AI, healthcare costs & the future of work

Chapter 1: The change already underway

I believe AI is going to change both sides of the health-insurance equation: the cost of delivering care and the work that helps people pay for it.

A business may be able to serve more customers without building the administrative workforce that earlier generations of companies needed. Healthcare could become less labor-intensive at the same time that fewer people obtain dependable coverage through a traditional job. For employers, insurance leaders, and policymakers, those are parts of the same problem.

My long-range argument begins with a decline in scarcity: diagnosis, treatment, prevention, and eventually the treatment of aging itself becoming more accessible. The question is how to turn that possibility into a system that works for people.

Healthcare is still expensive today. The Centers for Medicare & Medicaid Services projects national health spending to rise from $5.3 trillion in 2024 to $9.0 trillion in 2034. Employers are feeling that pressure now: KFF's 2026 Employer Health Benefits Survey puts the average annual premium for employer-sponsored family coverage at $28,303, up 5% from 2025, with workers contributing $7,471 on average. The survey covers employers with at least 10 workers. These are present-day costs, not a forecast of what AI might eventually save.

But a rising near-term cost curve does not tell us where technology can ultimately take the system.

My long-range view is that artificial intelligence, robotics, earlier detection, better prevention, and healthspan research will make many forms of care faster, more precise, and less expensive to deliver. If those gains reach patients and payers, healthcare could begin to look less like an unpredictable financial emergency and more like abundant infrastructure.

That would change employee benefits, the workforce, and the role employers play in providing access to care. It could also make a universal base layer of healthcare more economically practical—while leaving room for employers to compete through richer benefits, convenience, choice, and human support.

This transition will not happen all at once. It is not guaranteed. But the early signals are already visible.

AI will not merely change the tools people use. It will reduce the number of people required to operate healthcare, insurance, and administrative systems—and that makes job-independent healthcare essential infrastructure.

—Andrew Decker

The starting point is the work already moving into software. From there, the argument reaches the structure of a business, the future of employment, and the way we finance care.

Chapter 2: Insurance is already adopting AI

A health insurer brings together enormous volumes of documents, transactions, rules, questions, and follow-up. Those are precisely the areas where connected AI systems can absorb work. The strongest evidence is now coming from specific deployments, rather than general promises about transformation.

Cigna and Sierra: voice AI is already handling calls

In a November 3 discussion published by The Cigna Group, Sierra co-founder and CEO Bret Taylor said Accredo's AI agent was autonomously handling more than 40% of calls. That is a reported result for the specialty-pharmacy deployment—not a claim that AI handles 40% of every Cigna business or that Cigna has eliminated 40% of its staff.

The larger idea is continuity: an agent can carry relevant context across conversations instead of making a person start over each time. Cigna describes combining that capability with clinical expertise and human oversight. To me, the significance is that routine service can increasingly be delivered by software while people concentrate on exceptions, complex care, and responsibility for the outcome.

UnitedHealthcare: self-service with a human handoff

UnitedHealthcare introduced Avery on March 26, describing a generative AI companion for eligible members that helps with benefits, claims information, cost estimates, and care navigation. The company reported that members using Avery had not needed an advocate's assistance 90% of the time, while providing a transfer when live help was needed or requested.

That is UnitedHealthcare's reported experience with a particular tool and user population. It is not evidence that 90% of its customer-service jobs have disappeared. It does show why the staffing required for routine questions may change as self-service becomes more capable.

Claims, correspondence, and fraud detection

CareFirst has described AI improvements in inbound correspondence, including a process shortened from six days to a few hours. Aetna's May 26 announcement says its second-generation Claims Assist Manager reduces processing time by more than 20% for complex claims requiring manual review. These are company reports about defined workflows, not independently established reductions in total medical spending or staffing.

The pattern extends beyond health insurance. On November 5, Zurich UK reported that connected AI and advanced analytics helped identify 860 fraudulent claims worth £17.9 million. That was a portion of its £109.3 million total detected fraud, not the entire total. Zurich also described criminals using generative AI to fabricate evidence. Better automation can create new investigative work even as it removes routine processing.

Who receives the productivity dividend?

Profit is part of the incentive. Lower administrative costs can improve an insurer's margins, and the same tools may improve service, shorten delays, and help people find appropriate care. Those possibilities can coexist. A technology announcement alone does not tell an employer which benefit will materialize, or who will receive it.

My expectation is that successful automation will reduce the human labor required for routine insurance administration. That does not support a claim that Cigna, UnitedHealthcare, or any other carrier is replacing all of its staff. Clinical judgment, appeals, investigations, security, governance, and accountable leadership remain consequential work. Employers should ask for evidence of faster service, fewer errors, fair decisions, and lower costs—not accept an AI label as proof of value.

These deployments are the first part of the story. The next question is what happens when a company designs its operations around agents from the start.

Chapter 3: A company built around AI

Organizational design matters. A traditional business can add AI to existing departments, job descriptions, and approval chains. A business built around agents can instead design connected workflows first, then assign people to the responsibilities that require human judgment, relationships, and accountability.

In his May 20, 2026 essay, DFINITY's Pierre Samaties argues for a different software foundation: a live World Model connecting the organization's customers, projects, budgets, risks, and decisions. Agents could use that shared context to coordinate work and identify problems earlier. His account still includes decisions being brought to people; it does not establish that whole businesses already run without humans.

That argument changed how I approached my own business. After reading Pierre's essay, I scrapped the app I had been building and started over around this operating model. I am glad I did. I wanted the software to support connected work from the beginning, with deliberate decisions about what an agent can do and what remains a person's responsibility.

My expectation is that this model will let some businesses grow without creating the large administrative teams that earlier companies needed. The pressure could arrive as jobs never created, vacancies left unfilled, or smaller teams at competitors. A company could still add employees while needing fewer human labor hours for each customer served. That forecast depends on reliable execution and the costs of supervision; adopting AI alone does not establish that an employer will cut jobs.

What the strongest AI models can do now

As of November 12, 2026, Anthropic recommends Claude Opus 5.5 for most coding and knowledge work, with the newly released Fable 5.2 and Mythos 5.2 aimed at more demanding, open-ended technical problems. Fable and Mythos share an underlying model; Mythos access requires approval through designated verification programs. They should not be described as equally available to every business.

