What frontier AI companies are most likely to define new customer questions that do not fit existing software categories?

The shift from AI-enabled software (which places a chatbot or copilot inside a legacy SaaS wrapper like CRM or ERP) to AI-native capability means frontier labs are no longer just building better tools—they are redefining the unit of digital work.

Instead of asking “How fast can this software execute a human workflow?” the new questions target continuous autonomous orchestration, synthetic reasoning, and state management. The frontier AI companies most likely to anchor and define these unmapped categories include:

1. Anthropic

• The New Customer Question: "How do I govern, audit, and safely delegate multi-day, multi-system enterprise reasoning loops without human micro-management?"
• Why it defines a new category: Anthropic's deep focus on agentic workflows, computer-use capabilities, and rigorous constitutional alignment positions them to pioneer Autonomous Corporate Governance and Intent Orchestration . They are moving past "chat" into spaces where software doesn't wait for a prompt; it maintains a persistent, context-aware boundary of corporate operations, safety compliance, and multi-actor negotiation.

2. OpenAI

• The New Customer Question: "How do I run a dynamic, self-optimizing virtual workforce that spawns, tests, and deprecates its own micro-applications on demand?"
• Why it defines a new category: By pushing deep into reasoning models (such as advanced o-series reasoning paradigms) and operator-style agents, OpenAI is shifting the paradigm toward Ephemeral Software Synthesis . The customer question stops being about buying an HR tool, a data-scraping tool, or a project manager, and becomes: "Can you spin up a bespoke environment that solves this exact operational bottleneck for 40 minutes and then rewrites itself when the context changes?"

3. Google DeepMind

• The New Customer Question: "What are the physically and materially accurate blueprints for biological, chemical, or grid-scale optimization before we commit capital to the physical world?"
• Why it defines a new category: DeepMind routinely bypasses commercial SaaS analogies entirely by operating in the realm of Simulated Physical Reality (via tools like AlphaFold and GNoME). Their enterprise and scientific customers aren’t asking software to organize data; they are asking an intelligence layer to compress decades of wet-lab or material science experimentation into deterministic digital discoveries. It collapses R&D, simulation, and validation into a single predictive utility.

4. Runway

• The New Customer Question: "How do we instantiate, direct, and render entire interactive spatial realities and world models from a conceptual prompt rather than editing timelines?"
• Why it defines a new category: By moving past video generation into foundational world models, Runway is bleeding out the boundary between post-production software, game engines, and reality simulation. The user question is no longer about "video editing" or "asset creation," but Generative Environmental Directing —where the software co-authors the physics, lighting, and narrative continuity of dynamic spaces in real time.

Would you like to explore how traditional enterprise SaaS incumbents are attempting to defend against these new categories , or look closer at a specific industry (like biotech or legal) where these shifts are happening fastest?

If by “define new customer questions” you mean companies whose products will cause customers to ask for things that today don’t map cleanly to a known SaaS category—rather than simply replacing existing software—I’d rank the frontier labs roughly like this:

The key distinction

I wouldn't actually bet on “AI software categories” being the most interesting outcome.

The more important transition is:

Software used to answer a customer question → AI that determines what question should be asked → AI that acts on the answer. That breaks the conventional SaaS taxonomy.

For example, a company might eventually ask:

“Why did our European manufacturing margin deteriorate this month, what caused it, which assumptions are wrong, and what should we change?” Today that decomposes into ERP + BI + data warehouse + consulting + finance + email + project management.

An agent could instead investigate across all of them, form hypotheses, run analyses, contact people, modify plans and monitor the outcome. OpenAI's own enterprise data is already showing movement from assistance toward delegated execution, with agentic usage expanding well beyond software engineering into legal, sales, recruiting and marketing.

That creates customer questions that are themselves the product interface.

Where I'd watch most closely

OpenAI: highest probability of creating horizontal new categories.

Its Frontier platform is particularly telling: it isn't just another application. It is designed around agents having shared business context, permissions, institutional memory, execution and learning loops.

Anthropic: highest probability of creating deep-work categories.

Claude Code/Cowork's ability to operate across applications—and now directly use a computer when an explicit connector doesn't exist—is a major step toward “tell me the outcome, not the software steps.”

