How to build stuff that AGI cannot replace?

What does “AGI-proof” actually mean for a business?

An AGI-proof business is one that can keep capturing value even when AI becomes good enough to perform most of the intellectual work behind the product.

That definition gives us a much tougher test than asking whether AI can copy the software. Assume an extremely capable AGI can code, research, design, negotiate, analyze documents, generate marketing campaigns and operate computers better than most people. What would it still need from your company?

Sometimes the answer is almost nothing. If we sell a report, a logo, a spreadsheet model or a simple software feature, the useful output can potentially be recreated from information alone.

Other businesses control something the AGI still needs. A payment network has credentials, banking relationships and merchant acceptance. A marketplace has actual buyers and sellers. An enterprise system holds permissions and transaction history. A utility owns physical infrastructure. A licensed medical company has regulatory approval attached to a specific legal entity and product.

That difference is more useful than trying to predict which tasks future models will fail to master.

A tougher test is how much of the customer outcome depends on resources outside intelligence itself.

Is AGI already replacing enough work to worry about this now?

Yes. We do not have to wait for universally accepted AGI before building businesses around the assumption that cognitive work will keep getting much cheaper.

METR’s latest measurements of frontier agents are particularly hard to dismiss. On its software-heavy benchmark, the most capable systems are already reaching the point where the benchmark itself struggles to measure their 50% task horizon accurately above 16 hours of equivalent human work. In separate testing disclosed by METR, leading agents successfully completed some software reimplementation tasks estimated to take humans days or even weeks.

The labor data is moving too. Stanford’s Digital Economy Lab recently updated its analysis of millions of US payroll records through mid-2026. It still found no broad employment collapse, but workers aged 22 to 25 in highly AI-exposed occupations were about 19% below the employment level we would expect if they had kept pace with similarly aged workers in less-exposed jobs. The gap had widened from 15% in the researchers’ earlier measurement. Most of the adjustment came through weaker hiring rather than mass layoffs.

At the same time, using intelligence keeps getting cheaper. Stanford’s AI Index found that the inference cost of reaching roughly GPT-3.5-level benchmark performance fell from about $20 per million tokens to seven cents between late 2022 and late 2024, a decline of more than 280 times. Newer models have continued pushing capability upward since that measurement.

The extrapolation still needs care. METR itself says its results mostly cover software engineering, machine learning and cybersecurity, so a 16-hour software horizon does not mean AI can suddenly automate every 16-hour human task. But the direction is already strong enough that building a company around a permanent shortage of basic cognitive capability looks increasingly risky.

What is changing	Recent evidence	What it tells us
Autonomous software work	METR’s strongest agents are pushing beyond a 16-hour measurable 50% horizon on its current task suite	Longer chunks of technical work are entering agent territory
Entry-level knowledge work	Stanford finds a 19% employment shortfall for young workers in highly AI-exposed occupations	The effects are beginning to appear in actual hiring
Cost of basic intelligence	GPT-3.5-level inference fell more than 280x in Stanford’s measured period	Intelligence itself keeps becoming less scarce

Can a pure digital product ever be AGI-proof?

A pure digital product looks fragile when the whole customer outcome can be recreated by generating information or software.

We can see why by looking at what has happened to software production lately. Lovable says users have already created more than 50 million projects on its platform, with around one million additional projects being created every week. The company reached roughly $500 million in annualized revenue in less than three years, and many of the people building on it are not traditional programmers.

That tells us two things at once. There is obviously enormous demand for easier software creation. There is also an enormous new supply of people able to produce software. That second part is easy to underestimate.

A business selling “we can build this application” therefore owns a less scarce capability than it did a few years ago. The same pressure applies to generic research, copywriting, translation, design variations, data extraction, summaries and many other outputs that increasingly come straight from general models.

Some digital products can still become very valuable. The important question is what they accumulate after launch. A simple AI application can gradually acquire customer records, integrations, recurring workflows, unique transaction data, a community or distribution that makes it much harder to replace than its original feature.

