Which business models are hardest for AI to replace?

Is AI actually replacing business models yet?

AI is replacing chunks of work much faster than it is wiping out entire business models.

Stanford's latest AI Index puts the change in perspective. AI is already used somewhere inside 88% of surveyed organizations, and generative AI is used in at least one business function by 70%. Yet AI-agent deployment remains in the single digits across almost every business function. Companies are adopting the technology very quickly, while fully handing businesses over to autonomous agents is still rare.

The labor market shows the same split. Stanford found that employment among software developers aged 22 to 25 has fallen by nearly 20% since 2024, one of the clearest early signs that AI can reduce demand at the entry level in highly exposed digital jobs. At the same time, the latest Bureau of Labor Statistics projections still expect overall software-developer employment to add about 175,000 jobs over the next decade.

So we should be careful with the word "replace." AI can remove junior work, reduce headcount, collapse prices and radically change margins long before an entire industry disappears.

For a business owner, the more useful question is what remains after AI has done the easy 50%, 70% or even 90% of the cognitive work. If customers still need the company to hold money, move a package, fix a building, absorb a financial loss or keep an authoritative record, the business can survive a surprisingly large amount of automation.

What the business mainly sells	What better AI does	Replacement risk
Generic digital output	Produces the output directly	Very high
Advice and analysis	Does much of the thinking	High
Workflow software	Automates work inside the system	Medium
Marketplace infrastructure	Improves matching and discovery	Medium to low
Regulated responsibility	Helps the licensed operator	Low
Physical execution	Helps plan the work	Very low
Scarce physical capacity	Helps allocate the asset	Very low

Why are digital service businesses so exposed to AI?

Purely digital service businesses are among the easiest models for AI to attack because the technology already lives in the same environment as the product.

If a customer sends information through a computer and receives another piece of information back, AI can potentially sit between those two steps. That covers a huge amount of copywriting, translation, illustration, transcription, research, basic coding, presentation design, routine bookkeeping and low-level consulting.

The International Labour Organization's latest global analysis reaches a similar conclusion from occupational data. Clerical work remains the most exposed category, while exposure has also risen among financial analysts, programmers and other professional jobs whose work happens almost entirely on a screen.

The important point for a business is pricing. AI does not need to produce perfect work to wreck the economics of a $2,000 service. If a customer can get 80% of the result in five minutes for $20, the old provider has to explain what the remaining $1,980 buys.

That is already a harder argument for generic content agencies than it is for a plumber, insurer or logistics operator.

Some digital businesses will survive through specialization, distribution or exceptional brand strength. Generic production work still has one of the weakest structural positions because the customer can increasingly access the production technology directly.

Are plumbers and electricians really among the most AI-proof businesses?

Plumbing, electrical work and similar skilled trades remain some of the hardest ordinary businesses for AI to replace today.

The latest U.S. employment projections actually became more favorable for several of these occupations. The Bureau of Labor Statistics expects electrician employment to grow 9% from 2025 to 2035, creating roughly 75,900 net new jobs and 72,700 openings per year. Plumbing, pipefitting and steamfitting are projected to grow 7%, with around 42,000 openings a year.

The electrician numbers are especially interesting because the BLS now explicitly points to AI-driven data-center construction and rising electricity demand as one reason more electrical infrastructure will be needed. AI may indirectly create additional work for one of the occupations it struggles to automate.

AI can still transform the economics of a local contractor. It can answer calls, produce estimates, optimize routes, recognize equipment from photographs, prepare invoices, order parts and help technicians diagnose problems. A ten-person company may eventually handle the administrative work of a fifteen-person company.

The strongest local service companies should therefore become more productive as AI improves. Their back offices shrink while the part customers actually pay for remains stubbornly physical.

Will robots eventually replace those physical service businesses too?

Robots will eat into physical work, but the current deployment data still show a huge difference between predictable environments and messy ones.

The International Federation of Robotics recorded 542,000 industrial robots installed worldwide in 2024, more than twice the annual level a decade earlier. The technology is clearly scaling. Preliminary figures also show U.S. industrial-robot installations rising 11% in 2025 to about 38,000 units.