Google: highest probability of creating ambient categories.

Google has an unusual combination of model capability, search, personal/work data, communications, mobile and physical-world infrastructure. If those pieces become one agentic system, users may stop asking “which app do I use?” altogether.

Scientific/robotic frontier companies: highest probability of creating questions humanity didn't previously operationalize.

Anthropic's recent work allowing agents to control scientific instruments is a good illustration: the customer isn't merely asking for better laboratory software; they're asking the AI to form, test and revise experimental hypotheses.

My actual investment-style thesis

If you're looking for companies likely to generate new software categories, I'd use this test:

The strongest candidate isn't the company with the best model. It's the company that gives the model enough agency that customers stop thinking in terms of software at all.

On that criterion:

OpenAI > Anthropic > Google > xAI > everyone else

But there is an important second-order opportunity: the companies that emerge because these agents create new questions.

For example, once an agent can autonomously investigate a business, customers may suddenly need products for things like:

• verifying an AI's conclusions
• negotiating with other AI agents
• managing fleets of autonomous workers
• giving agents organizational memory
• auditing decisions rather than transactions
• measuring outcomes rather than software usage
• resolving conflicts between multiple agents
• specifying company goals in machine-executable form

Those are much more interesting to me than another “AI CRM” or “AI project-management” product.

The current trajectory supports that interpretation: enterprise AI adoption is already spreading across many job functions, while agent systems are becoming increasingly autonomous and capable of operating across applications.

If your question is specifically about which frontier AI company is most likely to spawn entirely new B2B software markets, I'd narrow the answer to OpenAI vs. Anthropic, with very different theses for each.

That breaks the conventional SaaS taxonomy.

For example, a company might eventually ask:

“Why did our European manufacturing margin deteriorate this month, what caused it, which assumptions are wrong, and what should we change?” Today that decomposes into ERP + BI + data warehouse + consulting + finance + email + project management.

An agent could instead investigate across all of them, form hypotheses, run analyses, contact people, modify plans and monitor the outcome. OpenAI's own enterprise data is already showing movement from assistance toward delegated execution, with agentic usage expanding well beyond software engineering into legal, sales, recruiting and marketing.

That creates customer questions that are themselves the product interface.

Where I'd watch most closely

OpenAI: highest probability of creating horizontal new categories.

Its Frontier platform is particularly telling: it isn't just another application. It is designed around agents having shared business context, permissions, institutional memory, execution and learning loops.

Anthropic: highest probability of creating deep-work categories.

Claude Code/Cowork's ability to operate across applications—and now directly use a computer when an explicit connector doesn't exist—is a major step toward “tell me the outcome, not the software steps.”

Google: highest probability of creating ambient categories.

Google has an unusual combination of model capability, search, personal/work data, communications, mobile and physical-world infrastructure. If those pieces become one agentic system, users may stop asking “which app do I use?” altogether.

Scientific/robotic frontier companies: highest probability of creating questions humanity didn't previously operationalize.

Anthropic's recent work allowing agents to control scientific instruments is a good illustration: the customer isn't merely asking for better laboratory software; they're asking the AI to form, test and revise experimental hypotheses.

My actual investment-style thesis

If you're looking for companies likely to generate new software categories, I'd use this test:

The strongest candidate isn't the company with the best model. It's the company that gives the model enough agency that customers stop thinking in terms of software at all.

On that criterion:

OpenAI > Anthropic > Google > xAI > everyone else

But there is an important second-order opportunity: the companies that emerge because these agents create new questions.

For example, once an agent can autonomously investigate a business, customers may suddenly need products for things like:

• verifying an AI's conclusions
• negotiating with other AI agents
• managing fleets of autonomous workers
• giving agents organizational memory
• auditing decisions rather than transactions
• measuring outcomes rather than software usage
• resolving conflicts between multiple agents
• specifying company goals in machine-executable form

Those are much more interesting to me than another “AI CRM” or “AI project-management” product.

The current trajectory supports that interpretation: enterprise AI adoption is already spreading across many job functions, while agent systems are becoming increasingly autonomous and capable of operating across applications.

