Ali Merchant speaks with healthcare operations leader Michael Fritts about where AI can create real impact in health information management without compromising accuracy, compliance, or patient trust. Michael explains why AI should act as an operational exoskeleton, helping teams process information faster, identify patterns, handle repetitive work, and scale high-volume workflows while humans remain responsible for judgment, oversight, and mission-critical decisions.
The conversation becomes especially compelling around voice AI. Michael argues that an accurate answer from an AI agent in 30 seconds can be more valuable than making a customer wait 10 minutes for a human. The session shows how AI can resolve routine inquiries, reduce hold times, improve quality coverage, transfer complex cases with full context, and create more meaningful work for human agents. It also offers a practical roadmap for AI adoption: start with a defined use case, establish governance, maintain traceability, and choose a true partner for mission-critical operations.
AI works best as an exoskeleton that improves speed, scale, and accuracy while humans provide judgment and oversight.
Fast, accurate AI support can deliver a better customer experience than long hold times for routine inquiries.
Status checks, general inquiries, follow-ups, and repetitive calls are strong early automation opportunities.
Urgent, emotional, exception-based, and high-risk situations still require human attention and decision-making.
Approved platforms, secure data handling, human review, documentation, and traceability are critical in healthcare AI.
Organizations should choose experienced partners that can deploy, support, scale, and share responsibility, not just provide software.
Michael Fritz:
To me, AI is going to be more like an exoskeleton that humans can strap on.
Is it better for somebody to speak with an AI agent that may not be 100% perfect but provides an answer within 30 seconds, or should they wait on hold for 10 minutes to speak with a human?
As you improve operations with other AI tools and shorten turnaround times, you also reduce the number of follow-up calls.
Ali Merchant:
If the process is mission-critical, you need a partner, and they need to behave like one.
Michael Fritz:
Exactly.
Introduction
Ali Merchant:
Hello, everyone. Welcome to Enterprise AI by CallBotics, a conversation series where we speak with enterprise leaders about how they think about AI and how it is shaping real-world operations.
I am Ali Merchant, and today I am joined by Michael Fritz.
Mike is a senior executive in the health information space and has spent his career leading complex healthcare and technology operations, including leadership roles at GE Healthcare and other high-growth organizations.
Mike, it is great to have you here. Thank you for joining us.
To begin, for those who may not be familiar with your background, could you share a little about your journey and how it has shaped the way you think about running large operations?
Forty Years From Scanners to Healthcare Operations
Michael Fritz:
My journey is an interesting one.
I am an electrical engineer by training, but I began my career carrying a tool bag and working on CT and MRI scanners at client locations.
My initial experience involved working around patients, clinicians, and hospital teams. That experience informed my journey over approximately 40 years in the healthcare industry.
I saw firsthand that when things go well, they can significantly improve outcomes. When things do not go well, they can hurt patients, clinicians, and hospitals.
That gave me a healthy respect for what it takes to run and scale a business, create products and services, and ensure that those products are repeatable, compliant, useful, supportable, and usable.
These solutions may ultimately be used across tens of thousands of facilities around the world.
My journey took me from being an individual contributor inside hospitals to making decisions and managing the operations that drive large businesses.
The main questions that continue to guide me are:
Is this going to help the patient? Is it going to help the clinician? Is it sustainable? Is it supportable? Is it repeatable? Is it scalable?
Ali Merchant:
Absolutely. I agree with that completely.
Let us start with the basics.
For someone who is not closely familiar with this industry, how do release of information operations actually work?
How Release of Information Operations Work
Michael Fritz:
Approximately half of hospitals and clinics still manage release of information operations internally, while the other half use specialized companies to manage the process.
When someone submits a lawful request for health information, the request must be presented correctly and include the required information.
The requester needs to have a valid reason for accessing the records, and the request must clearly specify what information is required. That may include particular dates, laboratory results, treatment records, or other categories of information.
It is extremely important that the correct information is released.
When you think about HIPAA and protected health information, you want the process to be fully compliant and secure.
You do not want health information to be hacked, disclosed incorrectly, or released to someone who is not authorized to receive it.
Hundreds of millions of pages of health information are shared every year. It is difficult to get every transaction right, every time.
However, when you introduce the right technology, train people properly, establish effective processes, and maintain strong compliance and cybersecurity controls, you can perform the work very effectively.
The requester may be an insurance company, an individual requesting their own health information, a hospital such as Mayo Clinic requesting records for continuity of care, a disability claims organization, a law firm, or someone acting under a subpoena.
