Every time a patient visits a new doctor, a school needs to verify a student's vaccinations, or a pharmacist has to check drug interactions, the same fragmented process plays out. Someone asks for a form. A record is faxed. A parent digs through a filing cabinet. A clinician works from incomplete information.

This is not a niche inconvenience. It is a systemic failure at the foundation of how modern health systems operate — and it carries a cost that goes far beyond inconvenience.

The Hidden Cost of Paper-Based Health Identity

The consequences of fragmented health records are well documented but consistently underestimated. Consider the chain of events that unfolds when a patient arrives at an emergency department unconscious, without identification, and without any way to communicate their medication history or known allergies. A clinician makes decisions with incomplete information. That is not a rare edge case — it is a daily occurrence in emergency medicine.

The same dynamic plays out in less dramatic but equally consequential settings:

• A school nurse cannot verify whether a student's vaccination record is current without contacting the family and waiting for a paper form
• A pharmacist filling a prescription cannot see the patient's other active medications unless the patient remembers to disclose them
• A specialist seeing a patient for the first time has no access to prior imaging, lab results, or treatment history without a formal records request that takes days or weeks
• An insurer processing a claim cannot verify medical necessity without a paper-based authorization workflow that adds days to every decision

What a Modern Digital Health Infrastructure Looks Like

The architecture of a genuine digital health spine is not complicated in concept — though it requires rigorous thinking about privacy, interoperability, and governance to execute well. At its core, it consists of three layers.

Layer 1: The Citizen Health Identity

Every person has a secure, portable digital health identity — accessible via a mobile app, a physical card with NFC capability, or a QR code. This identity contains a structured summary of the person's health record: vaccination history, active medications, known allergies, chronic conditions, and relevant clinical history. It is owned by the citizen, not by any single institution. The citizen controls what each party can see.

Layer 2: Role-Based Access Portals

Different stakeholders in the health ecosystem need different views of the same underlying data. A school administrator verifying vaccination compliance does not need to see medication history. A pharmacist filling a prescription needs active medications and allergies but not imaging results. An emergency first responder needs the critical safety information instantly, without requiring the patient to be conscious or present identification.

The architecture serves each stakeholder with a purpose-built portal that surfaces exactly what they are authorized to see — and nothing more. This is not just good design. It is a regulatory requirement in any serious privacy framework.

Layer 3: The AI Intelligence Layer

The raw digital health record becomes exponentially more valuable when an AI layer can reason over it — not just store and retrieve it. This layer enables population health surveillance that identifies vaccination coverage gaps before they become outbreaks. It flags drug interaction risks in real time at the point of dispensing. It surfaces care gaps for patients who have fallen out of routine screening schedules. It enables proactive intervention rather than reactive response.

The Implementation Reality

The technology to build this system exists today. The barriers are not technical — they are organizational, political, and governance-related. Interoperability standards, data sharing agreements between institutions, privacy impact assessments, and citizen trust all require sustained attention that is harder to fund and govern than the technology itself.

The organizations that succeed in building digital health infrastructure share a common trait: they treat governance as a first-class design problem, not an afterthought. They define the data sharing rules, the access controls, the audit mechanisms, and the citizen consent framework before they write the first line of code.

At Krellis, this is the work we do. We bring the domain expertise to navigate the organizational complexity, the AI architecture to build the intelligence layer, and the practitioner experience to design systems that clinicians, administrators, and citizens will actually use. The digital health spine is not a distant aspiration — it is an achievable design challenge. The question is who will build it, and how well.

When loss ratios deteriorate, the instinctive response in most carriers is to reach for pricing. Rates go up. Underwriting guidelines tighten. The problem appears to improve — until it doesn't. The next adverse development cycle begins, and the same conversation repeats.

This pattern is not a failure of pricing sophistication. It is a failure of data infrastructure. And until the industry addresses the root cause, pricing will continue to be a lagging corrective mechanism for problems that should have been identified and resolved much earlier in the value chain.

Where the Loss Ratio Is Actually Made

Consider where in the insurance value chain the loss ratio is fundamentally determined. It starts at submission — with the quality of risk information provided, the accuracy of the underwriting decision, and the appropriateness of the terms and conditions applied. A loss ratio problem that appears in year three of a commercial lines book was likely embedded in the underwriting decisions made in year one.

