What businesses get better as AI gets better?

Why does better AI suddenly change which businesses are attractive?

Better AI currently rewards businesses that can turn cheaper intelligence into more valuable work without giving away the same advantage to every competitor.

The economics have changed fast. Stanford’s latest AI Index found that frontier models gained roughly 30 percentage points in a single year on Humanity’s Last Exam, one of the harder reasoning benchmarks. At the same time, the leading models have moved much closer together. Anthropic, xAI, Google and OpenAI were separated by only 22 Arena Elo points in Stanford’s latest comparison.

That combination is important. Companies can buy much better intelligence than they could a few years ago, while depending less on one particular model provider.

Microsoft is already changing how it charges for software around this idea. Some Copilot and Dynamics products are moving beyond a simple per-seat subscription toward seat-plus-consumption pricing. Salesforce is doing something similar with Agentforce by charging around the amount of AI work performed. ServiceNow says deployments of its agentic AI increased ninefold in nine months.

These companies are trying to capture a very specific economic benefit: if an AI system can handle twice as much useful work next year, they want customer spending to rise with that work.

A business gets especially interesting when the model keeps improving while the harder-to-copy parts of the product stay under the company’s control.

Do all AI software companies get better when AI models improve?

No. Better AI can actually make a weak AI software business easier to replace.

Imagine a startup whose main product is a polished interface around a frontier model. A new model arrives and suddenly gives the product better reasoning, writing and coding. That looks great for the startup until every competitor integrates the same model.

This problem is becoming harder to ignore because the best models are converging. Stanford’s latest comparison puts four major providers inside an unusually narrow performance range. Customers also have more credible choices across OpenAI, Anthropic, Google, xAI and others.

So we should be skeptical when a company’s main advantage is simply “our AI gives better answers.”

The stronger businesses own something around the intelligence: customer data, permissions, workflow history, distribution, physical infrastructure, trusted professional information or the ability to carry out an action.

AI improvement can then strengthen an advantage that competitors cannot download from the same API.

Which business model benefits most directly from smarter AI today?

Enterprise workflow software currently has one of the best setups because smarter AI can perform more work inside systems companies already depend on.

ServiceNow is a good example. Its AI can work with IT tickets, employee requests, security incidents, configuration data and approval workflows already sitting inside ServiceNow. According to its latest quarterly results, ServiceNow’s AI products crossed $1 billion in annual contract value while total subscription revenue grew about 25%. The company also said agentic AI deployments had increased ninefold in nine months.

Salesforce is showing something similar on a larger customer-data layer. Agentforce and Data 360 recently approached $3.9 billion in annual recurring revenue. Agentforce alone exceeded $1.5 billion. Salesforce also recorded 3.2 billion Agentic Work Units during its latest quarter, almost twice the previous quarter’s level.

We should be careful with that last comparison because Salesforce broadened what it includes in Agentforce ARR. The usage figures are more interesting anyway. Customers are actually asking agents to perform billions of pieces of work inside Salesforce.

Microsoft may have the clearest version of the same strategy. Dynamics now exposes more than 650,000 actions that AI agents can perform across sales, finance, supply chains, HR and customer service. Those actions still inherit the company’s existing rules, permissions and audit trails.

That is a powerful position. As models improve, these platforms can let agents do more without rebuilding the databases, integrations and permissions underneath them.

Why do companies like Salesforce and ServiceNow gain from AI agents instead of getting bypassed by them?

Salesforce and ServiceNow can benefit from smarter AI agents because agents still need somewhere to find reliable company data and somewhere to take authorized actions.

A powerful model does not automatically know which customer owes money, whether an employee can approve a purchase, which version of a contract is valid or whether a support ticket has already been escalated.

Enterprise systems already contain that information.

This helps explain why Salesforce’s Data 360 usage is rising alongside Agentforce. In its latest quarter, Data 360 ingested 104 trillion records, up 355% from a year earlier. ServiceNow finished its latest quarter with $29 billion of remaining contracted revenue while large new contracts also increased strongly.

As AI agents become more autonomous, the value of controlling business actions could actually rise. Someone still has to decide which system is authoritative, what an agent can access and what it is allowed to change.

