1. AI Spending Is Moving Up the Stack

The earliest commercial winners of the generative AI boom were concentrated near the infrastructure layer.

Money flowed toward:

Cloud computing Graphics processing units Foundation models Data infrastructure Model training systems Vector databases Developer APIs Security and observability tools This was predictable. Before companies could build AI applications, they needed models, compute, storage, and development infrastructure. But infrastructure does not represent the final destination of most technology spending. Historically, the largest long-term software markets emerged at the application layer. Businesses ultimately pay for outcomes such as managing customers, processing payments, creating documents, designing products, hiring employees, operating supply chains, and closing sales. The application layer translates technological capability into useful work.

The Mercury spending data suggests that this transition is accelerating. Startups are no longer buying AI only as an experimental technical component. They are purchasing completed applications that fit into everyday activities. These applications write, summarize, design, code, transcribe, research, recruit, sell, support customers, manage compliance, and perform professional tasks. This shift matters because application businesses can capture value differently from infrastructure businesses. An infrastructure provider may charge for tokens, computing time, storage, or API calls. An application company can charge based on: Users Projects Documents Completed tasks Business outcomes Leads generated Cases resolved Revenue influenced

Hours of work replaced Entire workflows completed The closer a product sits to a measurable business outcome, the more pricing power it may eventually possess.

2. Horizontal AI Applications Hold the Early Advantage

The report divided applications into two broad groups. Horizontal AI applications can be used across departments, roles, and industries. Vertical applications focus on a specific function, profession, or market. Approximately 60 percent of the ranked companies were horizontal and 40 percent were vertical. That horizontal majority reflects the current stage of AI adoption. When a new computing platform emerges, broad tools often spread quickly because almost everyone can experiment with them. General assistants, writing tools, meeting products, search systems, and creative platforms require relatively little organizational change. An employee can begin using them immediately. A horizontal AI assistant can help a founder prepare an investor update, help a marketer draft a campaign, help a developer understand code, help a recruiter prepare interview questions, and help an operations manager summarize documents. The same core product may deliver value across dozens of use cases. This produces several advantages. Large addressable market A horizontal product is not restricted to one profession. It may sell to consumers, freelancers, startups, enterprises, educational institutions, and governments. Fast product-led adoption

Users can sign up without lengthy implementation. They can experience value before asking their company to purchase a larger plan. Cross-functional expansion Once one department succeeds with the tool, neighboring teams can adopt it. Frequent usage General assistants can become daily work interfaces rather than occasional utilities. Data and workflow accumulation As users connect files, messages, projects, and company knowledge, the assistant can become more deeply embedded in the organization. However, horizontal markets are also extremely competitive. General assistants face limited differentiation when multiple products can summarize documents, answer questions, generate text, write code, and analyze information. The report’s ranking included OpenAI, Anthropic, Perplexity, Merlin AI, Notion, Manus, and other products that overlap in certain use cases. This suggests that the market may not quickly become a simple winner-takes-all category.

Companies may use different systems for different purposes:

One model for coding Another for long-form analysis Another for research Another for document collaboration Another for private company data Another for lower-cost automation The future enterprise may operate with a portfolio of AI models and applications rather than one universal provider. This is similar to how businesses use several cloud platforms, communication tools, databases, and productivity products simultaneously.

3. General AI Assistants Are Becoming a New Workplace Layer

OpenAI and Anthropic occupied the top positions in the report, while Perplexity and other assistant products also appeared in the ranking. This is significant because general assistants are evolving beyond standalone chatbots. They are becoming an interface layer between employees and company information. Traditional business software requires users to navigate menus, fields, filters, dashboards, and predefined workflows. AI assistants allow users to express goals through natural language.

Instead of learning exactly where information is stored, an employee can ask:

Which customers are at risk of leaving? Summarize our last five product meetings. Compare these vendor contracts. Draft a response to this customer complaint. Build a financial model using these assumptions. Identify the most important issues in this codebase. Prepare a sales briefing for tomorrow’s meeting. This changes the role of the user interface. The interface becomes less about navigating software and more about communicating intent. OpenAI reported in its 2025 enterprise study that more than one million business customers were using its tools and that organizational usage was moving toward deeper workflow integration rather than isolated experimentation. The strategic opportunity is therefore larger than building a better chat window.

The winning workplace assistant may need to combine:

Model access Company knowledge Search Memory Permissions Workflow execution Data analysis Document generation Collaboration Security Auditability Human approval systems

The assistant becomes a control layer across many business applications. For startups, this creates both an opportunity and a threat. A specialized AI product may deliver greater domain-specific value. But it must also consider whether a general assistant could eventually absorb its most basic features. Defensibility may require proprietary workflows, integrations, data, trust, distribution, regulation, or outcome ownership.