By the time a loss event occurs, the economics are largely locked in. The question is no longer whether the carrier will pay — it is how much, how quickly, and with how much leakage. This is where the second major loss ratio lever exists: claims handling quality, fraud detection, and triage accuracy.

The Claims Data That Never Gets Mined

Most carriers have years of claims data that is structurally underutilized. The data exists in claims management systems, adjuster notes, medical records, and third-party databases — but it is rarely integrated in a way that enables real-time pattern recognition across the portfolio.

The consequences are predictable. Fraud rings exploit the gaps between claims systems that don't talk to each other. Legitimate claims are delayed because the triage process cannot distinguish them from suspicious ones. Medical record summarization that should take minutes takes days because it is done manually. Subrogation opportunities are identified late or missed entirely because no one is systematically scanning for them.

What AI Changes in the Claims Function

AI doesn't replace claims adjusters. It makes them dramatically more effective by handling the cognitive load that currently consumes most of their time — reading medical records, cross-referencing claim details against historical fraud patterns, routing claims to the appropriate handler, flagging anomalies for human review.

A well-designed claims AI system operates in real time at the point of first notice of loss. It ingests the claim, cross-references it against the carrier's historical claims database, external fraud network data, and third-party data sources, and produces a triage recommendation within seconds. The adjuster receives a claim that has already been scored for fraud risk, categorized by complexity, and enriched with relevant context from prior interactions with the same claimant or involved parties.

The Underwriting Data Problem

On the underwriting side, the data problem is different but equally consequential. Most commercial lines underwriters make decisions based on the information provided in the submission — which is a curated representation of the risk, provided by a party with an incentive to present it favorably.

Alternative data sources — satellite imagery, telematics, building sensor data, social and behavioral signals — can provide an independent view of the risk that supplements and challenges the submission data. Carriers that have deployed AI models trained on alternative data sources are making demonstrably better underwriting decisions, with lower loss emergence in the years following the policy inception.

Where to Start

The path from here to a materially improved loss ratio is not a single large transformation program. It is a sequence of targeted, high-ROI use cases that compound over time. In our experience working across P&C carriers, the highest-leverage starting points are consistently:

• Claims fraud detection — the ROI is immediate and measurable, and the data required is largely already available
• Intelligent FNOL triage — reduces handling time and leakage in the first 48 hours of a claim
• Medical record summarization — eliminates a massive manual bottleneck in complex claims
• Underwriting data enrichment — alternative data integrated at point of submission decision

None of these require a multi-year transformation program. Each can be proven in weeks against real data, before any commitment to full deployment. That is the model we use at Krellis — and it is why our clients see measurable loss ratio improvement within a single policy year.

When designing clinical AI platforms, the instinct is to focus on the model — the accuracy of the food recognition, the quality of the metabolic prediction, the sophistication of the clinical decision support. These matter. But they are almost never the reason a digital health platform succeeds or fails with patients.

The reason most digital health products fail to achieve sustained engagement is far simpler and far more tractable: the onboarding experience tells the wrong story to the wrong patient.

The Adaptive Onboarding Insight

When we designed a physician-led metabolic health platform, the single most impactful design decision we made was to treat onboarding as a diagnostic and personalization exercise rather than a registration form.

We built an adaptive onboarding flow around three questions that are simple for the patient but carry significant diagnostic weight for the platform:

• What is your primary health goal? — Weight optimization, cardiovascular health, longevity, energy and performance, or condition management
• How comfortable are you with health tracking technology? — A range from "I prefer simple check-ins" to "I want every metric tracked"
• What does success look like for you in six months? — An open-ended response that the AI processes to calibrate the tone and depth of subsequent interactions

These three data points generate a completely different platform experience for different patient profiles. A 58-year-old with low tech comfort and a diabetes management goal sees a radically simpler, more guided interface than a 35-year-old performance athlete who wants granular biometric tracking. Same underlying platform. Completely different presentation.