Anthropic reports improvements in coding, reasoning, vision, and tool-driven work. Its examples include a production software migration and an experimental, multi-day project in which agents built a working prototype operating system. Fable can also advise or coordinate other models. Those are company-reported demonstrations, not a guarantee that an unattended agent can safely run any business.

The business implication is substantial: AI can help build and change the software that performs work, as well as perform tasks within it. But a benchmark score is not a percentage of employees who can be dismissed. Real deployment still depends on accurate data, permissions, testing, supervision, and the cost of correcting mistakes.

Putting that outlook into practice at SSGI

This is the operating model I am building toward at SSGI. If I believe AI will change the human labor required to administer healthcare and benefits, I need to apply that thinking to our own business. The capability for fully autonomous execution of defined operational workflows is already in place. I can enable it today; deciding where to use it is a leadership responsibility.

SSGI One is designed to support businesses as they grow—from a team of a few employees to a national organization with 25,000 employees and beyond. It connects benefits administration and employee service in one portal, helping teams coordinate requests, documents and follow-through across locations and divisions. Each enterprise deployment requires configuration, integration review and capacity validation.

SSGI One puts agency staff, employers and employees into a connected benefits administration environment. Group benefits and ICHRA enrollment share one login, with enrollment windows, new-hire and life-event requests, waivers and allowance workflows. Employees and employer administrators can send requests and documents and see status and replies in the portal. That gives us a practical place to apply an agentic operating model.

Our chosen operating mode keeps human approval in consequential parts of that work. Inside SSGI One, replies to customer requests and advice currently require human approval. Fixed-template messages may be sent automatically. The available autonomous capability is broader than the authority we choose to delegate.

We already use an AI agent to answer our phone calls at SSGI Benefits. Our public Ask SSGI assistant also responds automatically, without individual review of every response. These systems handle general information and reception. Answering a general question and authorizing a consequential insurance action are different responsibilities.

Our commitment for carrier API write transactions is explicit: staff approval comes first. Technical capability does not cancel the approvals, professional responsibilities or agreements that govern an action. A business still has to stand behind the work it authorizes.

Clients can also bring their preferred AI tools for an optional integration review. We begin with the workflow and existing systems, then agree on compatibility, permissions, data handling, human review, provider terms and costs before implementation. The useful question is what the integration should accomplish and what authority it needs.

The Internet Computer, Open Cloud, and the software underneath

Our public website and SSGI One run on an SSGI-controlled Open Cloud engine with U.S.-based replica nodes. Open Cloud uses the Internet Computer Protocol to provide an application environment assembled from participating compute nodes. A cloud engine gives its operator choices over the nodes, providers, locations, and replication used for that environment.

The application unit is a canister: software bundled with persistent state—the information it remembers between interactions. The network replicates application updates, with consensus supporting agreement on the result. Canisters can also serve web content and connect to external services. Ordinary fast reads and consensus-backed updates are different operations; verifiable reads require the appropriate certification mechanism. Those distinctions matter more than simply attaching the word blockchain to a website.

For an agentic business, persistent state is useful because work continues after a conversation ends. An application needs to know which request exists, what information is missing, what has been authorized, and what happened next. My aim is to make that business context dependable enough for people and authorized agents to work from it, rather than reconstructing it from scattered messages.

The model and the application infrastructure perform different jobs. A large language model can interpret a request, propose a plan, or use an authorized tool. The application must enforce the permissions and record the resulting work. Blockchain does not make an AI answer correct, and replicated software is not a substitute for privacy controls, testing, or human responsibility. Model inference, telephone services, and other integrations should not be assumed to run entirely on the blockchain.

Open Cloud also describes an MCP connector for AI assistants, with per-app, time-limited permissions for reading information or taking actions. That is relevant to the operating model: an agent needs defined authority to interact with software. It is a platform capability, not a claim that every available integration is already enabled for every SSGI client.

Web3 principles without making clients manage crypto

This is the part of the Web3 ethos I value: greater control over software, open interfaces, verifiable execution, and less dependence on a single provider's closed application stack. Those are design goals, not an automatic guarantee of privacy, decentralization, or immunity from failure. The underlying protocol has its own resource and governance mechanics; clients should not need to master them to obtain benefits support.

Visitors use our website in an ordinary browser without buying cryptocurrency or connecting a wallet. The infrastructure still has operating costs, including the cycles used to pay for computation. Removing crypto friction from the visitor experience does not mean the technology beneath it stops using blockchain. It means the service should make that complexity someone else's job.

This work brings the wider argument into focus. AI can reduce the labor needed for administration while increasing the importance of sound judgment, clear responsibility and benefits that remain dependable through workforce change. I explore the business model behind those choices in a separate Insight.

Read Building an Agentic Business: Automation by Design, Accountability in Practice

Organizational design is where the technology argument becomes a workforce argument. If growth requires fewer administrative hours, the first employment effect may be a position that is never opened.

Chapter 4: The jobs we may never create

Most jobs will not survive in their current form

Here is the workforce conclusion I believe we need to say plainly: most jobs as they are currently designed will eventually be replaced, compressed, or fundamentally rebuilt by AI and robotics. In labor-heavy administrative fields, I expect the result to be fewer human jobs and lower employment.

That is not the same as saying most people will be permanently unemployed. A job is a bundle of tasks, not an indivisible object. The title may remain while the work underneath it changes completely. A smaller team may produce what once required an entire department. New occupations will appear, and people will continue to create needs, services, and forms of work that we cannot fully anticipate. But my thesis is not merely that everyone will use a new tool while employment remains unchanged. It is that organizations will need substantially fewer human labor hours to produce the same—or greater—output.

I believe it would be a mistake to underestimate the scale of the disruption. The evidence today does not prove that a majority of occupations will vanish. It does show that exposure is already broad. The International Labour Organization estimates that one in four jobs worldwide has some exposure to generative AI and says transformation is currently more likely than complete replacement because most occupations still include tasks requiring human input. The International Monetary Fund estimates that almost 40% of employment worldwide—and about 60% in advanced economies—is exposed to AI.