Whether hospitals perform the work themselves or use an external agent, it must be done correctly every day.
It is a difficult job, but when it is managed properly, health information can be shared quickly, accurately, securely, and compliantly.
That is what organizations in the release of information industry do. We act as agents for healthcare facilities and ensure that these processes operate continuously.
Doing the work correctly requires diligence, training, and increasingly, technology.
The question now is how we can use technology to complete this work more quickly while improving accuracy and compliance.
AI as an Exoskeleton
Ali Merchant:
AI has improved significantly over the past year. Even recent versions of systems such as Claude have become much more capable.
However, health information requires an extremely high level of accuracy.
The severity and probability of something going wrong should determine how AI is introduced into a workflow.
How do you think organizations should approach this?
Michael Fritz:
I think there are two broad categories.
The first involves developing AI tools that will be deployed operationally.
When doing that, you must clearly define the individual work steps and processes the AI will perform.
For example, is the AI looking for a needle in a haystack? Is it analyzing a specific type of information or handling a narrowly defined task?
If you are training a model to find a needle in a haystack, you need to be careful about the data used to train it.
We do not use our clients’ information for that purpose. We use synthetic data created specifically for model training.
Pattern detection and finding important information inside large datasets are areas where AI can be very effective.
However, you must think through the process, test it, document it, and create traceability around what the system is doing and why it is doing it.
Whether the AI is identifying unusual information or performing daily operational work, you should be able to go back and ask, “Why did the system make this decision?”
The problem with AI is that it can hallucinate.
In medical terms, it can produce false positives or false negatives. It may miss something it should have identified, or it may introduce information that should not be there.
That is why I strongly believe humans must remain in the loop.
Humans are needed during development, deployment, management, and ongoing monitoring.
Someone must confirm that the outputs and outcomes are correct.
You should be able to see what work was completed, how it was completed, and how the AI behaved after deployment.
Across a broader organization, you should also have an AI champion or group of AI experts who can help teams use AI properly.
Those people should ensure consistency, establish operating mechanisms, and prevent AI adoption from turning into the Wild West.
It is easy to say, “AI is a wonderful tool. Go out and use it.”
However, we do not allow people to place company information into any AI platform they choose.
Employees must use approved platforms that the company has contracted with and reviewed for security.
This applies not only to client information but also to the company’s own information.
Employees should not upload company data into personal versions of Gemini, Perplexity, Claude, or other public tools.
There must be a policy explaining which systems are approved and how AI should be used.
In operations, you also need to be careful that AI does not introduce errors or slow down the process.
We process large volumes of information very quickly. It is similar to a high-speed production line that must operate correctly every time.
You do not want to deploy AI and later discover that it has repeatedly produced incorrect results.
Without humans in the loop, the same mistake can be repeated at scale before anyone notices it.
Ali Merchant:
What are your broader thoughts on the relationship between AI and humans?
Michael Fritz:
To me, AI will be more like an exoskeleton that humans can strap on.
It can act as a second brain, take over repetitive tasks, and identify patterns across complex or fragmented datasets that would be difficult for a human to detect manually.
However, humans are still needed to evaluate the results, understand the context, and ensure the AI is not hallucinating or overlooking something important.
Humans will continue to play an essential role.
When it is implemented correctly, AI can help us scale, complete work faster, and improve accuracy.
In that sense, we are creating superhumans with second brains.
The people and organizations that understand and embrace that opportunity will be the ones that succeed in this new AI environment.
Even something simple, such as using AI to help draft emails, can be a reasonable starting point.
Highest-Leverage AI Opportunities in Records Operations
Ali Merchant:
Let us get more specific about release of information operations.
The operational cycle includes receiving medical record requests, identifying the correct records, arranging them chronologically, checking for commingled records, making calls, receiving calls, and ultimately releasing the information.
Based on your experience, where do you see the greatest opportunity for AI in release of information today?
Michael Fritz:
The easiest areas to identify and improve initially are the incoming information and the entry or keying of outgoing information.
The largest long-term impact will come when AI can process the health information itself because that is where most of the labor exists.
Ali Merchant:
That is an interesting point.
The largest part of health information management is nonvoice work. It involves receiving records, determining which records are required, reviewing them, and releasing the correct information.
CallBotics was created from our other company, AM Infoweb, where close to 2,000 people perform this type of work.
Our first instinct was to build AI for the nonvoice side of the operation.
