How may AI shorten the period of time {that a} declare is open? How may it present adjusters and examiners with real-time info to help their decision-making?

Associated:The 12 months of the Chief Integration Officer — SAS CIO

To drive sooner claims decision and ship the appropriate info to customers, Sidekick is designed to simplify AI prompting for examiners and adjusters by dealing with the prompting course of within the background.

“We give them a button that permits them to only get the data that they want, and we deal with the entire prompting within the background,” Safieh stated.

What modified, what stayed the identical: AI layered into present programs

Whereas Sidekick was designed to streamline claims workflows, Safieh and his staff nonetheless needed to take care of the architectural realities of constructing these instruments and integrating them into Sedgwick’s present ecosystem.

Key architectural priorities included:

Sedgwick’s present companies and APIs simplified the combination course of. 

“At Sedgwick, we have already got an ecosystem constructed on companies and APIs,” Safieh defined. “So, whether or not it was getting knowledge from the claims platform or pushing knowledge — a notice, a diary or info … [it] was really comparatively simple as a result of we had a basis of companies already out there to the generative AI, or Sidekick.”

Associated:What does reliable AI seem like in 2026?

Sidekick doesn’t display screen scrape like an RPA software. As a substitute, the platform sits in entrance of the corporate’s giant language fashions (LLMs) to provide a easy, front-end person expertise. 

On the again finish, Safieh and his staff should rigorously monitor efficiency to make sure that Sidekick can question legacy programs with out breaking them.

“What we wanted to do is make it possible for the answer labored and we weren’t overtaxing what you’ll be able to name our legacy programs or claims platforms,” Safieh stated. “Observability monitoring and efficiency monitoring had been issues that we put in place very early on.”

The Sidekick instruments are designed to assist examiners and adjusters do their jobs extra effectively inside Sedgwick’s present decision-making framework. 

Sedgwick has an consumption course of, dubbed sensible.ly, that determines the validity of consumer requests and documentation, resembling a photograph. If a declare meets particular consumer necessities, resembling the sort and quantity, will probably be despatched for automated processing. If it doesn’t meet these standards, it’s despatched to a human. 

The aim is to provide examiners extra time to work together with claimants and make their selections. 

“There’s all the time a human within the loop for decision-making after we’re taking a look at claims processing, and that can proceed as we go ahead,” Safieh stated. 

Sidekick reduces the time it takes for examiners and adjusters to sift by means of documentation. It offers a abstract in a minute or two for a doc that might take 10 to fifteen minutes to learn, based on Safieh. 

Sidekick additionally amplified Sedgwick’s skill to audit claims and supply insights. 

“Usually … you audit a subset of claims, and also you say, ‘Based mostly on this subset and this demographic, we imagine we’re doing the appropriate factor,'” Safieh stated. “Now, with these instruments, we’re auditing each declare.” This contains guaranteeing that the corporate’s selections and the data supplied are acceptable. 

Whereas Sidekick automates extra of the claims administration course of, it has not altered the corporate’s tradition, based on Safieh. “The consensus or tradition … of serving to our claimants of their time of want nonetheless stays and is deeply rooted,” he stated. 

The friction: Constructing belief in AI outputs

Safieh stated he knew he and his staff must construct belief with the individuals who can be utilizing the Sidekick instruments internally. Sedgwick approached constructing that belief by working pilots after which parallel testing.

By pilots, the staff works to make sure the appropriate prompts are in place and that the software offers the appropriate output. Then, by means of parallel testing, the output an examiner or adjuster offers is in contrast with the output from Sidekick.

“In one thing like a doc summarization, we’re seeing 98% to 99% accuracy from what an examiner or adjuster would output [compared] to … Sidekick,” Safieh stated.  

Sedgwick constructed a staff to facilitate conversations throughout its operations and completely different enterprise groups.

“We’ve got to know what the enterprise case is. We’ve got to know what the result is,” Safieh stated. “We constructed [a] staff that is ready to have these conversations in the appropriate method to deal with these use circumstances.”

The price of getting it proper: The guardrails behind the AI rollout

The corporate’s very first GenAI use case with Sidekick took roughly three months to implement.

Sedgwick was already an Azure Cloud person, enabling it to rapidly and securely deploy the OpenAI mannequin in its atmosphere, based on Safieh. 

He described establishing the structure to help safety, efficiency and entry as probably the most time-consuming a part of the method. 