Those estimates largely describe the opening phase: intelligence delivered through software. Robotics gives that intelligence a body. As perception, dexterity, mobility, and autonomous decision-making improve, automation reaches warehouses, transportation, construction, food service, agriculture, healthcare, maintenance, and other work that once appeared protected because it happened in the physical world.

My forecast goes further than today's institutional estimates. Once an AI system can perform enough of a job at an acceptable quality, it can be copied at extremely low incremental cost, operate continuously, and improve across every deployment at once. Employers do not need to wait for a machine that can do everything a person can do. They only need a system that performs enough of the valuable tasks faster, more consistently, more safely, or less expensively than the existing process.

Why replacement can move faster than people expect

Previous automation often required expensive, specialized machinery and a carefully controlled environment. AI is different because software can spread globally in days. The same underlying capability can read documents, produce analysis, write code, respond to customers, review claims, schedule work, monitor systems, and guide a robot.

The economic pressure compounds. When one organization redesigns a workflow and materially lowers its costs, competitors cannot indefinitely preserve the old staffing model simply because it is familiar. They adopt the technology, reorganize, or lose ground. What begins as an optional productivity tool becomes a requirement for remaining competitive.

This is why replacement will often look less like a dramatic announcement and more like steady compression:

• A ten-person team becomes a five-person team using AI.
• A call center retains people for difficult exceptions while software handles routine volume.
• One technician supervises a fleet of machines that once required many operators.

The BLS 2025–2035 projections provide an important current baseline: 5.9 million additional jobs overall, or 3.5% growth, including more than 2.2 million additional jobs in private healthcare and social assistance. Those projections anticipate net employment growth, not mass unemployment.

BLS also introduced occupational AI-exposure categories in August 2026. Exposure describes how an occupation's tasks overlap with AI capabilities and observed uses; it does not measure job losses, replacement probability, or productivity gains. My longer-term forecast of administrative contraction goes beyond those findings. Employers should plan for disruption without treating an exposure score as a staffing forecast.

What the hiring evidence shows

The evidence is uneven. Stanford's August 2026 research revision found that weaker employment among young workers in exposed occupations mainly reflected reduced hiring. The authors describe observational patterns, not causal estimates. A March 2026 Federal Reserve research note using data through November 2025 found no negative association between AI adoption and total job postings. These studies do not establish that the most technologically advanced businesses are universally cutting the most jobs.

Maryland offers a local signal. The state's October 2026 Business AI Benchmark announcement described a June–July survey of nearly 300 senior decision-makers. Among respondents, 64% planned to accomplish more with existing employees using AI rather than hire additional staff; 4% expected headcount reductions. Those are stated expectations, not measured job losses or a finding about every Maryland employer.

If business growth becomes less dependent on payroll growth, we should also reconsider how people obtain health coverage and financial security through work. Watching layoffs alone can miss a changing entry-level hiring pipeline and the benefits access that would otherwise come with those jobs.

Administrative work at every employer follows the same path

The same pressure extends to HR, payroll, benefits administration, finance, procurement, scheduling, customer service, and reporting. AI can assist with retrieving policies, comparing records, drafting responses, and routing exceptions. My expectation is that employers will redesign these connected workflows around smaller teams, with reductions arriving through attrition, fewer new hires, vendor consolidation, or layoffs.

Deployment will lag capability. Legacy systems, fragmented data, contracts, privacy, cybersecurity, labor agreements, and liability all affect whether automation works safely in practice. High-stakes decisions need testing, auditability, appropriate appeals, and a responsible human owner. Maryland insurers remain accountable for compliance with applicable insurance law when they use AI.

Those constraints affect timing; they do not make existing roles immune to change. Some employers will reinvest savings in growth and new services. The practical question is which tasks disappear, which roles grow, and how employees move into the work that remains.

Five kinds of work closest to near-term replacement

If I were planning staffing today, these are the five areas I would examine first. This is my practical watchlist of work that can be reduced or redesigned—not a measured ranking of five occupations guaranteed to disappear immediately.

1. Routine customer service and telephone reception. Basic questions, message-taking, routing, and repeat status checks are already targets for voice and chat agents. Cigna's Accredo example illustrates the change. People still need a dependable escalation route when a problem falls outside the agent's authority or ability.
2. Data entry and document-processing work. Moving information from forms, emails, and attachments into another system is vulnerable when extraction, validation, and routing can be connected. Exceptions, poor source documents, and identity errors still require attention.
3. Routine insurance processing and correspondence. Intake, eligibility checks, missing-information requests, and standard follow-up can increasingly run through software workflows. That does not give an agent unrestricted authority to deny care, settle a disputed claim, or provide licensed advice.
4. Basic bookkeeping and reconciliation. Matching transactions, preparing recurring reports, and investigating straightforward discrepancies can require fewer manual steps. Financial controls, unusual transactions, approvals, and responsibility for the accounts remain essential.
5. Routine coding, software testing, and junior technical production. Standard changes, test generation, documentation, and repeatable debugging are increasingly within reach of coding agents. The difficult work shifts toward requirements, architecture, verification, security, and deciding whether a change should ship.

Anthropic's March 2026 labor-market research places programmers, customer-service representatives, and data-entry workers among highly exposed occupations. But its observed AI use falls well short of theoretical capability, and it found no systematic unemployment increase in the most exposed occupations, with tentative evidence of slower hiring for younger workers. Exposure is not a layoff count.

My near-term concern is that some employers will stop replacing departing staff or creating entry-level positions before they announce major layoffs. A job can disappear from a hiring plan without ever appearing in a redundancy announcement. How quickly that happens depends on the actual workflow, demand, and the cost of reliable supervision.

New work does not guarantee a painless transition

Technology has always created jobs as well as destroyed them. AI and robotics will create demand for system designers, robot technicians, cybersecurity specialists, data stewards, patient advocates, caregivers, investigators, safety professionals, and people who provide judgment, creativity, trust, and human connection.

The problem is that a new job in one place, at one wage, requiring one set of skills is not an automatic replacement for the job lost somewhere else. A 52-year-old claims processor does not become a robotics engineer overnight. A worker cannot pay a mortgage with the promise that the economy will eventually become more productive. Even if total employment remains strong, millions of people may face broken career ladders, lower bargaining power, repeated retraining, or periods without stable work.