4. Creative AI Has Become a Major Business Software Category

Creative tools represented the largest individual category in the ranking, with approximately ten companies appearing among the top 50. These included platforms related to images, audio, video, editing, avatars, and advertising.

The category included names such as:

Freepik ElevenLabs Canva Photoroom Midjourney Descript Opus Clip CapCut Arcads Tavus The importance of creative AI goes beyond the design industry. Historically, high-quality creative production required specialized people, expensive software, agencies, equipment, and significant time.

A startup preparing a marketing campaign might need:

A copywriter A graphic designer A video editor A voice actor A photographer An animation specialist An advertising agency A localization team AI does not eliminate the value of all these professionals. It does, however, reduce the cost and expertise required for many common outputs. A founder can create a product mockup. A salesperson can generate a personalized presentation. A recruiter can produce an employer-branding video.

A customer-support team can localize tutorials. A small e-commerce company can generate product photographs. A marketing manager can transform one webinar into short videos, social posts, summaries, and advertisements. This is why creative software is becoming horizontal. The category is no longer limited to people with “designer” or “video editor” in their job title. Almost every employee communicates. Almost every department creates documents, presentations, visuals, recordings, or customer-facing material. AI transforms creativity from a specialized production department into a general workplace capability. The economic consequences This expansion may produce several second-order effects. Content volume will increase dramatically When production becomes cheaper, organizations create more variations, campaigns, formats, and experiments. Personalization becomes affordable

Companies can adapt creative material to industries, customer segments, countries, languages, and individual accounts. Creative testing becomes continuous Instead of producing one expensive campaign, companies can test dozens or hundreds of variations. Small companies gain agency-like capabilities A five-person startup can create output that once required a larger marketing operation. Human creative work moves upward Professionals may spend less time resizing, transcribing, removing backgrounds, preparing basic edits, or generating first drafts. More time may shift toward strategy, taste, narrative, direction, and quality control. The winners will not necessarily be products that generate the most content. They may be products that help companies create the right content, maintain brand consistency, manage approvals, measure performance, and connect creative production directly to business results.

5. Vibe Coding Has Become a Business Purchasing Category

One of the report’s strongest signals concerned AI-assisted software development. Replit, Cursor, Lovable, and Emergent appeared among the leading AI applications purchased by startups. Replit ranked behind only OpenAI and Anthropic in the reported list. The phrase “vibe coding” is often used to describe building software through natural-language instructions, rapid experimentation, and AI-generated code rather than writing every component manually. The label can sound casual, but the spending data shows that the underlying behavior has reached workplaces. AI coding platforms now serve several audiences. Professional developers

Developers use AI to:

Generate boilerplate code Explain unfamiliar systems Debug errors Write tests Refactor applications Create documentation Review pull requests Explore implementation options Accelerate repetitive programming tasks Technical founders Founders can build and test products with smaller engineering teams. Designers and product managers

Non-engineers can create prototypes, internal tools, and interactive demonstrations. Operations teams Employees can build simple automations, dashboards, forms, and workflow applications. Entrepreneurs without traditional technical backgrounds People can move from an idea to a functioning product without first hiring a complete software team. Anthropic’s research has consistently found that software development and technical tasks represent a major share of AI usage. Its Economic Index has also documented an increase in more directive automation, where users delegate larger portions of tasks to AI systems. The a16z report highlights an important difference between consumer popularity and enterprise spending. Lovable ranked strongly in consumer web traffic, but Replit reportedly generated approximately 15 times more revenue from the Mercury customers studied. The likely explanation is not simply brand strength. Businesses pay for complete operational capability. A visually impressive prototype may attract users, but a company needs more than interface generation. It may require: Databases

Authentication Hosting Security Deployment Version control Team collaboration Access management Monitoring Integration Enterprise governance Replit’s broader application-development environment may have aligned more closely with these business requirements.

The lesson for founders is clear:

Generating an output is valuable. Owning the complete path from idea to deployed business result is more valuable.

6. Vertical AI Is Dividing Into Augmentors and Substitutes

Vertical AI products focus on specific roles or workflows.

The report identified applications in areas including:

Customer service Sales Go-to-market operations Recruiting Human resources Compliance Accounting Legal services Information technology The authors separated vertical applications into two types. Augmentors These products help human employees perform their roles more efficiently.