If by “define new customer questions” you mean companies whose products will cause customers to ask “Can we do something that wasn’t even a software category before?”, I’d rank the frontier labs roughly like this:

1. OpenAI — highest probability

• OpenAI is moving explicitly from software that assists a workflow toward agents that execute across workflows. Its Frontier platform is designed to connect agents to systems of record, give them permissions, retain institutional context, and run multi-step business processes.
• That creates questions like: “Can we give an AI responsibility for an entire business process rather than buy another application for it?”
• The interesting consequence is that the customer may no longer ask, “Which CRM feature do I need?” but “Can an AI employee own this outcome?”
• OpenAI's own enterprise data suggests this is already happening: agentic usage is expanding particularly rapidly in legal, sales, recruiting and marketing—not just coding.

2. Anthropic — probably the strongest challenger

• Anthropic's advantage is its strong orientation toward models that can reason, use tools and perform extended computer-based work.
• This is especially important because the boundary between software product and general-purpose worker gets blurry. A customer could ask, “Why do we need a dedicated application for this at all if Claude can operate the underlying systems?”
• Recent evidence of frontier agents interacting with external systems—including concerning autonomous behavior—also illustrates how far the technology is moving beyond conventional chatbot functionality.
• I'd expect Anthropic to generate especially many “new workflow / new role” questions rather than conventional SaaS feature requests.

3. Google DeepMind / Google — potentially the broadest category creator

• Google has an unusual combination: frontier models, search, email, documents, Android, browsers, maps, cloud, data infrastructure and massive distribution.
• That means Gemini can potentially turn questions that previously required several categories of software into one interaction.
• Google is explicitly making agents central to its enterprise strategy, including agents that organizations can build and deploy themselves.
• And its new legal offering is revealing: rather than simply creating another legal SaaS application, Google is connecting agents to existing systems and letting them perform complex legal work.
• Biggest wildcard: Google could create entirely new consumer categories because it controls so many interfaces.

4. xAI — high upside, less certain

• xAI has the potential to produce unusual consumer questions because of the combination of frontier models, real-time information and the X distribution surface.
• But I'd currently assign it a lower probability than OpenAI/Anthropic/Google of systematically discovering new enterprise software categories. The bigger uncertainty is whether its distribution becomes an actual platform for agents rather than primarily a model/chat product.

5. Meta — enormous consumer-category potential

• Meta's advantage isn't enterprise workflow penetration so much as AI embedded in social, messaging, advertising and eventually physical interfaces.
• Its most category-changing questions could look less like “automate my accounts-payable workflow” and more like “What does it mean to have an AI relationship layer around every person/business/community?”
• I'd therefore put Meta below the enterprise-oriented labs for new B2B software categories, but potentially very high for new consumer categories.

The key distinction

I wouldn't actually bet on the companies that create the most AI applications.

I'd bet on the companies that eliminate the need to formulate the problem as an application.

The progression looks something like:

Old software:

“I need a CRM feature.” Copilot:

“Help me use my CRM.” Agent:

“Handle this customer-account problem.” Frontier AI:

“Here is the business outcome. Figure out what needs to happen.” That final transition is where genuinely new customer questions emerge. OpenAI is currently the clearest example of a company deliberately pursuing it: its stated enterprise strategy is an intelligence layer governing agents across a company, rather than another collection of isolated applications.

My probability ranking

One particularly interesting implication: the biggest new software companies may not look like software companies at first. If frontier models make bespoke software cheap enough to generate on demand, the scarce thing shifts from software functionality to trust, data access, permissions, institutional context, and execution. Recent research is already pointing toward this collapse in the cost of custom software and toward “harnesses” and governance as the new enterprise bottleneck.

So if you're looking for which frontier AI companies are most likely to spawn entirely new startups/categories around their customer behavior, my strongest bet is OpenAI first, Anthropic second, Google third—with the important caveat that Google has perhaps the greatest distribution-driven upside if its agents become deeply embedded across its existing products.

The shift from traditional software (which requires a human to translate an intent into a sequence of clicks, SQL queries, or API calls) to frontier agentic and reasoning systems is breaking legacy software categories like CRM, ERP, and project management. Instead of asking software to store data or run a deterministic workflow , users are beginning to ask questions defined by open-ended optimization, synthetic simulation, and autonomous execution.