We should also expect entry-level work to change first and fast. Junior employees have traditionally learned by doing research, drafting routine materials, reconciling information, handling basic customer requests, and observing more experienced colleagues. Those are precisely the tasks AI can absorb early. If employers automate the bottom rung without designing a new way for people to develop expertise, they may save money today and create a shortage of experienced judgment tomorrow.

What would change my view?

My prediction is substantial disruption, not a claim that the outcome has already been settled. A serious argument should say what would make it weaker.

If exposed occupations sustain hiring, wages, and credible entry-level pathways while AI use grows, that would weaken my expectation of broad employment contraction. New demand and new kinds of work could absorb people faster than existing tasks disappear. Fewer labor hours per customer do not automatically mean fewer employees across the economy. The BLS, Federal Reserve, and Anthropic findings discussed above are reasons to test the argument, not footnotes to dismiss.

If supervision, integration, security, and rework consistently consume the productivity gains, that would weaken the economic case for an agent-first organization. A convincing demonstration must survive ordinary operating conditions, including mistakes and exceptions.

And if better medicine expands access and utilization faster than it reduces unit costs, total spending could keep rising even while individual services become cheaper. Longer lives also create needs that do not disappear because diagnosis improves.

Those possibilities affect the pace and financing of my argument. They do not erase the value of reliable coverage between jobs. But they do mean that a portable healthcare foundation must work without assuming mass unemployment, free care, or a windfall from AI. I would rather make the case under demanding assumptions than win it with an optimistic spreadsheet.

The workforce argument would matter even if medicine stood still. It becomes more consequential when the same technologies also change how illness is prevented, detected, and treated.

Chapter 5: Beyond administration: medicine and prevention

AI is moving from administration into medicine

For employers, AI is already changing how benefits are explained, administered, and navigated. The larger transformation will come as AI moves deeper into clinical care.

AI will change every doctor's job

A better question than "Will AI replace doctors?" is "How will AI change what every doctor does?" Marc Triola, MD, Senior Associate Dean for Medical Education and director of the Institute for Innovations in Medical Education at NYU Grossman School of Medicine, framed the issue directly in a recent TEDxNYU Langone Health talk: AI Is Changing What It Means to Be a Doctor: Are We Ready?

That distinction matters. AI does not need to eliminate the physician to transform medicine. It can redistribute diagnosis, documentation, image review, continuous monitoring, research, treatment planning, and patient communication between people and machines. A doctor may spend less time retrieving information, producing routine notes, and screening ordinary cases—and more time validating machine recommendations, managing exceptions, explaining difficult choices, integrating a patient's values, and accepting responsibility for the outcome.

The medical degree will not insulate a job from automation. In fact, some highly paid cognitive tasks may be easier to automate than lower-paid physical work performed in an unpredictable environment because the medical information is already digital, structured, and measurable. Doctors who use AI effectively may first displace doctors who do not. Over time, health systems may need fewer professionals for routine work while expanding teams around complex care, prevention, oversight, patient trust, and cases that do not fit the model.

That evolution could reduce the labor required to deliver a unit of care. It could also expand access by allowing each clinician to serve more people. The ultimate effect on employment will vary by specialty and demand, but the direction is clear: every doctor's workflow, training, judgment, and economic role will change. Medical education will have to teach clinicians not only what to know, but how to interrogate an AI recommendation, recognize model failure, communicate uncertainty, protect patient data, and remain accountable when technology participates in a decision.

For employers and health plans, the promise is not "AI" as a marketing label. It is better outcomes delivered with less delay, duplication, and avoidable labor. Purchasers should insist that clinical productivity gains reach patients and payers—and that a named, qualified human remains responsible when the stakes are high.

The U.S. Food and Drug Administration now maintains an expanding list of authorized AI-enabled medical devices spanning radiology, cardiovascular care, neurology, surgery, and other specialties. These are not science-fiction demonstrations. They are regulated tools entering real clinical workflows.

The next wave reaches further. The Advanced Research Projects Agency for Health is funding programs designed to use AI to speed drug discovery, improve the safety and design of therapies, strengthen biomedical data infrastructure, and shorten the path from research to patient benefit.

AI can help a clinician find a pattern in an image, surface a risk before symptoms become severe, match a patient to a treatment, automate documentation, and reduce the time required to develop a therapy. Each gain may look incremental. Together, they can change the unit economics of care.

There is already measured evidence for some of those gains. A randomized trial published in NEJM AI involving 66 practitioners found that ambient AI reduced time spent on notes by 0.36 hours—about 22 minutes per eight hours of patient time—and improved measures of exhaustion and interpersonal disengagement. The full study describes a 24-week trial within one health system. This is a useful workflow improvement; it does not establish that total healthcare spending, premiums, or staffing fell.

The research frontier is moving too. In an October 27 update on drug development, Anthropic described work with six life-sciences organizations. One example involved Generate Biomedicines scientists using Claude Code to propose changes to a component of CAR T-cell therapy, followed by laboratory testing. That is an example of AI participating in scientific work beyond paperwork. These early research results do not establish an approved treatment, a patient benefit, or lower insurance premiums.

The key question for the benefits industry will be whether those productivity gains are passed through to employers, employees, and households. Technology can lower the cost of producing care without automatically lowering the total amount the country spends. Better access may increase utilization, and breakthrough therapies may arrive with high launch prices. Plan design, purchasing strategy, regulation, and competition will determine who captures the savings.

Prevention changes the risk pool

Insurance costs are downstream of human health. If fewer people develop preventable disease, the system has fewer severe and expensive claims to finance.

One encouraging signal is the decline in alcohol consumption. Gallup reported that 54% of U.S. adults said they drink alcohol in 2026, unchanged from 2025. The record low has held; the latest survey does not show another annual decline in the share of adults who drink.

Once I understood what alcohol actually is, I felt stupid for ever drinking it. I saw no good reason to keep putting it into my body, so I stopped. I have not had an alcoholic drink in years, and none of my friends drink either. That was a decision about health and what made sense to me. My own circle is not a national survey, but it shows me that social life does not need alcohol at its center.