However, we discovered that AI was not yet capable of handling the entire process reliably.
It may be able to manage smaller parts of the workflow, such as moving information from one place to another, but healthcare information cannot operate at 85% accuracy.
Michael Fritz:
Exactly.
Ali Merchant:
It needs to be close to 100%, or at least 98% to 99%, which is comparable to human accuracy. Otherwise, it does not work.
Michael Fritz:
A patient’s physician notes may be stored in one location while other information is stored elsewhere.
To properly support someone’s healthcare, prior authorization, referrals, and treatment, you need access to information across many different sources.
That is where the process becomes incredibly complicated.
A human may recognize that certain information is missing and understand that the available documentation will not satisfy the clinician’s requirements.
It is difficult to train an AI system to understand that completely.
Ali Merchant:
That is an excellent point.
We spent a lot of time exploring this when AI started becoming more widely available.
We had an excellent engineering team and 18 years of operational context, but we eventually paused the broader nonvoice automation effort.
That is when we began developing CallBotics.
A significant part of health information management and litigation support involves making calls, following up with facilities, and processing payments.
We already performed those activities for litigation support companies.
On the release of information side, organizations also receive high volumes of calls relating to inquiries and requests.
That is where voice AI proved especially effective.
Automating 80% to 85% of Calls
Ali Merchant:
We were able to automate approximately 80% to 85% of calls in certain operations.
The important part was that we could also conduct quality review across 100% of those conversations.
The AI had access to the relevant data and could use it during the interaction.
Michael Fritz:
Calls are an important area.
When I first started in this role, we had a high call abandonment rate and very long wait times.
The question is whether it is better for someone to reach an AI agent that may not be 100% perfect but responds within 30 seconds, or for them to wait on hold for 10 minutes to speak with a human.
When a large percentage of callers are waiting, immediate access is extremely valuable.
Even if AI resolves only 50% of the calls, that can still create a significant benefit.
We are introducing AI agents for initial call handling and general inquiries.
For example, if someone is asking, “What is the status of my medical record release?” an AI agent can often answer that question.
If half of your call volume consists of status inquiries, and AI can resolve half of those inquiries, that represents a major operational improvement.
It allows human agents to focus on calls that require deeper conversations.
Those may include billing questions, urgent requests, or situations in which a patient’s surgery has been moved forward and records are needed immediately.
Those situations may require a human.
However, for most routine status inquiries, AI can be the better solution.
Many people do not necessarily want to speak with another person. They are comfortable interacting with a chatbot or AI agent as long as the experience is smooth and they receive an accurate answer quickly.
Ali Merchant:
That is correct.
The AI agent also needs to have a certain personality.
It should respond quickly, demonstrate empathy, understand what the caller is saying, and avoid speaking over the caller.
It should not pause for 10 seconds before every response or create an uncomfortable delay.
Michael Fritz:
That lack of delay is important.
Another benefit is that an AI system can remain on the phone for half an hour without becoming frustrated.
You do not want a human agent spending half an hour waiting for someone else to answer a call.
Ali Merchant:
Exactly. That creates both operational and financial benefits.
Customer Satisfaction With AI-Handled Calls
Ali Merchant:
What excites me is the impact on continuity of care.
Sometimes records are being transferred from one provider to another. If the records do not arrive, someone’s treatment may be delayed.
In those situations, the financial benefit is secondary. The organization must ensure that the records arrive on time.
We knew that business users were often comfortable speaking with AI agents.
The next question was how individual consumers or patients would respond.
We tracked customer satisfaction scores and conducted surveys across a small call queue with one of our partners.
Most callers were satisfied, even when they knew they had interacted with an AI agent, because the process was fast.
Over time, we automated approximately 80% of the calls.
For the remaining 20%, human teams had more time to invest in complex cases and support customers who were struggling.
When the AI agent could not resolve the call, it transferred the conversation to a human agent with the full context.
The customer did not need to repeat everything.
The human agent could see the history of the interaction and the suggestions provided by the AI.
Sometimes the AI simply did not have access to the information needed to resolve the case, so a human needed to take over.
This allowed human agents to perform more proactive and meaningful work.
They were also more satisfied because they were helping people and solving real problems.
Michael Fritz:
Solving meaningful problems gives people a strong sense of satisfaction.
The alternative is that the human agent answers a call from someone who has already spent 10 minutes on hold and is frustrated, impatient, and demanding an immediate answer.
Repeated exposure to those interactions contributes to contact center burnout.