Look at where those robots work, though. Factories are designed around repeatability. Warehouse floors are flat. Parts arrive in known positions. Safety zones can be controlled. The same pattern appears in professional service robots: transportation and logistics alone accounted for about 103,000 units in the IFR's latest full dataset, more than half of professional service-robot sales.

A residential plumber faces a nastier robotics problem. Every house can be different. Pipes sit behind walls. Fittings corrode. Access points are awkward. Customers and pets move through the workspace. Something unexpected can happen five minutes into the job.

Robotics will keep pushing outward from structured environments, and portions of construction, cleaning, logistics and maintenance will become automated. Still, the commercial hurdle rises sharply once robots need excellent manipulation, mobility, safety and economics in thousands of unpredictable locations.

For now, physical work gets more defensible as the environment becomes less standardized.

Environment	Current automation level	Resistance to full replacement
Factory production line	Very high	Low
Warehouse transport	High and rising	Medium
Commercial cleaning	Growing quickly	Medium
Medical procedures	Growing quickly	Medium to high
Residential repair	Limited	High
In-home personal care	Very limited	Very high

Are businesses that own scarce physical assets even safer from AI?

Owning genuinely scarce physical capacity is one of the strongest business-model defenses against AI.

An AI assistant can find the perfect hotel room, but it cannot create another room in central Paris tonight. It can optimize warehouse usage without creating a warehouse next to a major port. It can plan electricity consumption without instantly adding transmission capacity to a congested grid.

We therefore see a big difference between businesses selling information about an asset and businesses controlling the asset itself.

Property-search websites, travel-planning services and brokers can lose part of their informational advantage as AI improves. Owners of airports, warehouses, specialized factories, energy infrastructure, spectrum, prime real estate and constrained transport capacity face a different problem. Their returns may rise or fall with supply, financing costs and competition, but software alone cannot reproduce the underlying capacity.

AI is actually increasing demand for some physical infrastructure. Stanford's latest AI Index notes that infrastructure spending by large technology companies has exploded alongside model adoption, with Google alone reporting more than $150 billion of annual capital expenditure in 2025.

That does not make every data center or power asset a good investment. Overbuilding remains possible. But it shows how strange AI disruption can become: the software gets dramatically smarter while certain physical inputs become more valuable because more of them are needed.

Are payment networks like Visa and Mastercard hard for AI to replace?

Payment networks are among the most AI-resistant large-scale business models because smarter software does not remove the need to authorize and settle transactions.

Visa processed 257.5 billion transactions and $14.2 trillion in payment volume in its latest annual reporting. Mastercard handled about $10.6 trillion of gross dollar volume and 175.5 billion switched transactions.

An AI shopping agent could completely change how those payments begin. We may increasingly ask an agent to compare products, choose one and buy it without manually visiting a merchant website.

Money still has to move.

The transaction needs an account, authorization, fraud controls, dispute handling, acceptance by the merchant and settlement between institutions. Machine-initiated commerce may even increase transaction frequency if agents start making more purchasing decisions automatically.

Visa and Mastercard certainly have competitors. Stablecoins, real-time bank payments and account-to-account systems could redirect volume from card networks. That is a genuine business threat, but it comes from competing financial rails rather than from intelligence becoming cheap.

This distinction appears repeatedly in AI-resistant businesses: software may change who decides, but the underlying transaction still has to happen somewhere.

Can AI replace insurance companies?

AI can replace a lot of insurance work, but replacing the insurer itself is much harder because someone still has to carry the financial risk.

Underwriting is highly automatable. Claims documents can be read automatically. Fraud models can improve. Customer support can move toward agents. Policies can be priced using much richer data.

None of those improvements pays a $300,000 fire claim.

The insurer collects premiums partly because it agrees to put its balance sheet behind uncertain future events. Capital, reserves, regulation and contractual responsibility remain attached to the insurance company even if AI eventually performs most of the analysis.

Regulators are reinforcing that responsibility rather than transferring it to model providers. The National Association of Insurance Commissioners has made clear that decisions supported by AI still have to comply with existing insurance laws. Its regulators have also been piloting an AI Systems Evaluation Tool across a group of states to examine how insurers govern and monitor higher-risk AI systems.

That creates a very different outlook for the insurer and for some jobs inside the insurer.