That is where many “AI wrapper” arguments go wrong. The wrapper itself may be easy to copy, but a company does not have to remain a wrapper forever.

The dangerous product is the one that remains stateless. If a customer can leave today, prompt another system tomorrow and recover almost the same result without losing anything important, AGI will make that position increasingly uncomfortable.

Will AGI kill SaaS?

AGI will probably wipe out a lot of standalone SaaS features, while software that holds a company’s real operating state could become even harder to remove.

The latest enterprise numbers are surprisingly strong for an industry that was supposed to be swallowed by AI agents. ServiceNow recently reported $29 billion in remaining performance obligations, up 21% from a year earlier, while subscription revenue grew about 25%. SAP’s current cloud backlog reached €22.9 billion, up 27%. Salesforce reported that Agentforce and Data 360 together had reached nearly $3.9 billion in annual recurring revenue, up more than 210%.

Customers are clearly buying AI, but they are often buying it inside the systems they already use.

That makes sense. An AGI can probably recreate a CRM screen. Recreating the company’s actual CRM is another problem. The customer records, sales history, access controls, workflows, integrations, contracts and internal rules have accumulated over years.

The interface may shrink dramatically. Employees could eventually stop navigating dozens of menus and simply ask an agent to perform the work. The underlying system can remain valuable because the agent still needs somewhere authoritative to read from and write to.

The hierarchy is pretty clear. A SaaS company selling a clever interface is vulnerable. A SaaS company that has become the authoritative database for an important process is in a much better position. A platform that also controls permissions and lets agents execute actions inside that process becomes stronger again.

Type of software	What AGI can easily attack	What can remain difficult to replace
Standalone feature	The feature itself	Very little unless the product accumulates something else
Workflow tool	Interface and routine actions	Integrations, configuration and workflow history
System of record	Querying and interface	Authoritative data, permissions and transaction history
System of action	Manual execution	Rights to execute real actions inside governed processes

Is proprietary data really a moat against AGI?

Proprietary data can protect a business, but a continuously refreshed data stream is much more convincing than a database somebody collected once.

Models are getting better at extracting structure from messy public information. A company whose main advantage is “we have a dataset with 100,000 records” should expect that advantage to erode if competitors can reconstruct much of it, buy an alternative or use models to infer the missing pieces.

Live operational data behaves differently.

Stripe, for example, processed $1.9 trillion in payment volume in 2025 across more than five million businesses using the company directly or through platforms. Stripe also manages more than 200 million active subscriptions through Billing. Every payment, fraud attempt, authorization and subscription event adds to an operational stream that changes continuously.

Cloudflare has a similar property at the infrastructure level. Cloudflare Radar draws on traffic observed across the company’s global network and its public DNS resolver to monitor internet traffic, bots, attacks and technology usage. The value comes partly from seeing what is happening across that network now.

An AI model trained on yesterday’s internet cannot magically know every payment attempt, fraud pattern, inventory change, account permission or network event that occurred five minutes ago.

The better test is simple: why will the company receive tomorrow’s proprietary data? If the answer is that customers keep conducting real activity through the product, the moat can renew itself. If the answer depends mostly on a dataset gathered in the past, we should be much more skeptical.

Can AGI copy a marketplace like Uber or Airbnb?

AGI could recreate much of the software behind a marketplace, but it cannot instantly recreate the people, inventory and transaction history already inside a large network.

Uber currently has 208 million monthly active platform consumers and handled almost 3.9 billion trips in its latest reported quarter. DoorDash processed 970 million orders during its latest quarter, up 27% from a year earlier. Airbnb says more than 5.5 million hosts have now welcomed over 2.5 billion guest arrivals.

Writing a ride-hailing application that looks like Uber is tiny compared with making sure a driver is available five minutes away in hundreds of cities. An Airbnb clone can have excellent AI search and still contain no apartment in Paris for next Friday.

Reputation adds another layer. Hosts, drivers, restaurants and customers build histories that influence whether other people are willing to transact with them. An AGI can analyze those histories brilliantly, but the histories still have to come from actual transactions.