If AI can handle straightforward inquiries quickly and effectively, the human agent only receives the calls that genuinely require attention.
The human then has the time and ability to calmly help the caller.
The caller is also less frustrated because they were able to reach the right person for their specific issue.
That is the experience you want to create.
Faster Turnaround Creates Fewer Calls
Michael Fritz:
Consider someone calling for the third time because they are about to undergo surgery and the required records have not arrived.
For provider-to-provider retrieval, those records may be needed for treatment.
If the information does not arrive, or the provider is uncertain whether it has arrived, physicians may personally begin calling to follow up because the information is important.
These calls become the frontline connection between patients and healthcare facilities.
Patient frustration can quickly reach the CEOs and operations leaders of those facilities.
If callers struggle to understand the agent, experience long hold times, or feel that their concerns are not being handled, the issue can escalate very quickly.
However, when the process is handled effectively, the situation becomes calmer.
As organizations use other AI tools to improve processing and reduce turnaround time, they also reduce the number of inbound calls.
There is a virtuous cycle.
You improve call handling, complete the underlying processing faster, release the information to the correct people sooner, and close the loop.
That reduces the need for follow-up calls.
Calls are handled more effectively, customers have a better experience, and clients hear positive feedback.
Poor call handling can cause an organization to lose clients.
Strong call handling may help it retain or gain clients.
In many cases, however, doing the work correctly is simply the minimum requirement for participating in the market.
Many organizations still do not manage release of information calls effectively.
Ali Merchant:
This principle also applies beyond healthcare.
When you look closely, implementing AI can reduce a significant amount of repetitive work created by operational inefficiency.
Michael Fritz:
Absolutely.
When operations have extended timeframes and inefficient processes, the organization pays an additional operational cost.
If you make the process faster, more effective, and more efficient, you increase throughput.
It is similar to lean manufacturing.
When you shorten cycle times, divide the work into proper steps, and create smooth operational flow, fewer errors propagate through the system.
You reduce wait times and lower the probability of errors at scale.
That creates another virtuous cycle.
Faster processing reduces the burden on the teams responsible for supporting those processes.
That is one of the significant benefits of AI.
Client Expectations Around AI Governance
Michael Fritz:
Clients are also becoming more prescriptive.
They want to understand how organizations are using AI, and they want reassurance that their information is being handled properly.
They ask how AI is being trained, how the system operates, and how data is protected.
Organizations need to educate clients about what they are doing and why they are doing it.
In some cases, AI usage also needs to be addressed contractually.
Companies must explain how AI is used and make sure the appropriate terms are documented.
Some client boards are encouraging leadership teams to adopt AI more aggressively.
In those cases, a service provider may help clients use AI through the provider’s own platform or managed services.
At the same time, clients’ legal, compliance, and security teams may be concerned about AI.
They may ask whether an external service provider is using AI properly, documenting its usage, and maintaining security and compliance.
These questions are increasingly part of conversations with both existing and prospective clients.
Where Organizations Should Start With AI
Ali Merchant:
Let us imagine that an organization is still early in its AI journey and feels overwhelmed.
Where should it start, and how should it approach AI adoption?
Michael Fritz:
AI has changed so rapidly that becoming an expert two years ago would be very different from becoming an expert one year ago or even six months ago.
In some ways, being early in the journey today is an advantage.
Now is a good time to begin because AI capabilities have improved and a large amount of educational information is readily available.
Start by learning the terminology, language, platforms, and differences between an agentic AI agent and a general AI tool.
The appropriate approach depends on the application.
The foundational question should be: What do I want to accomplish, and why do I want to use AI for it?
Then educate yourself about whether AI can actually help.
Map your processes.
It is difficult to use AI effectively when you do not understand the existing process or the problem you are trying to solve.
Once the objective is clear, determine whether you are developing something, using AI to manage an existing activity, or introducing a new operational capability.
Organizations should identify which tools, agents, and platforms will be used for which purposes.
They should appoint an AI champion or someone responsible for overseeing the program.
You cannot simply say, “IT will tell us what to do about AI.”
If everyone owns AI, no one owns it.
You need an owner, a process for evaluating tools, and a clear view of what the organization wants AI to accomplish.
You should also classify the use case according to its accuracy requirements and risk.
If AI does not draft an email perfectly, the consequence may be small.
If it creates an imperfect PowerPoint chart, that may also be manageable.
However, health information, financial information, audit data, or other mission-critical information must be handled correctly every time.