The interface may become less important over time. Employees may click through fewer CRM screens and fewer IT-service menus. But the databases, permissions and workflows underneath those screens become the environment where agents operate.

That gives well-embedded enterprise software a much better AI position than a standalone chatbot.

Does proprietary company data become more valuable as AI gets smarter?

Yes, when that data gives AI context it cannot easily get elsewhere and directly improves an important decision or action.

Glean is one of the clearest examples. The company connects information scattered across enterprise applications while preserving each employee’s access permissions. Glean reached $300 million in annual recurring revenue only 15 months after reaching $100 million, and its Fortune 500 customer count nearly doubled over the latest year reported.

The useful part of Glean’s position is not simply possessing data. In many cases, the customer still owns the underlying information. Glean has built the layer that can find the right piece of company context and make it usable by an AI system.

Legal AI shows the same principle in a different form.

Harvey recently said that 80% of the Am Law 100 now use its platform. Just days ago, the company raised $550 million at a $15.5 billion valuation. Harvey increasingly combines frontier models with legal workflows, security controls, customer-specific knowledge and tools built around how law firms actually work.

Thomson Reuters approaches the market from the other side. CoCounsel can pair improving AI with legal and regulatory information that Thomson Reuters has spent decades collecting and organizing. CoCounsel had already reached one million professional users across more than 100 countries and territories earlier this year.

A random database of old documents does not create this kind of advantage. The data becomes valuable when it is current, difficult to reconstruct, properly permissioned and closely connected to work people will pay to complete.

The same logic also makes model independence more valuable. A company like Glean or Harvey can adopt a better model when one appears while keeping the customer context and workflow layer that competitors cannot swap in so easily.

Is cybersecurity actually one of the biggest winners from better AI?

Yes. Cybersecurity currently looks like one of the strongest AI beneficiaries because smarter software gives defenders more automation while also giving attackers more ways to operate.

CrowdStrike’s latest results are unusually strong for a company of its size. Annual recurring revenue reached $5.84 billion, up 25%, while the amount of new ARR added during the quarter rose 51% to a record $333 million. Customers using its Falcon Flex platform represented $2.29 billion of ARR, up 101%.

Palo Alto Networks is seeing the same demand from another angle. Its Next-Generation Security ARR recently reached about $9.1 billion, while newer AI-focused security products have grown quickly enough to become meaningful businesses themselves.

This makes sense operationally. Companies are introducing more AI-generated code, more autonomous software, more machine identities and more connections between internal systems. Every one of those creates another place where permissions can be wrong, behavior can become suspicious or an attacker can find a path in.

Identity companies are already adapting. Okta, for example, has started building specifically around AI agents as identities that need authentication, authorization and revocation.

We are therefore looking at a market where the amount of software requiring security can grow faster than the number of human employees using that software.

Business	Latest useful evidence	How better AI helps
CrowdStrike	$5.84B ARR, up 25%	More endpoints, agents and machine activity to secure
Palo Alto Networks	About $9.1B Next-Generation Security ARR	AI creates additional cloud, SOC and agent-security workloads
Okta	Expanding identity products toward AI agents	Autonomous software needs permissions and access controls
Datadog	Revenue recently grew 36%	More AI software produces more systems to monitor

Does AI-generated code make developer tools less useful?

For the strongest developer platforms, AI-generated code is increasing the amount of software they can monetize.

GitHub now has 225 million users, and Microsoft says GitHub Copilot has reached 50 million users. More strikingly, one in three pull requests on GitHub currently involves an AI agent. Copilot revenue also increased more than 60% quarter over quarter after Microsoft introduced more usage-based charging.

Those numbers suggest AI coding is expanding GitHub’s role rather than removing it.

The reason is straightforward. Producing code is only one part of software development. Teams still need repositories, version control, pull requests, testing, security checks, access permissions and deployment processes. If AI helps developers produce much more code, all of those surrounding activities can increase too.

Datadog gives us the downstream version of the same story. Its latest quarterly revenue reached roughly $1.12 billion, up 36%, while the number of customers spending at least $100,000 a year increased from around 3,850 to roughly 4,720. Datadog has also become heavily used by AI-native companies.