The Physician-Patient Bridge

The second major engagement driver is the degree to which the platform genuinely connects the patient's experience to their clinical relationship. Most consumer health apps operate in a vacuum — they collect data, generate insights, and push notifications, but they do not close the loop with the physician who is actually responsible for the patient's care.

Platforms designed with the physician at the center of the experience — where AI-generated insights from patient-logged data surface directly in the physician's workflow — create a fundamentally different engagement dynamic. The patient is logging food, activity, and biometric data not just for their own benefit but because they know their physician will see it, contextualize it, and act on it at their next appointment.

What This Requires Architecturally

Building the physician-patient bridge requires solving several hard problems simultaneously. The data the patient generates must be structured in a way that is clinically meaningful — not just raw numbers but contextualized trends that a physician can act on in a three-minute appointment window. The physician-facing interface must be integrated into the clinical workflow, not a separate application that competes for attention. And the privacy and consent architecture must be airtight, because the stakes of getting it wrong are high in any regulated health environment.

AI Food Scanning: The Engagement Trigger

One of the most effective engagement triggers we have found in metabolic health platforms is AI-powered food scanning. The ability to point a phone at a meal and receive an immediate, accurate nutritional breakdown removes the single biggest barrier to food logging: the effort of manual entry.

But the engagement value of food scanning goes beyond convenience. It creates a moment of genuine delight — the experience of technology doing something that feels magical — at a point in the patient journey when that positive emotional association with the platform is most valuable. Patients who experience a high-quality food scan in the first week of using a health platform have dramatically higher retention rates than those who don't.

The Design Principles That Transfer

What we learned designing physician-led health platforms transfers to any AI product where sustained engagement is the success metric:

• Personalize from the first interaction — adaptive onboarding is not a nice-to-have
• Close the loop between the AI insight and the human relationship that matters to the user
• Find the engagement trigger that creates genuine delight in the first week
• Design for the least tech-comfortable user in your target cohort, not the most
• Measure retention at 7, 30, and 90 days — not just activation rate

The platforms that win in digital health will not be the ones with the most sophisticated models. They will be the ones that understand human behavior well enough to design AI that people actually want to use.

The concept of a digital health wallet — a secure, portable, citizen-controlled record of a person's health identity — is not new. Health information exchanges, personal health records, and patient portals have existed in various forms for two decades. None of them have achieved meaningful adoption at population scale.

The reason is not technological. It is motivational. The existing solutions were designed to serve the convenience of healthcare administrators, not the daily life needs of citizens. A digital health wallet that people actually use needs to solve real problems they encounter in real situations — not just make it slightly easier to transfer records between providers.

Here are five use cases that illustrate exactly what a well-designed digital health wallet makes possible — and why each one represents a genuine, daily-life motivation for citizens to adopt and maintain it.

Use Case 1: School Enrollment Without the Filing Cabinet

Every year, parents enrolling children in school are asked to provide proof of vaccination. In most jurisdictions, this means locating a paper immunization record, possibly having it verified by a physician, and submitting it to the school. The school then maintains a paper or manual digital record that no other agency can access.

With a digital health wallet, this interaction takes seconds. The parent opens the app, selects their child's profile, and shares a verified vaccination record directly with the school's compliance portal. The school sees exactly what it needs — vaccination status against required antigens — and nothing more. The record is cryptographically verified, so it cannot be forged. The parent never needs to find a piece of paper again.

Use Case 2: The Emergency Department Without Words

A patient arrives at an emergency department unconscious after a road accident. No identification. No family present. No ability to communicate. Under the current system, the clinical team works from zero — administering treatments that may interact badly with medications they don't know the patient is taking, unaware of allergies that could be life-threatening.

With an NFC-enabled health wallet, a first responder taps the patient's phone or health card within 2.3 seconds of arrival. The critical health summary appears on the responder's device: blood type, active medications with dosages, known allergies flagged in red, chronic conditions, treating physician contact. No login required. No consent workflow in a moment that requires speed. The data is structured to surface exactly what matters in an emergency.

Use Case 3: The Pharmacist Who Can Actually Check

A patient brings a new prescription to a pharmacy. The pharmacist is expected to check for drug interactions — but in most cases, they can only check against the medications dispensed at that specific pharmacy. If the patient fills prescriptions at multiple locations, or receives medications during a hospital stay, the pharmacist has no visibility.