The use case determines the required controls.
Organizations should familiarize themselves with available tools and decide which tools employees are permitted to use.
They should not allow a Wild West environment where everyone uses whichever platform they prefer.
You must ensure that employees are not placing sensitive data into free public AI tools where the platform may use that information for other purposes.
When an organization pays for an enterprise AI service, there is often a contractual expectation that the organization retains control of its data.
AI adoption should therefore include planning, control, documentation, and traceability.
You should be able to explain how the system is being used, why it is being used, what it produced, and how decisions were made.
That documentation may not seem important today, but it will become valuable later.
A year from now, someone may ask:
Why did we use that tool? What did we use it for? How did it reach that conclusion? How was the software created? Did we validate it correctly?
You will appreciate having that information.
My advice is to educate yourself about the terminology and platforms, understand why and how you want to use AI, establish operating mechanisms, and appoint champions who can develop expertise and train others.
In other words, train the trainers.
Ali Merchant:
That is excellent advice.
Choosing an AI Partner
Ali Merchant:
When adopting AI, organizations may also need to choose a partner that can help deploy it.
How should a company evaluate that decision?
Michael Fritz:
There is always a question of what should remain in-house and what should be outsourced.
Outsourcing may appear to cost more financially, but it can save time and effort while improving the quality of the result.
The organization needs to decide which capabilities are important enough to own internally.
Some capabilities may represent the company’s crown jewels and be critical to its long-term success. Those may need to remain in-house.
Other capabilities may not be commodities, but an external organization may still be able to deliver them better, faster, and more efficiently.
The external organization may have greater scale, global capabilities, stronger pricing, a better technology platform, or deeper expertise.
In those cases, building the capability internally may not be worthwhile.
The same principle applies to AI.
Organizations should determine which capabilities they truly need to understand, develop, and retain internally, and which capabilities can be provided more effectively by a specialist that has already built, scaled, supported, and maintained the technology.
In some cases, an organization may begin by developing an internal prototype.
That helps the team understand the problem and evaluate what good performance looks like.
Once that knowledge has been developed, the organization may choose to engage an external partner.
Therefore, the journey may begin internally and later transition to outsourcing.
In other situations, the capability is already well established, and it makes more sense to engage an expert immediately instead of trying to build everything internally.
Vendor Red Flags and Partner Green Flags
Michael Fritz:
There are also red flags and green flags when evaluating potential partners.
You should determine whether the organization is newly formed, lacks structure, or has limited process maturity.
You want a partner with established expertise, industry knowledge, operational capabilities, and experience.
You also want someone you can trust.
You can often tell from the way a company communicates and behaves whether it is approaching the relationship as a true partner.
Sometimes a vendor focuses heavily on contractual protections and obstacles designed to protect itself.
A partner behaves differently.
A partner says, “We are going to work together and figure this out.”
There is nothing inherently wrong with using a vendor for a simple commodity.
For example, the company restocking a vending machine can operate purely as a vendor.
However, if an organization is helping support patients, clients, employees, or mission-critical operations, you need a genuine partner.
A good partner will bring problems to your attention when they notice them.
They will work with you through the learning curve.
They will not spend all their time protecting themselves.
They will share the risks and rewards of the relationship.
You can evaluate whether they are professional, competent, capable, and trustworthy.
References are also important. Ask what other clients say about working with them.
Observe how they manage negotiations, how they perform the initial work, and how they scale after implementation.
Problems will occur.
The important question is how the partner responds.
Do they become defensive and protective, or are they open and willing to solve the problem collaboratively?
If you find a partner that works through problems with you and has the capacity to scale, you have probably found the right fit.
Mission-Critical AI Needs a Partner
Ali Merchant:
What I am hearing is that if the work is mission-critical, a traditional vendor relationship is not enough.
You need a partner, and that organization must demonstrate the qualities of a partner.
As the saying goes, actions speak louder than words.
The organization’s actions will show whether it truly behaves like a partner and whether it deserves your trust and the opportunity.
For someone listening to this conversation, the key takeaway may be:
If it is mission-critical, you need a partner, and they need to behave like one.
Michael Fritz:
Exactly. That is completely right and an excellent way to express it.
Closing
Ali Merchant:
Mike, thank you very much for sharing these insights.
This conversation was extremely informative, and I sincerely appreciate your time.
Michael Fritz:
Thank you for having me. I enjoyed the conversation.
Ali Merchant:
Likewise. Thank you so much.