Drinking is uncool. Alcohol should lose the glamour that cigarettes once had. I want the same cultural rejection of the idea that consuming a harmful product makes someone sophisticated, successful, or more fun. We should stop treating alcohol as an essential ingredient of adulthood, celebration, or business networking. A product that can damage the brain and cause cancer should not get a social free pass because it comes in an expensive bottle.

That is what I mean by alcohol becoming the new smoking: the normalization deserves to go. It is a cultural argument, not a claim that alcohol and cigarettes carry identical medical risks. The target is the glamour and the pressure to drink, not the dignity of people who need treatment or support.

Changing health perceptions are measurable. Gallup's 2026 survey found that 51% of U.S. adults believe one or two alcoholic drinks a day are bad for health, compared with 27% in 2001. In the same survey, 86% had heard at least something about research on alcohol’s long-term health effects, up from 79% in 2024. More people are paying attention to the science.

Alcohol is a neurotoxin and a toxic substance—a poison. Every alcoholic drink contains ethanol. Ethanol, the alcohol in beer, wine and spirits, can directly harm the nervous system. The World Health Organization identifies ethanol as toxic and dependence-producing. The National Institute on Alcohol Abuse and Alcoholism documents direct neurotoxic effects: heavy drinking over time can damage brain tissue and peripheral nerves, affecting memory, reasoning and sensation.

The harm extends beyond the brain and liver. Alcohol can damage the heart and pancreas, disrupt the gut and weaken immune defenses, particularly with heavy use. It also causes several cancers. The National Cancer Institute explains that the body converts ethanol into acetaldehyde, a toxic chemical that can damage DNA and proteins. Even light drinking can increase the risk of some cancers. The amount, pattern and duration of drinking affect the level of risk.

It is important to describe the impact accurately. Excessive alcohol use is a major preventable social and economic burden, but it is not documented as the nation's number-one healthcare expense. The Centers for Disease Control and Prevention estimates that excessive drinking cost the United States about $249 billion in 2010, its latest comprehensive estimate; healthcare represented 11% of that total, while most of the burden came from lost productivity.

The broader point remains powerful. When behavior changes at population scale—less smoking, less drinking, safer driving, better nutrition, more effective screening—the claims environment eventually changes too. Prevention is not a wellness slogan. Done well, it is risk engineering.

Aging may become a more manageable risk

The largest long-term opportunity may be healthspan: increasing the number of years people live in good health, rather than merely adding years of illness and dependency.

ARPA-H is funding clinical research aimed at understanding and targeting the biology of aging, while the National Institute on Aging's geroscience work examines how the mechanisms of aging influence multiple chronic diseases. This research is early. No responsible employer, adviser, or policymaker should treat it as a promise to cure aging.

But even partial success would matter enormously.

If one intervention could delay several age-related conditions instead of treating each disease only after it appears, the effect would reach far beyond medical claims. People could remain active longer. Caregiving pressure could ease. Disability could decline. Experienced workers could choose to stay in the labor force longer. Retirement, life insurance, disability, long-term care, and health benefits would all have to evolve together.

The future of employee benefits is therefore inseparable from the future of longevity.

Better information and better prevention are only part of the opportunity. Physical assistance, safer transportation, and care delivered closer to home could change costs that extend well beyond a medical bill.

Chapter 6: Robots and the cost of everyday life

Much of the burden of illness is practical: reaching a clinician, recovering at home, helping a parent, or arranging support around a working day. Robotics could change that burden if it becomes safe, reliable, and affordable in the places people actually live.

The auto-insurance preview

Auto insurance offers a useful, imperfect preview of what changes in risk and insurer economics can mean for customers.

The safety mechanism is real. The Insurance Institute for Highway Safety and Highway Loss Data Institute report that automatic emergency braking is associated with 50% fewer front-to-rear crashes, 56% fewer injury-producing front-to-rear crashes, and lower claim frequencies for both vehicle damage and injuries. Advanced sensors can make the crashes that still occur more expensive to repair, but preventing crashes changes the economics of insurance at the source.

That is the healthcare analogy: the greatest savings will not come only from negotiating a lower price after something goes wrong. They will also come from preventing the event, detecting it sooner, and treating it with less labor, less delay, and less physical intervention. Whether those improvements reach customers is a separate question that insurers should be expected to answer.

Robotics changes where care can happen

Robotics can extend scarce clinical expertise beyond the walls of a major medical center.

In August 2026, ARPA-H announced funding for teams developing autonomous systems for time-sensitive stroke procedures, including robotic platforms and microscale tools intended to make treatment less invasive and more widely accessible. The work is experimental, not a finished promise. But it points toward a future in which advanced procedures can be delivered with greater consistency in more places.

That matters because healthcare cost is not only the price of a device or a physician's time. It is also delay, travel, missed work, complications, repeated appointments, and the damage caused when treatment comes too late.

When robotics reduces the expertise bottleneck, care can move closer to the patient. When AI improves triage, the right patient can reach the right intervention sooner. When procedures become less invasive, recovery time can fall. Those are medical advances, workforce advances, and economic advances at the same time.

When the caregiver moves in: home robots and assisted living

I believe one of the most consequential uses of robotics will be caring for people in their own homes. Instead of moving a person into assisted living because they need help, move the help into their home. If a robot can safely and reliably provide much of that daily support, it could dismantle a substantial part of the economic case for facility-based assisted living.

Imagine a robot that lives with you: preparing meals, keeping the home usable, bringing essentials within reach, offering reminders, and contacting family or a care team when help is needed. More demanding personal care—helping someone transfer from a bed, bathe, or use the bathroom—would require much stronger proof of safety and reliability. These are capabilities I want this technology to achieve, not a description of a proven care robot available today.

Tesla's Optimus makes that possibility easier to picture. Tesla describes a general-purpose humanoid robot intended for unsafe, repetitive, or boring tasks. That ambition is not proof that Optimus can independently care for a frail adult, deliver medical treatment, or respond safely to every emergency. My argument concerns what capable home robots could change if those difficult problems are solved; it does not depend on a promised Tesla release date, price, or medical capability.

A challenge to the assisted-living business model

Assisted living combines housing, staffing, supervision, meals, and daily support in one place. If reliable help can be delivered in a person's existing home at a lower total cost, the alternative is no longer simply a different facility. It is staying home. Families could preserve familiar surroundings while purchasing support instead of a room and a share of a building.