“What might have taken a little bit bit extra time than possibly anticipated was ensuring that, one, we’re getting one of the best outcomes out of the mannequin, and two, how will we deploy this appropriately throughout the org and train customers find out how to use it,” he stated. 

Guaranteeing that the usage of a software like that is proper requires recognition of what may go mistaken and placing mandatory guardrails in place. The corporate handles a large quantity of delicate knowledge in the course of the claims administration course of and should be certain that Sidekick is not going to compromise knowledge safety. 

Key guardrails included:

The corporate ensures that its knowledge just isn’t saved anyplace exterior of its ecosystem. It isn’t used to coach third-party LLMs, and for coaching inner LLMs, the corporate makes use of anonymized knowledge, Safieh defined.

The fashions behind Sidekick additionally want controls for accuracy.

An AI resolution would possibly battle to interpret a doc, resembling a messy handwritten notice, throughout consumption. AI programs are incentivized to offer solutions to prompts, a lot in order that these solutions could also be inaccurate and even hallucinated.

“The best way we have designed the answer is that if it isn’t capable of interpret info, it is going to outline that [inability] as a part of the output,” Safieh defined.

He and his staff have labored to scale back the danger of hallucinations, first by guaranteeing the Sidekick instruments have entry to solely clear knowledge units and second by managing prompting. 

“Sidekick actually manages what we’re asking and the way we’re asking it in order that we now have repeatable responses or repeatable units of knowledge that cut back the general likelihood of an AI hallucination,” he stated.

Provenance is one other key space the corporate has to get proper. If a declare resolution is challenged, Sedgwick should be capable to present how AI was used within the decision-making course of.

“The Sidekick resolution shops not solely the data that is being supplied but additionally the abstract that is output,” Safieh stated. “What we will do at any time is take a look at the 2 items of knowledge that had been supplied to the examiner — what did the precise doc present, what did the abstract present — and what was the advice that the examiner made.”

Safieh, like different CIOs, is cognizant of the vendor lock-in threat within the AI house. With the know-how transferring so rapidly, the power to alter distributors is interesting. Sedgwick opted to architect its Sidekick options with that flexibility in thoughts. 

Sedgwick’s proprietary platforms fulfill the should be nimble, based on Safieh.

“We personal them. We personal the roadmap. We personal the design. We personal the structure,” he stated. “Once we take a look at these options and architect them, we make it possible for they’re versatile sufficient to both swap fashions or leverage a number of fashions in order that we do not get caught in a vendor lock-in.”

Up to now, Safieh hasn’t run into any surprising prices associated to Sidekick, however he stated he’s conscious this might change as extra AI options come on-line and the price of compute climbs. 

“How is the market going to take a look at monetizing these options, and what’s going to we have to do from there?” Safieh requested.

Getting ready to scale AI brokers: Inquiries to reply 

Safieh stated the Sidekick resolution has a stable basis that Sedgwick can use to roll out extra AI capabilities.

“From a scalability perspective, as we introduce increasingly brokers, what we have to do is proceed to verify we now have observability in place, efficiency monitoring in place, and that knowledge safety is in place,” he stated.

Scaling AI additionally raises broader architectural questions on orchestration and workflow design, Safieh added:

The tradeoff: Choices should be made rapidly 

Sidekick has taught Safieh that discovering worth with AI is completely different from different know-how initiatives. As a CIO, he has much less time to get there. It’s all about effectivity.

“Many instances with different initiatives, they’re constructed and executed in a different way, whether or not it is a several-month venture, multi-year initiatives, the place there is a spend allotted to it. You may have time to make sure selections all through the lifecycle of that venture to make things better,” he defined. 

“With such a know-how, you decide in a short time that it really works or it would not work, and you progress on.”

What different CIOs would possibly get mistaken: Do not chase each new AI software 

Safieh stated that in his conversations with different CIOs, he hears loads of pleasure in regards to the many new instruments out there available in the market. He suggested his friends to take a measured method somewhat than chasing the most recent and best providing. Ripping out and changing instruments too usually really slows an enterprise’s velocity, he cautioned.

“You probably have a software that’s working, that meets your wants, keep it up so long as you’ll be able to as a result of it is going to proceed to evolve and alter. If it would not meet your wants, then swap,” he stated.

The one factor he would do in a different way: Construct belief early on

Wanting again, Safieh stated the one factor he would do in a different way is concentrate on person adoption earlier within the course of.

“The sooner you may get folks on board and utilizing the options and trusting them, the higher adoption you are going to have, the higher scalability you are going to have as a result of everybody’s on board,” he stated.