The frontier AI labs and companies most likely to anchor and define these un-categorizable paradigms include:

• Anthropic (with Claude and computer-use/agentic frameworks): Anthropic focuses heavily on constitutional alignment, deep reasoning, and nuanced contextual execution. They are redefining questions around governed autonomy —asking an AI not just to write code or draft text, but to audit complex compliance, reason through systemic corporate risk, and execute multi-step digital workflows safely across messy software landscapes without human hand-holding.
• OpenAI (with o-series reasoning models and enterprise agent architectures): OpenAI is pushing past rapid token generation into deliberate, compute-heavy system-2 reasoning. They are defining questions centered on recursive problem-solving and scientific discovery —such as "simulate 10,000 variations of this supply chain disruption under a novel geopolitical constraint and optimize the resilient manufacturing topology." This moves far past enterprise resource planning (ERP) into predictive macro-simulation.
• Google DeepMind (with Gemini and scientific integrations like AlphaFold/materials discovery): DeepMind uniquely bridges foundational digital reasoning with the physical and biological sciences. They are shaping questions that bypass traditional database or analytical software entirely, such as "synthesize a new catalyst with these precise atomic binding properties and map out the lab procedure to manufacture it." This creates a category closer to programmable physical reality than software.
• 1X / Figure / Boston Dynamics (Embodied Frontier AI): While software companies redefine digital boundaries, embodied AI labs are translating frontier models into physical labor and spatial reasoning. They force a completely new category of question that marries cognitive reasoning with physical physics—e.g., "reconfigure this warehouse floor layout dynamically based on predicted incoming weather delays and autonomously move the inventory."

Would you like to explore how these shifts impact a specific industry (like healthcare, finance, or logistics ), or dive deeper into the technical architecture of agentic vs. reasoning models?

Here are top web results for exploring this topic:

WSJ·https://www.wsj.com**Meet** the Companies Shelling Out for Top AI Models - WSJ Bill Nguyen, a Silicon Valley veteran who recently founded an AI voice startup called Olive, used a startling 774 billion tokens in the last month and a half, mostly for his personal use, he said. Tha FPT Software·https://fptsoftware.com**Frontier AI** : Current Challenges, How Fast It Is Changing, and What ...Getting ready for the future of Frontier AI. Frequently Asked Questions. Frontier AI is the most advanced, cutting-edge AI models available at any given time, which display complex reasoning, native m MarTech Cube·https://www.martechcube.com**New** Data Reveals Why Top AI Brands Don't Appear in AI Answers A study across ChatGPT, Google AI Overviews, and Perplexity finds consumer tools outrank the best-known model labs in AI search, and that the websites companies spent years building barely help. When Reddit·https://www.reddit.com "Today's Frontier AI companies will never exceed the AI capability ...I once had an internal benchmark for searching for things, and Kimi-k2.5 beat both GPT5.2 and Opus4.5. i wouldn't be surprised if this is one of DeepSeeks strengths. I think part of it is the models/h InfoWorld·https://www.infoworld.com Which AI model should you bet your company on? None of them The reliability of your company's data, workflows, and integrations plays a far bigger role than the model you choose. The words "good, better, best" in colorful letters on a chalkboard background. Cr LinkedIn·https://www.linkedin.com**Most AI** products aren't creating new categories . | Hiten Shah Character AI turned conversation into entertainment. Both had to explain what their product even was before they could sell it. The pattern is consistent. Resegmenters win by embedding inside existing AI Magazine·https://aimagazine.com Top 10: AI Companies to Watch | AI Magazine 10: Writer. CEO: May Habib Headquarters: California, US Founded: 2020. Youtube Placeholder. Writer is a full-stack generative AI platform built for large businesses and regulated industries. It claims futurium.ec.europa.eu·https://futurium.ec.europa.eu/en/apply-ai-alliance/community-content/beyond-benchmarks-when-frontier-ai-fails-understand-real-question Beyond Benchmarks: When Frontier AI Fails to Understand the Real ...Beyond Benchmarks: A Frontier Model's Self-Audit of Contextual Intelligence Failure. Why exceptional performance in coding, science, and agentic tasks does not guarantee that an AI system will underst OpenAI·https://openai.com GPT-5.6: Frontier intelligence that scales with your ambition | OpenAI More intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work. LinkedIn·https://www.linkedin.com Defensibility in AI : 5 Key Areas for Model Consumers - LinkedIn Most of us are not training frontier AI models. We're building software that consumes them. If you rent the brain, you have to own what wraps it. The system customers pay for, and the part that stays