That is a threat to occupancy, pricing power, and the value of care-facility real estate. Some operators could become providers of home support, remote supervision, maintenance, and human response. Others could face a shrinking market. In my view, a mature home-care robot would be far more disruptive than another efficiency tool inside an assisted-living center: it could remove the need for many people to move there at all.

This would not make every residential setting unnecessary. Complex dementia, skilled nursing needs, unsuitable housing, isolation, and personal preference can still make human support or residential care essential. A machine's availability is not the same as companionship, clinical judgment, or a safe home. The opportunity is to make staying home a realistic choice for more people, with reliable human backup when needed.

Could home healthcare become free to the patient?

That is the future I want us to work toward: essential care at home with no charge when a person needs it. If AI and robotics reduce the full cost enough, governments, health plans, or employers could fund a basic level of home support as shared infrastructure. Free at the point of use is a financing choice that lower costs could make more achievable—not a claim that robots themselves cost nothing.

Equipment, repairs, energy, insurance, cybersecurity, and human response still have to be paid for. Daily living assistance also differs from skilled medical care. For example, Medicare's existing home-health benefit can cover eligible home-health services with no patient charge, but it does not cover round-the-clock home care or personal care when that is the only care needed. My vision would go beyond those current limits; it is not a statement of present coverage for Optimus or other household robots.

Making that vision useful requires consent, privacy, safe physical assistance, emergency escalation, and backup when a machine or connection fails. The FDA's guidance for home-use medical devices recognizes that the home, the user, and the device create a distinct safety challenge. A demonstration on a controlled stage is not enough for a system that someone depends on to live safely.

For employers, the implications reach well beyond medical claims. Employees often organize their working lives around a parent's care or a family member's recovery. Dependable support at home could reduce that burden, preserve independence, and change how we design caregiving leave and long-term-care benefits. The question should be whether people are safer and better supported at a lower full cost—not how impressive the robot looks.

All of these possibilities lead to the same financial question: when the underlying work becomes less expensive, how much of that improvement reaches the person who needs help?

Chapter 7: Who receives the savings?

There are three different measures to keep in view: the cost of delivering a service, the total amount spent on healthcare, and the premium an employer or household pays. Improving one does not automatically improve the others.

An insurer can process a claim more cheaply without changing the price charged for the treatment. A clinical tool can reduce the cost of one procedure while increased access raises the number performed. Savings can also be absorbed by implementation costs, rework, or new services—or retained as margin. Employers should ask which cost actually fell and where that reduction appears in their own results.

This is why I see AI's effect on healthcare financing as a choice as well as a technological possibility. We need to decide what a more productive system should make possible for the people relying on it.

First, identify who actually bears the claims risk

The name on an insurance card does not by itself tell us who pays the medical claims. In a self-funded employer plan, the employer bears that responsibility, often hiring an insurer or third-party administrator to run the plan. KFF's 2026 survey reports that 71% of covered workers are in self-funded plans.

That changes the question a business leader should ask. If an administrator becomes more efficient, does the employer's fee fall? If medical claims fall, does the employer receive that benefit after fees, risk-protection costs, and other plan expenses? If a vendor promises savings, what baseline, exclusions, and reconciliation determine whether those savings are real? An insurer's efficiency announcement and an employer's financial result are different measures.

For fully insured coverage, the insurer bears the covered claims risk in exchange for premiums. Federal medical loss ratio rules generally require individual and small-group insurers to spend at least 80% of premium revenue on medical care and quality improvement, and large-group insurers at least 85%, with rebates when applicable. The remaining share covers more than profit; it also includes administration and other costs. These rules do not mean every dollar saved by AI automatically becomes a dollar off an employer's next renewal. They are not a universal cap on profit, and they should not be applied to a self-funded employer's administrative contract as though it were an insurance premium.

The practical test is a reconciliation: show the change in administrative charges, medical spending, employee costs, service quality, and access. Separate what the technology did from changes in prices, enrollment, benefits, or the health of the covered population. That is how a productivity story becomes an accountable purchasing decision.

What happens to employee benefits?

Employer-sponsored coverage is not about to disappear. CMS projects employer-sponsored insurance enrollment to remain roughly stable through 2034, and the current system still places employers at the center of how millions of Americans access and finance care.

What will change is the value employers need from benefits partners.

From transaction to navigation

When plans, clinical options, funding arrangements, and technology tools multiply, employers need fewer order takers and more navigators. The benefits adviser's job becomes helping an organization decide what is credible, what integrates with benefits administration, what protects employees, and what actually changes outcomes.

From one-size-fits-all to personal pathways

AI can make benefits guidance more responsive to an employee's circumstances without turning the workplace into a surveillance system. The opportunity is personalized help; the obligation is strict data stewardship, transparency, and human accountability.

From paying claims to preventing them

Benefits strategy will increasingly connect medical coverage with prevention, mental health, medication support, navigation, leave, disability, and caregiving. Employers should evaluate whether a program removes friction and improves health—not whether it merely adds another vendor logo.

From employment lock-in to greater portability

As work becomes more flexible and individual-market technology improves, some benefits may become more portable. Arrangements such as ICHRA/CHOICE Arrangements already point toward a model in which employers can fund coverage without forcing every worker into the same group plan.

CMS now uses the name CHOICE Arrangements for ICHRAs. Employees generally can keep their individual insurance after leaving a job if they continue paying the premiums, but the employer's reimbursement does not automatically continue. Portability of the policy is different from portability of the funding. Compare local networks, employee costs, and administration before changing an employer's approach.

From universal access as ideology to universal access as infrastructure

If technology materially lowers the marginal cost of diagnosis and treatment, a universal base layer of essential care could become easier to finance and administer. That is a possibility, not a prediction of a particular law or a claim that one political model is inevitable.

In that future, employer benefits would not necessarily vanish. Employers might differentiate through faster access, broader networks, supplemental coverage, income protection, mental-health resources, family support, and a better service experience. Public infrastructure and private benefits can coexist.

What remains to insure when healthcare gets cheaper?