Routine underwriting, document review, customer support and claims administration can shrink dramatically. The risk-bearing institution has a stronger reason to exist because the customer is ultimately purchasing protection against a financial loss.

The same logic applies more broadly to businesses that use their own balance sheet. AI can improve decisions made by lenders, warranty companies and other risk-taking businesses while leaving the need for capital intact.

Does regulation really protect a business from AI?

Regulation creates a serious moat when customers still need a licensed organization to take responsibility for the result.

Healthcare gives us one of the clearest examples. The FDA continues to authorize AI-enabled medical devices, and it recently opened a new regulatory discussion specifically around generative-AI-enabled devices. The questions revolve around risk assessment, premarket evaluation and postmarket monitoring.

Better models are being pulled into the regulated healthcare system rather than making the regulated system disappear.

Europe is following a similar principle for high-risk AI. Under the EU AI Act, deployers of high-risk systems have obligations around monitoring and human oversight, including assigning oversight to people with the appropriate competence and authority.

Regulation will not preserve every job underneath a licensed business. Hospitals can automate documentation. Banks can automate review. Law firms can automate research. Insurers can automate underwriting.

The protection sits higher up the stack. Customers and regulators still need an organization that is legally responsible when something goes wrong.

That makes a licensed clinic much harder to replace than a health-information website, even though both may increasingly use the same underlying AI models.

Are healthcare businesses actually AI-proof?

Healthcare delivery is unusually resistant to full AI replacement even though healthcare contains some of the best tasks for AI to automate.

Medical AI has moved far beyond experiments. The FDA maintains a growing list of AI-enabled devices that have passed applicable premarket requirements, and medical robotics is expanding rapidly. The International Federation of Robotics recorded roughly 16,700 medical robots sold in its latest full dataset, up 91% in one year within its supplier sample.

But diagnosis and documentation are only pieces of healthcare.

Patients still need scans, blood tests, medicines, procedures, rehabilitation, physical examinations, hospital beds, emergency care and help inside their homes. Many of those services combine physical work with strict accountability.

The latest U.S. labor projections underline the scale of the remaining human need. Home health and personal care aides already account for roughly 4.7 million jobs, and the Bureau of Labor Statistics expects another 847,300 positions to be added over the next decade. That is the largest absolute employment increase of any occupation in its current projections.

AI should remove huge amounts of paperwork around those workers. It can improve scheduling, monitoring, clinical decision support and communication.

Helping an elderly person get out of bed safely remains a different problem.

Healthcare therefore contains both extremes: information-heavy jobs that AI can attack aggressively and delivery businesses where physical presence, equipment, regulation and responsibility give the provider a much stronger position.

Are marketplaces like Booking and DoorDash safe from AI agents?

Large marketplaces with real supply and fulfillment are fairly resilient, although AI agents could weaken their control over the customer interface.

Booking Holdings shows why scale matters. Its platforms processed $186.1 billion of gross travel bookings and more than 1.2 billion room nights in 2025. Booking.com alone has access to millions of properties and tens of millions of reported accommodation listings.

An AI travel agent can produce a much better hotel recommendation than an old search filter. Rebuilding live room availability, pricing, merchant relationships, payments, cancellations, customer support and years of marketplace reputation is another challenge.

DoorDash is even more operational. Its marketplaces handled about 3.2 billion orders and $102 billion of gross order value in 2025. An AI agent can decide that we want Thai food tonight, but the order still has to reach a real restaurant and then travel to a real doorstep.

The weaker position belongs to thin marketplaces that mainly collect information already available elsewhere. If an AI agent can query suppliers directly and complete the transaction without the aggregator, that intermediary has very little leverage.

Strong marketplaces therefore need more than search traffic. Supply, liquidity, reputation, payments, guarantees and fulfillment make bypassing the platform much harder.

AI could take over the front door while leaving much of the machinery behind that door intact.

Will AI kill SaaS, or are systems like QuickBooks and Salesforce different?

AI puts ordinary software features under pressure, but deeply embedded systems of record are proving much harder to dislodge.

QuickBooks is a useful case because accounting contains so many automatable tasks. AI can classify expenses, chase invoices, reconcile transactions, explain cash flow and increasingly execute workflows by itself.