Examples include tools that:

Summarize recruiting interviews Prepare sales research Draft customer responses Automate compliance evidence Organize accounting information Recommend next actions Reduce administrative work Substitutes These products attempt to complete an entire workflow that might otherwise be assigned to an employee, contractor, agency, or professional-services firm.

The report identified five companies in this more autonomous category:

Crosby Legal Cognition 11x Serval Alma These companies were associated with legal work, engineering, go-to-market execution, IT service operations, and immigration services. Most vertical products in the ranking were still augmentors. Twelve of the seventeen vertical applications identified in the report primarily enhanced human workers, while five were positioned more like AI employees or end-to-end service providers. This distinction may define the next phase of AI software. Traditional software helps a person manage work. AI-native software can increasingly perform the work.

Consider the difference:

Traditional sales software stores leads and activities. An AI sales system may identify prospects, research them, write messages, send outreach, respond to replies, update records, schedule meetings, and recommend follow-up actions. Traditional accounting software records financial activity. An AI accounting service may classify transactions, reconcile accounts, identify anomalies, prepare reports, request missing documents, and escalate exceptions. Traditional legal software organizes cases. An AI legal service may gather information, draft documents, prepare filings, communicate with clients, and manage a standardized process from beginning to end. This changes what the customer is buying. The customer is no longer paying only for a tool. The customer may be paying for a completed outcome.

7. AI-Native Services Could Become Larger Than AI Software

One of the most important implications of vertical AI is the emergence of AI-native service businesses.

Professional services represent enormous markets:

Legal services Accounting Consulting Recruiting Marketing Customer support IT management Insurance administration Financial analysis Compliance Real estate operations Healthcare administration

Traditional software companies sell tools to these industries. AI-native service companies can compete directly with the service providers themselves. This expands the addressable market. A legal software company may compete for part of a law firm’s technology budget. An AI legal service may compete for the client’s entire legal-services budget. An accounting application may charge $100 per month. An AI-operated accounting service may charge thousands of dollars annually while still costing less than a traditional provider. This is sometimes described as the movement from selling software to selling work. The economics can be attractive. A conventional service company often scales by hiring more people. Revenue and headcount rise together. An AI-native service company may increase revenue faster than headcount if software completes a growing portion of delivery.

However, these businesses face challenges that pure software products can avoid:

Professional liability Regulation Licensing Quality assurance Customer trust Exception handling Human escalation Data security Industry-specific accuracy Responsibility for outcomes The strongest model may therefore be hybrid. AI handles standardized, repetitive, and high-volume work.

Human specialists manage unusual, sensitive, regulated, or high-stakes cases. Over time, the percentage handled autonomously may increase.

8. Consumer AI Products Are Entering Companies From the Bottom Up

Nearly 70 percent of the companies on the a16z and Mercury list could be adopted by individuals without requiring an enterprise contract. Twelve of the ranked companies also appeared in a16z’s consumer AI rankings. This signals a major change in enterprise software distribution. Historically, many business software companies sold from the top down.

The process often involved:

Executive approval Procurement review Security assessment Legal negotiation Annual contract Implementation Employee training AI products frequently spread in the opposite direction. An individual discovers the product. The person pays for a personal subscription. The product becomes part of daily work. Colleagues notice the results.

A small team adopts it. The company requests collaboration, administration, security, and governance features. The vendor introduces a business or enterprise plan. This consumer-to-prosumer-to-enterprise journey is faster because the employee does not have to wait for a company-wide technology decision before experiencing value. It also changes the role of product quality. In traditional enterprise software, a mediocre user experience could survive because the buyer was not always the user. In product-led AI software, users can compare alternatives immediately. They may personally pay for the product. They can cancel easily. They can recommend competitors.

This puts pressure on vendors to deliver:

Immediate value Simple onboarding High-quality output Reliable performance Clear pricing Easy sharing Natural collaboration A smooth path to team adoption Enterprise features still matter, but they often become the second sale rather than the first.

9. AI Is Turning Specialized Skills Into General Capabilities

A central theme in the spending data is the horizontalization of expertise. Some tasks were historically associated with distinct professions. Coding belonged to engineers. Graphic design belonged to designers. Video production belonged to editors. Data analysis belonged to analysts. Legal drafting belonged to lawyers. Sales research belonged to business-development teams. Meeting documentation belonged to assistants or operations staff. AI does not make every user an expert in these fields. It does make basic and intermediate capabilities more accessible. A marketer may build a simple internal application.