Insurance is most valuable when a loss is uncertain, infrequent, and too large for an individual or employer to absorb. It is a less efficient way to finance predictable services that become inexpensive enough to purchase directly. If AI makes screening nearly continuous, routine diagnosis faster, administration largely automatic, and many treatments less labor-intensive, part of healthcare could move from complex claim reimbursement toward simpler access and direct payment.

That would not mean the end of health insurance. Catastrophic injuries, rare diseases, advanced therapies, long-term care, and other severe risks could remain expensive even in a far more productive system. Carriers and reinsurers would still pool those risks, finance unpredictable costs, protect households, and manage reserves. But the portion of healthcare requiring traditional insurance could shrink as ordinary care becomes less scarce.

The health plan of the future may have three layers:

Universal access does not require one predetermined political structure. The base layer could be financed and administered through different combinations of public programs, regulated private plans, individual coverage, or new arrangements that have not yet been designed. The principle matters more than the label: essential care should remain available when a person changes jobs, works independently, experiences displacement, or leaves the labor force.

For carriers, the strategic direction is clear. They become smaller, more automated platforms for risk, payment, navigation, governance, and catastrophic protection—not vast organizations built around manually moving claims and correspondence. Some capabilities will remain inside carriers; others will move to specialized technology companies or shared infrastructure. Competition and regulation will determine whether the productivity dividend appears as lower premiums, better coverage, improved service, or simply higher margins.

This creates an important obligation for employers, advisers, and policymakers. Lower production costs do not automatically become lower healthcare costs for the public. Purchasers must measure administrative expense, demand transparent outcomes, and insist that gains from automation reach the people and organizations financing care.

What I would put on the table

I am proposing a direction, not presenting a finished legislative package. Before asking a family or a taxpayer to rely on it, I would want five choices made explicit.

Define the foundation. My starting point is dependable access to essential prevention, primary care, mental-health care, diagnosis, necessary treatment, and medicines, with coordinated protection against catastrophic costs. The boundary between routine care and catastrophic coverage must not become another gap for a patient to fall into. An evidence-based benefit definition, independent review, adequate provider access, and a meaningful appeal would matter as much as the coverage label. The promise should be essential care, not an unlimited entitlement to every service at any price.

Specify who pays. Compare transparent combinations of general revenue, employer contributions, and income-related household payments. Each has tradeoffs. A system financed mainly through payroll would still depend heavily on the labor market whose stability I am questioning. Broader financing would move costs onto other taxpayers and requires its own distributional analysis. Existing public and employer spending cannot simply be counted as available cash without showing which obligations it currently funds. I would require an independent estimate of total costs, taxes, household payments, employer costs, and provider effects—including a scenario with no additional AI savings.

Protect the transition. Map how the foundation would coordinate with Medicare, Medicaid, ACA coverage, and employer plans before changing any entitlement or funding stream. Preserve treatment, prescriptions, and coverage while people move between arrangements. Specify who is responsible when public and private coverage overlap. A phased transition with measurable access and continuity standards is more credible than declaring that every existing institution becomes unnecessary overnight.

Keep a constructive employer role. Employers could provide supplements, income protection, caregiving support, broader choice, and personal service above the foundation. Supplements should add value without making essential care depend again on whether someone has a generous employer. The financing proposal must also state whether and how employers contribute to the base layer.

Publish the scorecard. Judge the result by access, wait times, health outcomes, financial protection, total costs, and the distribution of those costs—not enrollment alone. Test assumptions, report failures, and specify corrective action. Technology may make the proposal easier to deliver. It does not excuse us from proving that the financing and the care actually work.

Benefits must become transition infrastructure

This workforce future strengthens the case for essential healthcare that is less dependent on one job. If AI reduces the number of stable administrative roles, an employment-based financing system becomes harder to rely on precisely when families need continuity.

A universal foundation could provide that continuity while employers compete through broader networks, supplemental coverage, income protection, mental-health support, and service. My argument is that AI strengthens the need for this foundation while potentially lowering the cost of delivering care. The policy choices and financing would still have to be worked out.

Portable coverage, transition support, and training tied to real roles should accompany workforce redesign. Human accountability remains essential: an algorithm can organize choices, but a responsible person must help resolve consequential health and benefits decisions.

Related reading

Andrew Decker in The Daily Record: MD employers should plan for fewer administrative jobs now ICHRA is now called CHOICE: what employers should know

A more portable foundation would still require financing, administration, and difficult policy choices. It would also require institutions capable of using powerful technology without abandoning responsibility for the people affected by it.

Chapter 8: Who remains accountable?

Maryland is participating in the oversight, too

Maryland was one of 12 participating states in the NAIC's 2026 AI Systems Evaluation Tool pilot. The pilot plan scheduled use from March through September and included California, Colorado, Connecticut, Florida, Iowa, Louisiana, Pennsylvania, Rhode Island, Vermont, Virginia, and Wisconsin alongside Maryland. The purpose was to test a structured way for regulators to examine insurers' AI use and governance through market-conduct and financial oversight. Participation was not a certification that an insurer's AI was safe or an endorsement of a particular product.

This work is still developing. As of this article's November 12 research cutoff, the NAIC's November 16 meeting agenda called for consideration of an updated AI Risk Evaluation Supplement and a report on the pilot. Consideration on an agenda is not adoption.

Maryland also has an existing statutory boundary. For the state-regulated carriers and utilization-review arrangements covered by Insurance §15-10B-05.1, AI must not replace the healthcare provider's role, and the statute prohibits an AI, algorithm, or other software tool from denying, delaying, or modifying healthcare services. It also requires attention to individual clinical information and oversight. This is a rule for covered utilization review, not a ban on every administrative use of AI.

A newer development is the October 30 draft report of Maryland's adverse-decisions workgroup. Its detailed Recommendation 6 proposes surveying providers and payers about AI use. The report discusses concerns on both sides of the transaction, including coverage decisions and coding. It remains a draft recommendation—not a new law or proof that AI caused the rise in denials.

Other states are moving beyond broad principles

• Texas: Its June 12 AI bulletin states the department's expectation that a person review and agree with consequential AI decisions before action. It also points to the statutory prohibition on using AI to make adverse utilization-review determinations.
• South Carolina and North Dakota: South Carolina's September 24 bulletin sets expectations for a written AI program and governance. North Dakota's October 5 bulletin encourages such a program and explains that AI use remains subject to examination. The wording and legal context differ by state.
• California: An October 27 announcement describes finalized long-term solvency-planning rules addressing risks that include AI and other emerging technologies. That concerns insurer financial resilience, a different issue from who may make a medical-necessity decision.