We should not assume every marketplace is safe. Agents could make it easier to compare competing networks, and marketplaces with weak liquidity or heavy multi-homing can lose their grip quickly. A directory does not suddenly become defensible because we call it a network.

The strongest marketplace position exists when a customer cares about who is available right now, how many counterparties are nearby and what happened in previous transactions.

Marketplace	Current scale	The difficult part to reproduce
Uber	208M monthly active consumers; ~3.9B quarterly trips	Local driver liquidity and transaction history
DoorDash	970M quarterly orders	Dense networks of merchants, couriers and customers
Airbnb	5.5M+ hosts; 2.5B+ historical guest arrivals	Real inventory, availability, host reputation and reviews

Will distribution become more valuable when AI can build almost anything?

Yes. Distribution is becoming relatively more valuable because the supply of things competing for attention is exploding much faster than human attention itself.

Lovable’s roughly one million new projects per week gives us one indication of what is happening on the production side. Add AI-generated videos, newsletters, mobile apps, online stores, games and advertisements and the cost of creating another thing for people to look at has fallen dramatically.

The price of reaching people has not followed the same path.

Meta recently reported a 14% year-over-year increase in ad impressions across its apps while the average price per ad still rose 12%. More advertising inventory was available, yet advertisers were simultaneously paying more per unit.

Reddit offers another useful datapoint. Weekly active users recently passed 514 million, up 24% year over year, while quarterly revenue rose 61% to $805 million. A large audience made of people voluntarily gathering around specific interests remains economically valuable even while AI makes content production easier.

There are exceptions, obviously, but the basic imbalance is hard to escape. We can create another product in seconds. We cannot create another hour in a customer’s day.

For builders, that pushes owned audiences, recurring customer relationships, communities, strong brands, default placements and existing installed products higher up the priority list.

Product creation used to be one of the biggest bottlenecks. These days, getting chosen can be much harder.

Can trust and reputation protect a business from AGI?

Trust can create a strong AGI moat when it comes from infrastructure, accountability and a proven record; simply being human offers much less protection.

Waymo is an excellent reality check. People once assumed passengers would never trust a vehicle without a human driver. Waymo’s latest safety analysis now covers more than 220 million fully autonomous miles. Compared with human drivers operating in the same areas, Waymo reported 94% fewer crashes causing serious or fatal injuries and 82% fewer crashes involving any reported injury.

Customers can clearly learn to trust machines when the machines repeatedly produce good outcomes.

The more durable opportunity sits underneath that trust. Waymo has accumulated years of safety data, operational experience, regulatory relationships and insurance arrangements. Those things help determine whether a robotaxi can legally and commercially pick somebody up.

Payments show a similar pattern. Visa currently connects roughly 4.8 billion payment credentials with more than 175 million merchant locations. Its recent work on AI commerce focuses on giving agents tokenized credentials, authorization rules, fraud monitoring and controls over what an agent is allowed to spend.

An intelligent agent can decide that we should buy a flight. Somebody still has to establish whether the agent was authorized to spend our money and whether the merchant should accept the transaction.

That is the trust layer worth owning: one attached to identity, history, permissions, guarantees or liability.

Can regulation create a business that AGI cannot easily replace?

Regulation can create a durable barrier when the business controls a legal permission or approved product, while businesses that merely help people complete regulatory paperwork are much more exposed.

Medical devices make the distinction easy to see. The FDA has continued issuing guidance for AI-enabled medical software, including requirements around evidence, cybersecurity, changes to models and management across the product lifecycle. AI is being allowed into regulated medicine; the approval process still attaches obligations to identifiable manufacturers and specific products.

Europe is going in the same direction. Under the EU AI Act, high-risk systems face requirements around risk management, data quality, documentation, logging, human oversight, robustness and cybersecurity. Some of the transparency requirements for AI-generated or synthetic content are already beginning to apply.

An AGI could make compliance work dramatically cheaper. It could draft the documentation, monitor rules and prepare submissions better than many consultants.