A founder may generate a first version of a product video. A recruiter may analyze interview transcripts. A salesperson may produce an account-specific presentation. An operations employee may automate a reporting workflow. This has two major consequences. Organizational boundaries weaken Departments become less dependent on one another for small tasks. Employees can complete more work independently. Specialists move toward higher-value judgment

Professionals may focus more on:

Complex decisions Quality control Original strategy Taste Risk management Stakeholder communication Ethical responsibility Exception handling The most valuable specialists will not necessarily be those who manually perform every task. They may be those who know how to direct AI, identify mistakes, combine outputs, and apply domain judgment.

10. Meeting Intelligence Is Becoming Essential Workplace Infrastructure

Meeting products formed another visible group in the ranking.

The report mentioned applications such as:

Fyxer Happy Scribe Plaud Otter.ai Read AI Cluely These products record, transcribe, summarize, organize, and increasingly provide real-time assistance during conversations. At first glance, meeting notes may appear to be a narrow feature. In reality, meetings contain a large share of an organization’s unstructured knowledge.

Important information is frequently spoken but never entered properly into company systems:

Customer objections Product decisions Hiring feedback Operational problems Strategic priorities Commitments Deadlines Sales signals Risks Institutional knowledge AI meeting systems turn spoken conversation into searchable and reusable data. The product can eventually do more than produce a transcript.

It can:

Extract action items Assign owners Update a CRM Draft follow-up emails Identify customer sentiment Detect repeated objections Compare decisions across meetings Generate project updates Measure participation Prepare future agendas Alert leaders to unresolved commitments This creates a path from passive recording to active workflow management.

The winner in this category may not be the best transcription tool. It may be the system that converts conversations into coordinated organizational action.

11. Startups Are Becoming the Best Early Market for AI Applications

Large enterprises possess enormous budgets, but startups may be better early adopters for many AI-native products. Why? Startups have fewer legacy systems They are less constrained by old software, long-term contracts, and established processes. They have stronger pressure to operate efficiently A startup may use AI to delay hiring, extend runway, or compete with larger companies. Decisions can be made quickly A founder or department leader can approve a product without months of procurement. Teams are more willing to redesign workflows A new company can build its operations around AI rather than adding AI to a decades-old process. The cost of experimentation is lower Small teams can test a new product with fewer organizational dependencies.

Mercury’s broader 2025 survey of 1,500 early-stage U.S. companies found that 60 percent of companies reporting significant AI adoption said their confidence in their financial prospects had improved significantly from 2024. The equivalent figure among businesses that had not adopted AI tools was 28 percent. That does not prove AI caused the optimism. More confident or better-performing companies may also be more likely to invest in technology. Still, the relationship indicates that AI adoption is increasingly associated with how founders think about efficiency, competitiveness, and growth. The survey also found that 91 percent of Gen Z and millennial entrepreneurs and 87 percent of Gen X and baby boomer entrepreneurs had incorporated at least some AI into their businesses. AI adoption among startups is therefore broad rather than confined to younger founders or purely technical companies.

12. AI Spending Is Partly Replacing Other Budget Categories

It is tempting to treat AI spending as a new software category layered on top of existing expenses. That will not always be the case.

AI applications may compete with several existing budget lines:

Employee salaries Contractor spending Agency retainers Outsourcing Consulting Traditional software subscriptions Training Administrative support Research services Content production Professional services This means founders should not calculate the value of an AI product only by comparing it with another software subscription.

A $1,000 monthly AI platform may appear expensive relative to a $50 SaaS product. But it may be inexpensive if it eliminates $5,000 in agency work, delays a hire, produces more sales opportunities, or completes hundreds of hours of repetitive work. The appropriate comparison depends on the job being performed. AI founders should therefore define their economic value clearly.

Possible value metrics include:

Hours saved Cost per completed task Revenue generated Cases resolved Leads qualified Content variations produced Engineering time reduced Time to deployment Error rate lowered Compliance work completed Customers retained Hiring needs delayed

Products that demonstrate measurable value will be more resilient when companies review budgets.

13. The AI Market May Support Multiple Winners in the Same Category

Traditional technology analysis often asks which company will win a market. AI may produce a more fragmented reality.

Users can switch between models and applications depending on:

Accuracy Speed Price Context length Privacy Specialization Integrations Output style Geographic availability Regulatory requirements A company may use OpenAI for general knowledge work, Anthropic for coding or analysis, Perplexity for research, Notion for internal documents, and several specialist applications for particular workflows. The cost of adopting an additional AI tool can also be relatively low.