The direction is clear: insurers are being asked to explain how their AI works, how they test it, and who is responsible when it affects a person. Faster processing does not remove the obligation to make a fair decision or provide a meaningful path to challenge it.

Three decisions for your next benefits review

• Review renewal costs, provider networks, and employee expenses using today's numbers. Do not budget for AI savings that a vendor has not demonstrated.
• Ask each technology vendor for measurable results, data-access controls, and a named human who handles errors and appeals.
• Compare group coverage with ICHRA/CHOICE Arrangements where appropriate, and explain how coverage and funding change when an employee leaves a job.

What employers should do now

The long-range future may be dramatic, but the right actions today are practical:

1. Build a clean benefits data foundation. You cannot evaluate outcomes, navigation, or AI tools when eligibility, enrollment, and utilization data are fragmented or unreliable.

2. Demand evidence from technology vendors. Ask what problem is solved, how results are measured, how data is protected, and whether savings are independent and verifiable.

3. Invest in prevention that employees can actually use. Access, simplicity, trust, and follow-through matter more than the size of a vendor catalog.

4. Protect the human layer. Employees need a knowledgeable person when a diagnosis, claim, leave, or coverage decision becomes complicated.

5. Design for change. Funding strategies, administration, ICHRA/CHOICE Arrangements, leave programs, and communication should be evaluated as parts of one workforce system—not isolated annual renewals.

6. Plan honestly for lower administrative employment. Model which departments will require fewer people instead of assuming every productivity gain can be absorbed without workforce reduction.

7. Map tasks before roles are disrupted. Identify which work will be automated, which work will be augmented, and where employees can move as processes change.

8. Rebuild the entry-level ladder. Create supervised pathways for people to develop judgment even when AI performs many of the traditional beginner tasks.

9. Share the transition, not only the savings. Give employees advance notice, credible pathways into remaining roles, and meaningful support when reductions cannot be avoided.

10. Make benefits more portable. Evaluate healthcare, income protection, training, and transition support for a workforce likely to change jobs and employment arrangements more often.

Three requests to Congress

First, commission a public, independently costed comparison of ways to preserve essential coverage across job loss, self-employment, and transitions between jobs. Test current arrangements alongside a portable foundation. Show costs and distributional effects under conservative productivity assumptions, including no further AI savings. Congress should be able to debate a real financing choice rather than a technology promise.

Second, establish clear accountability for consequential insurance AI within the relevant federal and state jurisdictions. People should have understandable reasons for adverse decisions, a usable human review and appeal, and a responsible organization that cannot shift blame to its software vendor. Oversight should examine accuracy, inappropriate denials, access, and error correction as well as processing speed. Use the state pilots as evidence to learn from, not a certificate of safety.

Third, connect workforce-transition policy to benefits continuity and credible routes into paid work. Track hiring, vacancies left unfilled, entry-level opportunities, and displacement—not just announced layoffs. Judge retraining by whether people reach real jobs and maintain essential coverage, rather than by how many course certificates a program issues.

Three decisions for a corporate board or executive team

Within 30 days, establish the baseline. Give named leaders responsibility for a task-level workforce map and a benefits-cost map. Identify the funding arrangement, the administrative charges, the medical spending, and the points where employees lose time or access. Document what is measured and what is only estimated.

Within 60 days, test one bounded workflow. Compare it with current performance, including supervision, mistakes, human handoffs, implementation expense, and effects on employees. For carriers, measure the appeal and the unresolved case as carefully as the automated interaction. Keep a responsible person in control of consequential decisions and a workable way to stop the deployment.

Within 90 days, decide how gains will be used. Set a documented threshold for expansion. Explain whether verified gains reduce fees, improve access, fund transition support, or become margin. Do not build budgets around unverified savings or treat a successful pilot as permission to automate every adjacent function.

These are my proposed management milestones, not a claim that every organization can complete procurement or deployment in three months. The point is to turn a broad AI ambition into decisions someone owns, evidence others can inspect, and obligations employees can understand.

A more abundant future

Healthcare will not become cheap because someone declares it so. It will become less expensive where technology removes delay, error, scarcity, and unnecessary intervention—and where the market and public policy pass those gains through to the people financing care.

The near-term numbers still point upward. The long-term tools point somewhere more hopeful—and more disruptive.

My long-range thesis has four connected parts. I expect prevention, AI, robotics, and healthspan science to reduce the frequency, severity, and delivery cost of many medical events. The same systems will reduce the number of people required to operate healthcare, insurance, and the administrative departments of nearly every employer. Lower and less stable employment will weaken a benefits system organized around the traditional full-time job. A universal base layer of healthcare will consequently become both more necessary and, as the cost of delivering care falls, more achievable.

Employer benefits will not disappear in that future. They will evolve from being the only reliable gateway to care toward organizing supplemental protection, access, choice, income security, and human support above a portable foundation.

For employers, the goal is not to predict the exact year or policy model. It is to build a benefits strategy resilient enough to improve as the technology does.

That future is not here yet. But it is close enough to plan for.

Evidence at a glance

This is a guide to the weight each kind of evidence can carry, not a substitute for the linked material. The research cutoff is November 12, 2026.

Company results remain attributed. My judgments about future employment, prevention, robotics, and healthcare financing remain arguments to evaluate against evidence, including evidence that cuts against them.

About Andrew Decker

Andrew Decker is President of Strategic Sequoia Group Inc. His work spans employee benefits, insurance, real estate, artificial intelligence, blockchain, cloud computing, and the future of work. He holds licenses in life and health insurance, property and casualty insurance, and real estate. Through SSGI Benefits, he helps employers evaluate benefit strategies, administration, funding, compliance, and workforce change — including benefits for entrepreneurs building their first plan.

Perspective note

This article reflects Andrew Decker's forward-looking perspective. It is not medical, legal, tax, or investment advice; it does not predict a particular law or guarantee future healthcare or insurance costs.