But giving an AGI a prompt does not grant a banking license, medical authorization, spectrum allocation, insurance approval or other legal permission.

For founders, the distinction is simple. Selling hours of compliance labor looks vulnerable. Owning the licensed infrastructure that everybody else has to use looks considerably better.

Is building in the physical world safer from AGI?

The physical world slows AGI down, but simply requiring manual labor will not protect a business forever.

Industrial robotics has already moved far past niche adoption. The International Federation of Robotics counted 542,000 new industrial robot installations worldwide in 2024, more than twice the annual number installed ten years earlier. The global installed base reached roughly 4.7 million robots.

Humanoids are now starting to enter the same environment. BMW disclosed that Figure 02 accumulated about 1,250 operating hours during its Spartanburg deployment, moved more than 90,000 components and supported production of more than 30,000 BMW X3 vehicles. The task involved removing and positioning sheet-metal parts for welding during repeated ten-hour shifts.

Those numbers are still tiny relative to conventional industrial automation. They nevertheless weaken the argument that a business is protected simply because a person currently has to use their hands.

Physical automation faces problems software agents do not: robots cost money, break, need energy, have to navigate unpredictable spaces and can injure people. Deployment can take years rather than minutes.

That gives physical businesses more breathing room. It does not create permanent immunity.

A better position is to own something physically scarce that robots will also need.

Which physical assets could become more valuable because of AGI?

Power, grid access, well-located infrastructure and other hard physical bottlenecks could become more valuable as AI improves because stronger AI can increase demand for them.

Data centers give us the clearest current example. CBRE found that supply across the world’s 16 largest data-center markets had grown around 25% year over year by Q1 2026. Even after that huge expansion, global vacancy fell to only 6.7%, with available capacity close to nonexistent in some major hubs.

The International Energy Agency expects global data-center electricity consumption to roughly double from around 485 TWh in 2025 to about 950 TWh by 2030. Electricity used by AI-focused facilities is expected to grow much faster than overall data-center consumption.

You cannot prompt additional transmission capacity into existence. That sounds obvious, but it changes the economics. New grid infrastructure involves transformers, cables, land, permitting, financing and construction. The IEA has repeatedly pointed to those supply constraints as a limit on how quickly new compute can come online.

The interesting cases are where better intelligence increases demand for the underlying asset. That is the inverse of what happens to a generic AI writing tool, where improvements in the base model can absorb the product.

Scarce asset	Why AGI still needs it	What stronger AI can do to demand
Electricity	Compute consumes real energy	Increase total inference and training load
Grid connections	New capacity requires physical infrastructure and permission	Intensify competition for available power
Data-center sites	Power, fiber, cooling and land must meet in one location	Push utilization higher
Specialized factories	Software cannot manufacture physical goods by itself	Generate more designs and production demand
Logistics infrastructure	Physical goods still move through space	Increase and optimize transaction volume

Will people still pay extra for something made by a human?

Yes, but human-made products are likely to command a premium mainly when the identity, provenance or experience is part of what the customer is buying.

We should be skeptical of the broad claim that consumers will always reject AI output. Waymo already shows that people can accept machine execution in situations far more consequential than choosing an image or reading an article. In many digital categories, consumers mostly care about the result.

The stronger evidence appears where authenticity itself creates value.

Bain’s latest global luxury research found that experiential luxury is currently growing faster than physical luxury goods. Demand for experiences has been rising about 1.5 times as fast as demand for tangible products, while immersive bookings across dining, leisure and entertainment were reported up roughly 30% year over year.

You can see why. AI can generate a song resembling somebody’s style. It cannot retroactively make us one of the people who attended a particular live concert. It can create an image in seconds, but that does not give the image the provenance of a specific artist. It can explain Japanese cuisine brilliantly without becoming the chef cooking in front of us tonight.

This leaves room for restaurants, live entertainment, craftsmanship, sports, personal access, collectibles, travel experiences and other products where customers partly pay for the story of where something came from or for having actually been there.