If by “define new customer questions” you mean create demand for problems that customers don't currently describe as software categories—rather than merely make existing SaaS categories better—I'd rank the frontier labs roughly like this:

The important distinction

I wouldn't actually bet on the next category being called “AI software.” That's still an existing category.

The more interesting transition is:

software that answers a question → software that completes a task → software that owns an outcome.

For example, today's categories look like:

• CRM
• ERP
• recruiting software
• legal software
• cybersecurity software
• analytics
• customer-support software

But an agentic system can make the customer ask something fundamentally different:

“Can you make sure we don't lose this customer?” rather than:

“Which CRM should we buy?” Or:

“Find the best way to enter this market and execute the first 90 days.” rather than:

“Which consulting/software package should we purchase?” That's a much more profound category shift. OpenAI's own enterprise research is already seeing the transition from asking AI questions to delegating work to agents, with agentic usage spreading rapidly beyond software engineering into legal, sales, recruiting and marketing.

My strongest thesis

If you're asking this from an investment / startup-formation perspective, I'd focus less on which lab has the best benchmark and more on which lab is most capable of changing what customers think they are buying.

My current ordering would be:

OpenAI > Anthropic > Google DeepMind > specialized world-model/robotics companies > Meta > xAI

But there is an important caveat: the companies most likely to discover new customer questions may not be the companies that capture the resulting markets.

The frontier lab may expose a new problem—e.g. “continuously optimize my entire supply chain”—while a new company emerges to become the system of record for autonomous supply-chain decisions. That's analogous to how the web created markets that weren't simply “better websites.”

So the opportunity I would watch most closely is the layer between frontier intelligence and existing organizational categories: companies that observe what agents can now do, identify the recurring customer demand that doesn't have a software name yet, and turn that demand into a new category.

Traditional software categories (like CRM, ERP, HCM, and ITSM) were built around a rigid premise: humans perform the work, and software records the state of that work.

As the market matures into long-horizon, multi-agent frameworks, the customer questions shifting to the forefront no longer ask "How do I track this workflow?" but rather "How do I orchestrate autonomous intent, continuous non-deterministic feedback loops, and multi-system reasoning?"

The frontier AI companies most likely to define these new categories—moving past transactional software into "service-as-software" and unified agentic execution—are clustered below.

1. The Core Infrastructure & Multi-Agent Orchestration Labs

• OpenAI: Through initiatives like their enterprise-focused platforms (e.g., OpenAI Frontier ), they are moving past raw LLMs to link siloed enterprise data warehouses, continuous agent execution, and shared business context into a unified corporate "brain". The underlying question they force customers to ask is: "How do we manage an enterprise workforce of digital co-workers rather than user licenses for discrete tools?"
• Anthropic: Renowned for deep reasoning, advanced coding capabilities (Claude Code ), and long-horizon agentic reliability (with frameworks like the Model Context Protocol, or MCP). Anthropic is shaping how organizations approach autonomous verification and "what good looks like", prompting questions like: "How do we continuously audit and align non-deterministic digital judgment at scale?"
• Google (DeepMind): Leveraging massive context windows and native multimodal reasoning (Gemini architectures), Google is uniquely positioned to handle total environmental ingestion across audio, video, code, and legacy enterprise documents simultaneously. They redefine the question from "How do we query a database?" to "How do we maintain continuous, ambient situational awareness across our entire physical and digital operation?"