There is more risk for standalone coding products whose main attraction is generating code. Lovable, for example, has built a huge business and says users now create around one million projects a week. But coding models available to competitors are improving very quickly too.

So code generation itself is getting cheaper. The businesses around code creation, collaboration, testing, deployment and monitoring have a better chance of benefiting every time more software gets produced.

Do observability and IT monitoring get better as AI creates more software?

Yes. Observability becomes more useful when companies are running more software, more models and more autonomous processes that can fail in unfamiliar ways.

Datadog is a useful test because AI-native companies have adopted it heavily. At its investor day, the company said it served around 650 AI-native customers and 19 of the 20 highest-valued AI-native companies in the comparison it presented.

Chronosphere provides another unusual clue. Before Palo Alto Networks agreed to acquire it, the observability company reportedly went from more than $200 million to more than $500 million in ARR in roughly two quarters as large AI workloads expanded.

The growth makes sense once we look at what an AI application actually involves. A production system can include a model provider, retrieval infrastructure, databases, vector stores, APIs, agent tools, conventional software and cloud infrastructure. When an answer suddenly becomes slow, expensive or wrong, a company needs to know where the failure occurred.

AI can automate some of that monitoring. It also creates far more machinery to monitor.

This category probably becomes more important as agents move from answering questions to continuously running business processes.

Won’t cheaper and more efficient AI eventually reduce demand for data centers?

For now, the opposite is happening: AI is becoming dramatically more efficient per task while total electricity use keeps climbing.

The International Energy Agency says global data-center electricity demand increased 17% in 2025. Consumption at AI-focused data centers grew even faster.

The IEA’s current central forecast has total data-center electricity consumption rising from about 485 TWh in 2025 to roughly 950 TWh in 2030. AI-focused facilities alone are expected to roughly triple their electricity use over that period.

That is happening while AI computations themselves become more efficient. More efficient AI is making many new uses economical, and some of those uses are far heavier than a basic chatbot query. Reasoning models, autonomous agents, image generation and video generation can consume far more computing power.

The cloud companies are behaving as though demand will stay enormous. Microsoft, Amazon, Alphabet and Meta are collectively planning hundreds of billions of dollars of annual infrastructure investment.

Efficiency therefore has not produced lower aggregate demand so far. Cheaper intelligence is encouraging people to consume much more intelligence.

That pattern is especially good for businesses supplying the physical constraints around AI.

Could power, cooling and electrical equipment be better AI businesses than some AI apps?

Yes. Power and cooling currently have one of the cleanest links to AI growth because every new data center eventually runs into physical limits.

Vertiv sells equipment such as power systems and data-center cooling rather than frontier AI models. Its fourth-quarter organic orders increased 252% from a year earlier. Backlog reached $15 billion, up 109%, and the company recorded a book-to-bill ratio of roughly 2.9 times.

Those are extraordinary numbers for an industrial supplier.

The IEA explains why demand has become so intense. AI server racks have become far more power-dense, forcing data centers to rethink cooling, electricity distribution and grid connections. The agency also says transformer, turbine, chip and grid-connection bottlenecks are already slowing some projects.

This gives AI infrastructure an interesting scarcity profile. A software competitor can often reproduce a feature in months. Building substations, manufacturing transformers, securing several hundred megawatts of grid capacity or constructing a large data center is much slower.

The spending also spreads well beyond Nvidia and the cloud providers. Electrical equipment, cooling systems, grid infrastructure and power generation increasingly sit in the same AI supply chain.

AI bottleneck	What is getting harder	Businesses that can benefit
Compute	Supplying enough accelerated computing	Chips, servers, cloud providers
Cooling	Removing heat from denser racks	Liquid cooling and thermal systems
Power equipment	Delivering electricity inside facilities	Switchgear, transformers, power systems
Grid capacity	Connecting very large new loads	Utilities and grid infrastructure
Data-center sites	Finding land with usable power	Data-center developers and operators

Do vertical AI businesses get stronger when general-purpose models improve?

The good vertical AI businesses can get much stronger because every model upgrade lets them automate a larger share of an expensive professional workflow.

Legal AI is giving us an unusually clear experiment.