With access to the patient's medication list via the health wallet, the pharmacist sees the complete picture. The AI layer surfaces any interactions between the new prescription and the patient's existing medications in real time, before dispensing. The pharmacist can have an informed conversation with the patient and, if necessary, contact the prescribing physician with a specific, evidence-based concern — not a generic flag.

Use Case 4: The New Employer Occupational Health Screening

Many employers — particularly in healthcare, food service, and regulated industries — require occupational health screening as a condition of employment. This typically involves a paper-based process of collecting vaccination records, health declarations, and fit-for-work certifications, coordinated through an occupational health provider.

The health wallet enables the candidate to share a verified occupational health summary directly with the employer's HR system. The employer sees exactly what they are authorized to see under the employment contract. The candidate controls the disclosure. The occupational health provider's certification is cryptographically attached to the record, so it cannot be altered after the fact.

Use Case 5: The Chronic Disease Patient Who Doesn't Start From Zero

A patient with a complex chronic condition — Type 2 diabetes managed with multiple medications, a history of cardiac events, and a treating team that includes an endocrinologist, a cardiologist, and a primary care physician — sees a new specialist for the first time. Under the current system, the specialist begins with whatever records were successfully transferred in advance, supplemented by what the patient can recall.

With a complete longitudinal health record accessible via the health wallet, the specialist has immediate access to the full clinical picture: lab trends over time, medication history including discontinued treatments and the reasons for discontinuation, prior imaging with AI-generated summaries, and a structured problem list. The appointment can focus on the clinical question — not on reconstructing history that already exists somewhere in the health system, inaccessible.

The Design Principles That Make It Real

All five use cases share a set of design requirements that any serious health wallet implementation must satisfy:

• Citizen control — the person decides what each party can see, with granular consent
• Verified credentials — records are cryptographically signed by the issuing institution
• Offline capability — critical information must be accessible when connectivity is unavailable
• Role-based access — each stakeholder sees only what they are authorized to see
• Interoperability — the system must work across institutions, jurisdictions, and technology platforms
• Privacy by design — the architecture must protect against misuse, not just prohibit it

These are hard design problems. But they are solved problems — the cryptographic standards, the identity frameworks, the consent management architectures all exist. What has been missing is the organizational will and the practitioner-grade implementation capability to bring them together into a system that citizens actually trust and use. That is the work we do.

Commercial underwriting has a timeline problem. The average time from submission to binding decision in commercial lines ranges from three to six weeks for mid-market risks. For complex specialty risks, it can be months. In a market where brokers have options and clients have expectations shaped by their experience in every other industry, this is not a sustainable competitive position.

The instinctive response from many carriers has been to add headcount. More underwriters, more support staff, more analysts. The result is a larger team still taking three to six weeks, because the bottleneck is not the number of people — it is the structure of the process itself.

Where the Time Actually Goes

A detailed analysis of commercial underwriting workflows consistently reveals the same pattern. Of the total elapsed time from submission to decision:

• 30–40% is waiting — for information from the broker, for data from third parties, for review queues to clear
• 25–35% is administrative — data entry, document processing, templated correspondence, workflow management
• 15–20% is preliminary analysis — risk scoring, exposure identification, referral decisions
• Only 15–25% is actual underwriting judgment — the work that requires experience, expertise, and authority

AI addresses the first three categories decisively. The last category — the judgment — is where the underwriter's value lies, and it is where they should be spending their time.

What AI-Assisted Underwriting Actually Looks Like

A well-designed AI-assisted underwriting system intercepts the submission at first receipt and immediately begins work that would otherwise wait for an underwriter's attention. It extracts and structures the submission data, cross-references it against the carrier's appetite and guidelines, pulls relevant third-party data, and generates a preliminary risk assessment with supporting rationale.

By the time the underwriter opens the submission, they are not starting from a blank page. They have a structured risk summary, a preliminary score, flagged concerns requiring their attention, and suggested terms — all generated in minutes, not days.