Intuit has responded by putting agents directly into the product. In its latest annual report, the company says its agents can increasingly move from producing insights to autonomously completing business workflows. Yet QuickBooks Online Accounting revenue still reached about $5.05 billion for the year, up 23%.

Salesforce gives us another live test. Its latest quarterly results showed Agentforce ARR above $1.5 billion, growing more than 240% year over year under its updated reporting definition. At the same time, Salesforce still had $66.3 billion of remaining performance obligations across the broader business.

Customers are attaching agents to systems that already contain their business data rather than ripping those systems out en masse.

That makes sense. A CRM contains customer histories, permissions, integrations and active workflows. Accounting software holds financial records, bank connections and tax information. Payroll systems contain employee identities, payment instructions and compliance rules.

Interfaces can change quickly. Moving the authoritative state of a company is harder.

The SaaS products in the most danger are lightweight tools whose main value comes from generating something AI can already generate independently. Systems that store and execute the company's actual operations have a much stronger position.

Software model	What really keeps the customer there	AI replacement risk
Generic writing tool	Little beyond the output	Very high
Basic AI wrapper	Interface around another model	Very high
Simple analytics tool	Queries and charts	High
Project-management software	Team history and workflows	Medium
CRM	Customer state and integrations	Medium to low
Accounting platform	Financial records and transactions	Low
Payroll/core transaction system	Money, identity and compliance	Very low

Is proprietary data still enough to protect a business from AI?

A pile of proprietary data is no longer a particularly convincing moat on its own; data generated continuously by operating the business is much stronger.

Static databases age. Competitors can license alternatives, scrape adjacent sources or use models to reconstruct a surprising amount of the information.

Operational data behaves differently.

A payment network sees attempted transactions and fraud. An insurer sees claims. A marketplace sees real conversion and cancellations. An accounting platform sees cash moving through businesses. A logistics company sees actual delays, routes and failed deliveries. A maintenance company learns which machines break after which warning signs.

Every new transaction creates information that can improve the next decision.

This gives some established businesses a useful AI feedback loop. More activity generates better proprietary data, better data improves automation, better automation improves the service, and the improved service can attract more activity.

The advantage comes from operating the underlying network or workflow day after day.

A publisher owning 100,000 generic articles does not get the same protection merely by calling that archive "proprietary data."

Are consultants, agencies and coaches safe because clients still want humans?

Generic advice businesses are much less protected than the "people will always want people" argument suggests.

AI is exceptionally well suited to the first half of many advisory jobs: gathering information, spotting patterns, drafting recommendations, preparing slides, writing emails and answering follow-up questions.

A coach who mainly gives generic productivity advice therefore has a weak moat. So does an agency selling routine content production or a consultant whose deliverable is mostly a report.

The stronger version of these businesses takes ownership of implementation.

Consider two consulting firms hired to reduce a manufacturer's energy bill. One charges for 100 hours of analysis and delivers recommendations. If AI reduces the research to 15 hours, the customer has a good reason to question the old fee.

The second installs monitoring equipment, renegotiates contracts, changes operating processes and takes a percentage of the verified savings. Faster analysis helps that company make more money because the customer pays for the result.

We see the same distinction in cybersecurity, revenue-cycle management, recruiting, performance marketing and IT services. Producing advice is becoming cheaper. Taking responsibility for implementation, operating the workflow and being measured against a result remains much more defensible.

This is one of the biggest changes service businesses should make: move closer to the outcome customers actually care about.

Does customer trust still protect businesses from AI?

Trust protects a business when the relationship involves real consequences and responsibility; familiarity alone will not do much.

Someone asking for a restaurant suggestion can switch from a human recommendation to an AI answer without taking much risk. A family choosing someone to care for a vulnerable parent faces a very different calculation.

The same applies to wealth management, healthcare, childcare, legal representation, major construction work and other decisions where mistakes can be expensive.

Even here, we should avoid assuming that a human conversation automatically justifies a premium. AI can already provide patient explanations, remember preferences and respond instantly. The conversational part of trust is becoming easier to reproduce.

What remains harder is trust backed by something concrete: credentials, continuity, physical presence, insurance, reputation, legal responsibility or a record of handling real outcomes.