Harvey has gone from an AI legal assistant used by a smaller group of early adopters to a platform used by 80% of Am Law 100 firms. Its latest $15.5 billion valuation shows how aggressively investors are pricing that position, although the valuation obviously runs ahead of what we can verify from revenue alone.

The more interesting change is inside the product. Legal AI is moving beyond drafting and document summarization toward research, due diligence, contract analysis and customer-built agents. Better reasoning models steadily widen the amount of legal work these systems can handle.

Thomson Reuters is pursuing the same opportunity with a very different advantage: professional information and an existing distribution network. CoCounsel had reached one million professional users earlier this year.

The strongest vertical AI companies therefore do not need to beat OpenAI or Anthropic at building a foundation model. They need to turn improving models into reliable work inside a difficult industry.

Healthcare, accounting, insurance and other professional markets could follow the same pattern, but our confidence should be lower there until we see similarly strong evidence of paid adoption and repeated workflow use.

Can AI make human service businesses better instead of replacing them?

Yes. Service businesses can become much better when AI cuts the amount of human time needed for routine work while people remain available for difficult cases.

Intuit is explicitly building this model around bookkeeping, tax and financial services. Its latest full-year revenue reached $21.4 billion, up 14%, while the company has been pushing customers toward a mix of AI agents and human experts.

The economics can become attractive very quickly.

If a professional service previously required an hour of employee time and AI reduces the human portion to 15 minutes, the same professional can theoretically supervise four times as much work. The real improvement will be smaller once we include review, exceptions and customer communication, but the direction is still powerful.

This should work best where customers care about outcomes more than the number of human hours they receive. Tax preparation, bookkeeping, legal work, insurance claims and some healthcare administration fit that pattern.

Human experts also solve a problem that fully autonomous AI still struggles with: accountability. Customers may be happy for software to handle 90% of a task as long as a qualified person can step in for the remaining 10%.

Businesses that control both sides of that workflow can gradually push more work toward AI without asking customers to accept a completely autonomous service overnight.

Are humanoid robots finally an example of a business that improves when AI improves?

Humanoid robotics is becoming a credible example, although the business case remains much less proven than enterprise software or cybersecurity.

BMW gives us unusually concrete evidence. Figure 02 worked for roughly 1,250 hours at BMW’s Spartanburg plant, moved more than 90,000 components and contributed to the production of more than 30,000 BMW X3 vehicles.

The job itself was narrow: picking and positioning sheet-metal parts for welding. That limitation is useful rather than embarrassing because it tells us where physical AI currently stands. These robots can already handle some repetitive industrial work, while broad human-level factory versatility remains unproven.

BMW has since expanded its physical-AI testing in Europe, including battery-assembly and component-manufacturing work.

The interesting economics come from software improvement. A factory that buys a conventional machine usually gets roughly the capabilities it purchased. A sufficiently general robot could gain additional tasks after deployment as its perception, reasoning and motor-control software improve.

That creates the possibility that already-deployed hardware becomes more productive over time.

We should still keep robotics below enterprise software, cybersecurity and infrastructure in confidence. Hardware failures, maintenance, cycle times, safety and manufacturing costs can wipe out gains from better models. BMW’s deployment proves that real factory work is possible; it does not yet prove broad commercial economics.

Which businesses only look like AI winners but may get weaker?

Generic AI wrappers and commodity digital services look the most vulnerable because better models improve their output while also making it easier for competitors to reproduce.

Consider a company charging $100 for a task that costs $20 in AI inference. If better models reduce the AI cost to $2, the company appears to gain $18 of extra margin.

But suppose competitors can make the same improvement immediately. The market price may fall from $100 to $30. Customers capture most of the benefit, while the provider keeps very little.

That pressure is particularly strong in generic copywriting, basic image creation, simple research, uncomplicated translation and undifferentiated software generation. Better AI makes these services cheaper to produce and cheaper to enter.

The model convergence we discussed earlier makes the problem worse. By itself, access to excellent intelligence is becoming less unusual.

As seen above, stronger categories such as ServiceNow, Salesforce, Glean or Thomson Reuters combine AI with assets that remain difficult to reproduce. Commodity AI services often lack that second layer.