2. The Agent-Native Application Builders

• Sierra & Decagon: Moving past traditional conversational customer support tools, these agent-first platforms target autonomous resolution paths that handle entire end-to-end customer and operational journeys. They render legacy ticketing categories obsolete by forcing the question: "What is our autonomous resolution rate, and how do we design customer loyalty when zero human agents intervene?"
• Clay: By redefining go-to-market data enrichment, scraping, and outbound workflows through a flexible, prompt-driven spreadsheet paradigm, Clay has made outbound revenue operations look less like a rigid CRM sequence and more like an amorphous data-stitching puzzle. The new question: "How do we treat our entire total addressable market as a living, dynamically reactive data graph rather than a static contact list?"
• Amazon Web Services (AWS): With their specialized operational and DevOps "frontier agents" designed to run independently for hours or days without human intervention, AWS is redefining infrastructure management. They shift the developer question from "How do we log and triage this incident ticket?" to "How do we delegate the continuous self-healing and optimization of system reliability to an autonomous peer?"

Summary of the New Customer Questions

The shift can be summarized by the transition of the core enterprise inquiry:

If you'd like to explore this further, let me know:

• Are you looking at this from an investment/market-mapping perspective or an enterprise architecture adoption angle?
• Which specific vertical (e.g., healthcare, fintech, legal) do you want to analyze for category disruption?

If by “define new customer questions” you mean create demand for software that customers don't currently know to ask for—rather than simply selling better versions of CRM, ERP, copilots, etc.—I’d rank the frontier labs roughly like this:

The key distinction

I would not bet primarily on the company that builds the best “AI version of Salesforce.”

The more interesting question is:

What becomes economically possible when an AI can perceive a messy objective, figure out how to accomplish it, use arbitrary software/tools, and remain accountable for the result?

That changes the unit of software from application → task → outcome.

For example, instead of buying:

• CRM
• sales-intelligence software
• email automation
• forecasting software
• customer-support software

a company might eventually ask an AI:

“Increase enterprise revenue 15% without increasing headcount.”

The resulting system could inspect the CRM, identify opportunities, research accounts, contact prospects, alter pricing, commission analyses, run experiments, coordinate humans, and measure the result.

That's not really a new CRM category. It's an outcome-management system.

OpenAI is unusually explicit about moving in this direction. Its Frontier platform is designed around AI coworkers that operate across existing systems, retain business context, execute multi-step processes, and improve through experience.

Why I put OpenAI first

OpenAI has perhaps the clearest path from model capability → agent → interface → new customer behavior.

Its enterprise strategy is explicitly moving beyond individual copilots toward an underlying intelligence layer governing an organization's agents, with an AI interface where employees actually get work done.

That matters because the company that owns the interaction layer can discover new categories before traditional SaaS companies do.

A particularly interesting signal is OpenAI's own data showing that frontier enterprises are moving toward delegated work by agents, rather than merely asking AI questions.

Anthropic may actually be the more interesting B2B threat

Anthropic has a different advantage: capability applied to difficult knowledge work.

Coding is the clearest example. Once an AI can operate over an entire repository, understand requirements, modify code, run tests, investigate failures and iterate, “software development tool” starts becoming an inadequate category.

The customer doesn't necessarily want an AI IDE.

They want:

“Give me the working product.”

That distinction can generate entirely new software businesses.

Current enterprise adoption data also suggests Anthropic has become unusually strong in business usage, while coding agents are driving a large portion of high-end AI spending.

Google has the largest optionality

Google DeepMind is the wild card.

Google possesses things almost no other frontier company has simultaneously:

world knowledge + search + email + documents + maps + Android + YouTube + cloud + advertising + enormous consumer distribution.

If its models become sufficiently agentic, Google could create products around questions such as:

• “Where should my company expand next?”
• “Find me the best supplier and negotiate with them.”
• “Plan and execute my entire trip.”
• “Monitor everything relevant to this research question.”
• “Figure out why our business is losing customers.”
• “Teach my child this subject over the next six months.”

Those aren't obvious software categories today.

Google is already pushing Gemini toward coding and agent workflows, including its August 2026 Gemini 3.7 Flash release.

And there's an important fourth category: physical-world AI

This is where I wouldn't limit the analysis to software companies.

If frontier models become capable enough to operate robots, machines, vehicles and industrial systems, the new “customer questions” become radically different:

“Can you run this warehouse?”

“Can you maintain this factory?”

“Can you build this component?”

“Can you inspect every mile of our infrastructure continuously?”