That is a narrower moat than “humans will always prefer humans,” but it is much more believable.

Are services safer from AGI than software?

Most knowledge-work services are highly exposed to AGI, especially when clients mainly pay for a document, analysis or other deliverable that AI can produce directly.

PwC’s latest Global AI Jobs Barometer analyzed more than one billion job advertisements across 27 countries and found a useful split. Jobs where AI removes routine work while leaving substantial judgment were growing about twice as fast as jobs where AI makes the underlying expertise easier for non-experts to perform. Wage growth was also 42% faster in the first group.

Upwork’s freshest marketplace data points in a similar direction. Jobs asking people to improve AI-generated work have risen around 70% year over year, and in software development the number has increased more than eightfold since 2023. Creative categories show almost the same increase.

For now, that is creating plenty of work for humans around AI. But today’s cleanup work is a shaky long-term moat. If models keep improving, some of today's proofreading, debugging and correction tasks should disappear too.

A service becomes much more interesting when the customer is buying execution and responsibility rather than an answer. Installing hardware across 200 stores, obtaining local permits, negotiating supplier contracts, managing a regulated process or guaranteeing that a physical system stays operational includes layers that remain after the intellectual work has been automated.

This is why a ten-page strategy report worries us more than an operator who is contractually responsible for keeping an important process running.

Should we build infrastructure for AI agents instead of competing with them?

Yes. One of the strongest positions today is to build something increasingly capable AI agents will need in order to act in the real world.

Microsoft offers a useful picture of where enterprise software is heading. Dynamics 365 now exposes more than 650,000 actions through MCP across sales, finance, supply chain, HR, field service and customer service. An external agent can perform those actions while using the same underlying business data, rules, permissions and audit trails that govern a human employee.

What holds value is the system authorized to perform the actual transaction, not the intelligence deciding which action to take.

We see the same architecture emerging around money. As seen above, payment networks are now building tools that let agents transact while preserving identity, spending limits, authorization and fraud controls.

Agents will need similar infrastructure in many other areas. They need fresh data to know what has changed. They need identity so another system knows who is acting. They need permission to alter records. They need access to inventory and counterparties. They need mechanisms for payment, verification and dispute resolution. In regulated environments, somebody still has to carry legal responsibility.

The smartest agent on Earth remains surprisingly limited if every useful system tells it, “You are not authorized to do that.”

That creates a large category of businesses whose customer may increasingly be software rather than a person sitting in front of a screen.

So what should you build if you want AGI to make your business stronger?

Build around something scarce that intelligence needs rather than selling intelligence itself.

Across software, labor, marketplaces, payments, regulation, robotics and physical infrastructure, the same pattern keeps showing up.

If AGI becomes dramatically better, generic intellectual output becomes abundant. Another article becomes cheap. Another analysis becomes cheap. Another piece of code becomes cheap. Another basic SaaS feature becomes cheap.

Other things do not scale at the same speed.

A customer relationship still has to be won. A marketplace still needs real counterparties. A payment has to clear. A company still needs an authoritative record of what happened. An agent needs permission before changing a bank balance. A regulated product needs approval. Electricity has to be generated. A transformer has to be manufactured. A warehouse occupies land. A concert happens at a particular place and time.

The best businesses therefore have a useful asymmetry: AGI makes their operation cheaper or creates more activity for them, while the thing customers ultimately need from the business remains scarce.

A marketplace can use better AI to match more transactions. A payment network can process purchases initiated by agents. A system of record can let agents perform far more actions on top of the same authoritative data. A data center can sell capacity into growing compute demand. A licensed platform can automate its compliance work while retaining the permission that competitors still need.

There is no credible way to guarantee that a business survives a truly superhuman AGI. Anyone promising an eternal list of safe industries is pretending to know too much about technology that does not exist yet.

But we can make a much stronger bet than “AI probably cannot do this.”

We can build where better AI creates more demand for something we already control.

That is currently the closest thing to an AGI-proof strategy.