For standard risks that fall clearly within appetite and guidelines, the system can generate a referral recommendation and draft terms without the underwriter needing to touch the file. The underwriter reviews and approves. For complex or non-standard risks, the AI has done the preparation work that allows the underwriter to focus their expertise on the genuinely difficult questions.

The Alternative Data Advantage

Beyond process automation, AI creates a genuine underwriting quality advantage through access to alternative data. Satellite imagery reveals physical risk characteristics not captured in the submission. Telematics data provides an objective view of commercial auto risk that is more predictive than the self-reported data in the application. Building sensor data from IoT devices can provide near-real-time risk monitoring for property risks.

Carriers that have integrated alternative data into their underwriting models are making better risk selections — with lower loss emergence in subsequent years — than those relying solely on submission data. This is not a marginal improvement. It is a structural competitive advantage that compounds over time as the models improve with more data.

Starting the Transformation

The path to AI-assisted underwriting does not require a multi-year platform replacement program. The highest-leverage entry points are submission intake automation and preliminary risk scoring — both of which can be implemented against existing systems without displacing the underwriting platform. The ROI is immediate and measurable, the implementation risk is low, and the capability built becomes the foundation for more sophisticated applications over time.

The underwriters who will define the industry in five years are the ones who today are learning to use AI as a force multiplier for their judgment — not the ones who are worried about being replaced by it.

The term "agentic AI" has migrated from AI research papers into vendor pitch decks with impressive speed and corresponding loss of precision. If you have sat through a technology presentation in the past twelve months, you have almost certainly been told that agentic AI will transform your operations — without a clear explanation of what that means in practice for an underwriting team, a claims department, or a clinical operations function.

This piece is an attempt to close that gap. Not with technical definitions, but with concrete descriptions of what agentic AI does differently from the AI systems most organizations already have, and why the distinction matters for regulated industries in particular.

What Makes an AI System "Agentic"

A conventional AI system responds to a query. You provide an input — a document, a question, a dataset — and the system produces an output. The output informs a human decision. The human then acts.

An agentic AI system can plan, sequence actions, use tools, check its own work, and iterate toward a goal without requiring human intervention at each step. It can be given a task — "process this claims batch and flag anything requiring human review" — and complete it autonomously, making decisions along the way about how to proceed, what information to gather, and when a situation requires escalation.

Three capabilities define an agentic system:

• Planning — the ability to decompose a goal into a sequence of steps and execute them in order
• Tool use — the ability to call external systems, databases, APIs, and services to gather information and take actions
• Self-critique — the ability to evaluate its own outputs, identify errors or gaps, and correct them before presenting a result

What This Looks Like in a Claims Department

Consider the workflow for processing a complex personal injury claim. Under a conventional AI-augmented workflow, the adjuster receives AI-generated summaries of medical records, a fraud risk score, and suggested next actions. The adjuster reviews these outputs and decides what to do. They then execute each step: requesting additional documentation, contacting the claimant, consulting a specialist, updating the claims system.

Under an agentic AI workflow, the system receives the new claim and autonomously executes the initial processing steps: extracting and summarizing medical records, cross-referencing against fraud databases, identifying any coverage questions that need legal review, and drafting the initial contact letter to the claimant. It flags the claim for adjuster attention with a structured summary of what it has done, what it has found, and what decisions remain for the human to make.

What This Looks Like in a Clinical Setting

In a digital health platform, an agentic AI layer can monitor patient data streams continuously and take structured actions when predefined thresholds are crossed. A patient's continuous glucose monitor shows a concerning pattern over three days. The agentic system detects the pattern, reviews the patient's recent food logs and activity data for context, drafts a message to the patient with specific, personalized guidance, and flags the case for the treating physician with a structured summary. All of this happens without a human initiating the workflow.

The physician sees a prepared summary with actionable context, not a raw alert that requires them to pull and interpret data before they can respond. The patient receives timely, relevant guidance. The system has done the cognitive work that would otherwise have required a clinical coordinator to monitor, notice, and act.

The Governance Question

The capability that makes agentic AI powerful — autonomous action — is also what makes governance essential. In regulated industries, the design of the boundaries within which an agentic system operates is as important as the design of the system itself.