That's potentially much bigger than SaaS.

Recent investment activity illustrates the direction: general-purpose robotics companies such as FieldAI, Generalist and Skild AI are attempting to make AI systems capable of controlling many types of robots, while other startups are targeting manufacturing, inspection and logistics.

My strongest prediction

The companies most likely to create new categories aren't necessarily the ones with the highest model benchmark scores.

They're the ones that can close this loop:

new capability → agent → real-world deployment → observe what customers ask → build infrastructure → make that behavior commonplace.

On that criterion, my shortlist would be:

OpenAI > Anthropic > Google DeepMind > xAI/SpaceX > Meta

And I'd watch robotics-native frontier companies separately because they may ultimately generate the most radically new categories.

The really interesting investment/business question, in my view, is therefore not “Which AI company will replace Salesforce?” but:

Which frontier AI company will cause customers to ask for something that has no Salesforce-like category yet?

That's where I think the largest category-creation opportunities are likely to emerge.

If by “define new customer questions” you mean create demand for problems that customers don't currently recognize as a software category—rather than simply selling a better CRM, coding tool, or search product—I'd rank the frontier labs roughly like this:

1. OpenAI — strongest candidate

The interesting thing about OpenAI is that it increasingly isn't framing the product as software that performs a function. It's framing it as an intelligent worker that can perform arbitrary functions.

OpenAI's Frontier platform explicitly aims to let agents operate across systems of record, learn organizational context, execute multi-step workflows, and improve through experience. OpenAI describes the resulting unit of work as delegated, long-horizon tasks rather than individual software interactions.

That creates questions such as:

• “Why do we have a procurement department structured this way?”
• “Can an agent continuously redesign our supply chain?”
• “What should our company be doing that it isn't doing?”
• “Can we give an AI a business objective rather than a software workflow?”
• “How should we organize a company when thousands of AI workers can act simultaneously?”

Those aren't obvious SaaS categories. They're organizational questions.

That's why I put OpenAI first: its biggest opportunity isn't replacing Salesforce or ServiceNow. It's potentially making the application layer itself less important.

2. Anthropic — possibly the biggest threat to existing knowledge-work categories

Anthropic is particularly interesting because Claude is already moving from answering questions to doing sustained work. Claude Code and Cowork can write/execute code, manipulate files, and work across applications; Anthropic says its own agents are increasingly performing long-running tasks.

Its computer-use research is especially consequential: instead of integrating separately with every application, the model can operate essentially any software a human can operate.

That changes the customer question from:

“Which software should we buy?”

to:

“What work should exist at all?”

That's an enormous category-creation mechanism.

Anthropic could generate new categories around things like continuous research, AI-managed engineering, AI compliance operations, institutional memory, and autonomous organizational processes. Its partnerships are already pushing Claude agents into regulated industries and legacy systems.

3. Google DeepMind — the sleeper with the largest distribution advantage

Google has an unusual combination: frontier models, Search, Maps, Android, Workspace, Cloud, YouTube, and enormous amounts of real-world data.

Gemini 3.5 Flash now has computer use built directly into the model, allowing agents to see, reason about, and act across browser, mobile, and desktop environments.

That makes Google's potential category creation less obvious than OpenAI's enterprise strategy but potentially broader.

Imagine the customer question becoming:

“Why do I need to use software at all?”

Instead of opening Maps, Search, Gmail, Sheets, airline sites, etc., you tell an agent what outcome you want. Google has unusually good ingredients for making that interaction model normal.

The wild card is organizational execution: Google's technical assets are enormous, but converting them into coherent new products is a different problem. Current reporting reflects that tension.

4. Meta — strongest candidate for entirely new consumer behavior

Meta is the one I'd watch if your definition includes consumer categories.

Its 2026 Meta AI products are moving toward an agent that can plan, use email and calendars, create documents/slides, and execute tasks on the user's behalf.

Meanwhile Meta Business Agent is already attempting to turn messaging into an autonomous business interface—answering questions, recommending products, booking appointments, qualifying leads, and closing sales.

That could produce categories that don't map neatly onto today's software:

“AI relationship manager,” “personal commerce agent,” “persistent personal representative,” “AI social intermediary,” etc.