Lower production costs are useful. They become a real business advantage only when competitors cannot force most of the savings back to customers.

Which businesses have the best “AI gets better” setup right now?

Enterprise workflow software, cybersecurity, AI infrastructure and businesses controlling scarce professional context currently have the best combination of stronger demand and defensibility.

Enterprise workflow platforms sit at the top because they already own the data, permissions and actions AI agents need. Smarter models let ServiceNow, Salesforce and Microsoft automate more work without forcing customers to rebuild the systems underneath them.

Cybersecurity and observability come next. More AI means more software, more machine identities, more autonomous activity and more complexity. CrowdStrike’s $5.84 billion ARR and 51% jump in quarterly net-new ARR, Datadog’s 36% revenue growth and Palo Alto Networks’ expanding AI-security products show that this is already translating into spending.

Physical infrastructure has an equally convincing demand story, although the economics are more capital intensive. The IEA expects AI-focused data-center electricity consumption to roughly triple by 2030, while Vertiv’s 252% order growth shows how abruptly that demand can reach suppliers far away from the model layer.

Professional AI built around difficult workflows also looks strong. Harvey, Thomson Reuters, Glean and Intuit show four slightly different ways to combine improving models with valuable context, distribution or expertise.

Robotics belongs in the promising rather than proven category. BMW’s deployment shows meaningful progress, but we still lack enough large-scale commercial deployments to rank humanoids beside enterprise software.

Generic AI wrappers land at the bottom. They benefit technologically from every model upgrade while becoming easier for competitors to imitate.

Business model	What happens when AI improves	Our confidence
Enterprise workflows and systems of record	Agents can complete more valuable work using existing data and permissions	Very high
Cybersecurity, identity and observability	More AI creates more systems, identities and attacks to manage	Very high
Power, cooling and data-center infrastructure	Higher AI usage increases physical infrastructure demand	Very high
Professional data + vertical workflows	Better reasoning makes scarce context more useful	High
AI-enabled expert services	Human labor required per job can fall	High
Developer platforms	More generated software increases surrounding development activity	High
Robotics	Better models can unlock new tasks on existing hardware	Medium
AI compute providers	Demand remains huge but capital needs are enormous	Medium-high
Deep vertical AI apps	Strong when embedded in real workflows	Medium-high
Generic AI wrappers	Product improves but competitors receive similar upgrades	Low

So what businesses actually get better as AI gets better?

The clearest winners are businesses that own something scarce around increasingly abundant intelligence.

Today, enterprise workflow software looks like the strongest example. Better models can enter Salesforce, ServiceNow, Microsoft Dynamics or Intuit and immediately gain access to years of customer data, permissions, integrations and established workflows. Each improvement in reasoning expands the amount of work those systems can potentially automate.

Cybersecurity and observability have another strong setup. AI adoption itself creates more software and more autonomous activity to secure and monitor. That means the same technological improvement threatening some software categories can directly increase demand for CrowdStrike, Palo Alto Networks, Datadog and similar platforms.

Physical infrastructure benefits for a different reason. AI may be virtual, but large-scale inference eventually requires chips, cooling, transformers, substations, electricity and land. Those constraints are getting tighter even as computing becomes more efficient.

Professional platforms such as Harvey, Thomson Reuters, Glean and Intuit can also improve quickly because stronger models let them extract more value from information, trust and customer relationships they already control.

Robotics could eventually become one of the purest versions of this idea. Software improvements could make an installed machine capable of additional work without requiring a completely new machine. BMW’s factory data makes that possibility much harder to dismiss today, although the commercial evidence is still early.

The easiest way to test any business is to imagine that next year’s best AI becomes twice as capable and substantially cheaper.

A strong business suddenly performs more valuable work for customers it already has. It still controls the data, actions, distribution, infrastructure or trust needed to deliver that work.

A weak business simply produces the same output more cheaply while dozens of competitors receive the exact same upgrade.

That distinction explains why some of the best AI businesses may turn out to be security companies, enterprise databases, legal-information providers, electrical-equipment manufacturers and robot operators.

They do not need to own the smartest model.

They need to own the place where smarter models become valuable.