Which actions can the system take without human approval? Which require review before execution? Which must always be human-initiated? How is the system's decision-making audited? What happens when it encounters a situation outside its training distribution?

These are not hypothetical concerns. They are design requirements that must be addressed before any agentic AI system goes into production in an insurance or healthcare environment. The organizations that get this right will build durable competitive advantage. Those that skip the governance design in the rush to deploy will create operational and regulatory risk that outweighs the efficiency gains.

The good news is that this is a solvable design problem — not a reason to avoid agentic AI. It requires practitioners who understand both the technology and the regulatory environment well enough to draw the right boundaries. That combination of expertise is rare, and it is exactly what the industry needs right now.

Every organization pays a hidden tax on growth. It is not on the balance sheet. It rarely appears in board presentations. But it compounds annually, consuming resources that could be invested in capability building, and it grows more expensive the longer it is left unaddressed.

The hidden tax is the cost of organizational complexity — the legacy systems, manual processes, and structural inefficiencies that accumulate over years of growth and underinvestment in the operational foundation. In insurance and healthcare, where the regulatory environment, the risk profile, and the data volume are all unusually demanding, this tax is particularly high.

The Three Lenses We Use to Find the Tax

At Krellis, we diagnose organizational hidden tax through three analytical lenses before recommending any AI investment. Each lens reveals a different dimension of the problem.

The Technology Lens: Is Your Stack an Accelerator or an Anchor?

Legacy technology creates AI drag in ways that are not always obvious. The most visible form is data silos — claims data that cannot be joined to underwriting data, patient data that cannot be connected across care settings. But the subtler forms are often more damaging: data quality issues that require extensive cleaning before any model can be trained on the data; API architectures that cannot support the real-time data flows that agentic AI requires; vendor lock-in that prevents integration with best-of-breed AI tooling.

A technology audit in preparation for AI investment is not about replacing systems. It is about understanding which parts of the existing stack are AI-ready and which require remediation before AI can deliver value. This distinction determines the sequencing of the AI roadmap.

The Operational Lens: Where Does Work Turn Into Meetings?

Manual processes are the most visible form of operational hidden tax, but they are often symptoms of a deeper problem: the absence of clear data flows and decision rights. When a claims team spends 40% of its time in status meetings, the root cause is usually that no one has a real-time view of where each claim is in the process — which is a data architecture problem, not a meeting culture problem.

AI cannot automate a process that is not defined. Before implementing any AI-driven automation, the current process must be mapped with enough precision to identify what decisions are being made, by whom, on the basis of what information, and with what authority. This mapping is often the most valuable output of a diagnostic engagement — not because it reveals where to apply AI, but because it reveals what the AI will actually be doing.

The Culture Lens: Why Do Decisions Take Three Weeks Instead of Three Hours?

The third dimension of hidden tax is the hardest to quantify and the most important to address. Organizational culture — specifically, the structures and incentives that govern how decisions get made — determines whether AI investments deliver their promised value or get absorbed by the organization without changing anything.

The most common cultural failure mode is what we call fear-based bureaucracy: decision-making processes that optimize for avoiding blame rather than achieving outcomes. In this environment, AI outputs become one more input to a review process that ultimately requires sign-off from multiple layers of management — and the cycle time savings from the AI are consumed by the approval workflow.

Changing this requires leadership commitment to redesigning decision rights alongside the AI implementation — not just deploying the tool and expecting the behavior to follow.

The Sequencing Principle

The diagnostic work across all three lenses produces a prioritized view of where the hidden tax is highest and where AI investment will deliver the fastest, most durable return. The sequencing principle we apply is consistent: prove the value on a use case where the data is cleanest, the process is best-defined, and the organizational readiness is highest. Use that proof to build the credibility and the infrastructure for the next, more complex use case.

This approach is less exciting than a bold transformation vision. But it is dramatically more likely to produce measurable results in a twelve-month window — and measurable results are what create the organizational appetite for the larger transformation that follows.

The carriers and health systems that will lead their industries in five years are not the ones with the most ambitious AI strategies today. They are the ones that are doing the unglamorous work of understanding their hidden tax, addressing it systematically, and building the operational foundation that makes AI an accelerator rather than another layer of complexity on top of an already complex system.