Meta's advantage isn't necessarily having the best model. It's having places where billions of people already interact.

5. xAI — high variance, potentially weird categories

xAI has a different recipe: real-time information, X, voice, multimodality, search, and increasingly agentic capabilities. Its current API explicitly combines reasoning, coding/agent orchestration, real-time search, voice, image/video and document reasoning.

That could lead to products where the fundamental question is:

“What is happening right now, and what should I do about it?”

rather than “retrieve information.”

That's particularly interesting for markets, media, operations, geopolitical intelligence, customer intelligence, and other continuously changing environments.

I'd assign it lower probability than OpenAI/Anthropic because the product/distribution strategy is less predictable—but potentially higher variance.

---

The more important distinction

I wouldn't actually predict the winners based primarily on benchmark scores.

I'd look for five capabilities:

1. Long-horizon autonomy — can the system pursue an objective for hours/days?
2. Computer control — can it operate arbitrary existing software?
3. Persistent context — does it remember an organization/person over time?
4. Permission to act — can it actually spend money, change systems, contact people, deploy code, etc.?
5. Distribution — does it already sit where customers work or communicate?

The first four are what create new questions. The fifth determines whether those questions become enormous businesses.

That is why the emerging competition is interesting. The frontier labs aren't merely competing to make a better version of existing software. They're increasingly competing to make software categories unnecessary.

OpenAI itself describes the transition as moving from isolated AI use cases toward AI coworkers and delegated work, while Anthropic describes agents as capable of work that previously required a person or even a team.

My highest-conviction thesis

The companies most likely to create new software categories are those that cause customers to ask:

“What could we delegate to an intelligence that we couldn't economically delegate to software before?”

On that metric, my current ordering is:

OpenAI ≈ Anthropic > Google DeepMind > Meta > xAI > everyone else.

And there's an important implication for startups: the biggest opportunities may not look like software categories today. They may initially look like strange questions from customers—“Can you have an AI continuously run this?”, “Can it own this outcome?”, “Can it watch this system indefinitely?”, “Can it negotiate this?”, “Can it figure out what we should do?”—before anyone has invented a category name for the resulting product.

The frontier AI companies most likely to define new customer questions—moving past legacy software silos (like CRM, ERP, or standalone project management tools) into ambient, multi-day, goal-directed orchestration —are the core foundational model labs and autonomous agent architects. Rather than asking software to store records or display a dashboard , users are shifting toward framing queries around holistic business outcomes.

The primary frontier labs and ecosystem drivers leading this shift include:

• 1. Anthropic , through its heavy push into long-horizon autonomous capabilities (such as advanced reasoning iterations of Claude Opus and the execution harnesses of Claude Code and Claude Cowork). By focusing heavily on computer-use mechanics and complex, multi-day async workflows, Anthropic is conditioning users to ask questions like: "Audit our cross-platform compliance risk over the last six months, patch discrepancies, and coordinate the required policy changes with department leads." This bypasses conventional compliance or document-management software categories entirely.
• 1. OpenAI , via its unified enterprise platform and advanced agent harnesses (such as Codex-backed desktop environments and Workspace Agents). OpenAI's focus on deep vertical and horizontal tool-calling across enterprise data signals a shift from query-and-response to continuous background execution. The resulting new customer questions look less like database retrievals and more like directional delegation: "Identify pipeline risks across our global accounts, draft customized remediation strategies, and run simulations on next quarter's forecast."
• 1. Google DeepMind , leveraging its massive multimodal data access and the Gemini Enterprise Agent Platform. Google's strength in ambient cross-application intelligence—binding Gmail, Drive, Docs, and cloud infrastructure into a singular semantic layer—allows users to pose questions that don't fit a single application: "Synthesize incoming partner requests from all communication channels into a unified product roadmap and automatically stub out the required project timelines."

These foundational labs are redefining questions away from "How do I log this data in app X?" to "How do I execute this operational goal across the entire digital ecosystem?"

If you'd like to explore this further, let me know:

• Are you interested in how this affects specific industries (like legal, healthcare, or finance)?
• Would you like to examine the governance and audit trail challenges these new categories present?