Businesses built around that deeper form of trust should hold up much better than businesses relying on the vague idea that customers prefer talking to humans.

Which business models actually get stronger as AI gets better?

Some businesses should become more profitable as AI improves because intelligence is one of their operating costs rather than the thing customers buy.

A plumbing company can automate dispatch and invoicing while continuing to charge for the repair. An insurer can improve underwriting while continuing to earn premiums for taking risk. A marketplace can improve matching while earning fees on transactions. A hospital can reduce administrative work while continuing to perform procedures. A logistics network can optimize routes while moving more packages with the same resources.

Booking Holdings provides a useful real-world example. During a period in which the company was investing heavily in AI and broader operational changes, its transformation program produced roughly $550 million of annual run-rate savings. At the same time, 2025 gross bookings rose 12% to $186.1 billion, revenue increased 13% and adjusted EBITDA rose 20%.

More companies will pursue that combination: automate internally without giving up the part of the value chain customers still need.

That makes AI exposure almost the wrong metric for judging some businesses. A company may automate 70% of its internal tasks and become stronger if those savings stay with the company.

The dangerous situation is when the customer can use the same technology to bypass the company altogether.

Which business models look most vulnerable to AI right now?

Commodity digital production, labor-arbitrage services, thin software wrappers and weak information intermediaries currently sit closest to the danger zone.

Commodity content is the clearest case. When customers mainly pay for text, images, basic videos, translations, presentations or routine research, the marginal cost of creating a plausible alternative has collapsed.

Labor-arbitrage businesses face a similar problem. A company that charges $100 because it employs somebody for $30 to complete a repeatable digital task depends on that labor gap. AI can shrink the required labor so quickly that either prices fall or competitors capture much of the margin.

Thin AI wrappers have another weakness. If nearly all of the underlying capability comes from OpenAI, Anthropic, Google or another model provider, the wrapper needs its own workflow, distribution, proprietary state or execution layer. A nice interface will rarely be enough for long.

Pure information aggregators also deserve skepticism. When the information is public and an agent can contact the actual supplier directly, an extra layer becomes difficult to defend.

The most exposed models share the same uncomfortable characteristic: the customer can increasingly obtain the core result without the business.

Which business models are hardest for AI to replace?

The hardest business models for AI to replace today are businesses that control something AI cannot cheaply generate: scarce physical capacity, regulated responsibility, financial risk, transaction infrastructure, real-world execution or deeply embedded operational state.

At the top, we would put scarce physical infrastructure, licensed physical services, risk-bearing businesses, transaction networks and care businesses requiring real-world presence. Strong marketplaces and systems of record come just behind them because AI can change the interface without easily reproducing the supply, liquidity, integrations or authoritative data underneath.

Outcome-based service companies can also become unusually strong if AI cuts their cost of delivering the result. The vulnerable service companies are the ones still charging primarily for human hours spent producing information.

The line between the two groups becomes clearer if we imagine AI becoming ten times smarter and ten times cheaper.

A company selling generic research has a problem because the customer can buy far more intelligence for almost nothing.

A plumber still has to fix the pipe. Visa still has to settle the transaction. An insurer still has to pay the claim. A hospital still has to treat the patient. A warehouse owner still controls the warehouse. QuickBooks still contains the financial record that the agent needs to act on.

That is the ranking we would use. The safest business models are the ones where better AI makes the company cheaper or better to operate without eliminating the scarce thing customers ultimately need from it.

Business model	What customers still need	AI resistance today
Scarce physical assets and infrastructure	Access to constrained capacity	Very high
Licensed physical services	Real-world work + authorization	Very high
Home healthcare and care networks	Presence + responsibility	Very high
Insurance and other risk-bearing models	Capital that absorbs losses	Very high
Payment and transaction networks	Authorization + settlement	Very high
Deep marketplaces	Supply + liquidity + fulfillment	High
Systems of record	Authoritative business state	High
Outcome-based managed services	Execution + measurable result	High
High-stakes professional services	Responsibility + implementation	Medium to high
General SaaS	Workflow and integrations	Medium
Thin marketplaces	Search and aggregation	Low
Commodity digital agencies	Digital production	Low
Generic information products	Information itself	Very low
Thin AI wrappers	Model capability someone else supplies	Very low