Unlike replacing a core enterprise resource planning system, subscribing to a new AI assistant may require minimal implementation. This makes multi-product usage likely. However, tool proliferation creates a new problem. Companies may accumulate dozens of overlapping AI subscriptions.

That leads to:

Wasted spending Duplicate functionality Fragmented data Security concerns Inconsistent quality Weak governance Unclear ownership Difficulty measuring return on investment A new category of AI management platforms may emerge to help companies monitor, secure, provision, compare, and optimize their AI applications. The AI application boom may therefore create opportunities not only for applications, but also for an operating layer that manages them.

14. What Founders Should Learn From the Spending Report

The ranking is useful, but copying the visible winners is not a strategy. Founders should study the underlying purchasing behavior. Lesson 1: Build around frequent work Products used daily or weekly have more opportunities to demonstrate value and become habits. Lesson 2: Reduce time to first value Users should experience a useful result quickly, ideally before a sales call or complex implementation. Lesson 3: Start with a narrow problem but design for expansion A focused use case can attract early users. The product should then expand into adjacent tasks, teams, or workflows. Lesson 4: Own more of the workflow A feature generates an output. A platform moves the user from intention to completed result. Lesson 5: Make individual adoption easy A self-service plan can create internal champions who later bring the product into a team.

Lesson 6: Prepare for enterprise requirements early

As usage expands, customers will ask for:

Security controls Administration Role-based access Audit logs Data retention policies Compliance Procurement support Integration Service-level commitments Lesson 7: Connect pricing to value Seat-based pricing may not fit an autonomous product that reduces the number of human users required. Consider usage, workflow, output, performance, or outcome-based models.

Lesson 8: Build trust into the product As AI moves from suggestion to execution, customers need transparency, review, permissions, and control. Lesson 9: Design for human escalation No autonomous system will handle every case reliably. Products should know when to request approval or transfer work to a person. Lesson 10: Measure retention, not curiosity A viral launch may attract attention. Durable spending requires recurring operational value.

15. Where the Next Large AI Application Companies May Emerge

Based on the spending patterns and broader market direction, several categories appear especially promising. AI workflow operating systems Platforms that coordinate multiple agents, tools, databases, and approval steps. AI finance departments Applications that manage bookkeeping, forecasting, payments, collections, reporting, and financial operations. AI legal and compliance services Systems that complete standardized filings, reviews, monitoring, and documentation. AI customer operations Products that combine support, success, retention, onboarding, and account intelligence. AI sales execution Systems that move beyond drafting messages and manage larger portions of the revenue workflow. AI-native business-process outsourcing

Companies that replace conventional outsourced service teams with software-led operations. AI creative operations Platforms that connect generation, brand management, approvals, distribution, testing, and performance measurement. AI security and governance Systems that monitor how employees and agents use models, data, tools, and permissions. AI employee infrastructure Identity, email, phone numbers, wallets, accounts, access controls, memory, audit trails, and operational systems designed specifically for agents. Industry-specific agent platforms End-to-end products for insurance, logistics, healthcare administration, construction, real estate, manufacturing, and government services. AI application procurement and optimization Tools that help businesses understand which AI subscriptions they use, what they cost, who accesses them, and whether they produce measurable value.

16. The Deeper Transformation: Software Is Becoming Labor

The most important conclusion from the AI application spending report is not about one ranking. It is about the changing economic identity of software. Traditional software stores information and helps people complete tasks. AI software can interpret instructions, generate outputs, make recommendations, operate tools, and increasingly complete workflows. The unit being purchased is shifting. Companies once purchased software seats.

Now they may purchase:

Completed research Resolved tickets Generated videos Qualified leads Reviewed contracts Deployed applications Processed documents Completed reconciliations Managed cases Autonomous working time This blurs the line between software and labor. It also changes competitive strategy.

An AI application may compete against:

Another software company An employee A contractor An agency A consulting firm An outsourcing provider An internal department A traditional professional-services company The largest AI application companies may therefore be much larger than conventional SaaS businesses because they can address both technology budgets and labor budgets. But that opportunity comes with greater responsibility. When software performs work, customers care not only about features. They care about accuracy, accountability, reliability, safety, and outcomes.

Key Takeaways

Startup spending shows real demand. Transaction data offers a stronger signal than website traffic, downloads, or social media attention. Horizontal AI currently holds a slight advantage. General assistants, creative tools, coding platforms, and meeting products can spread across many departments. The general-assistant market may support several major providers. Companies are likely to use different models and interfaces for different purposes. Creative AI is becoming horizontal business software. Image, video, audio, and advertising tools are now used across companies, not only inside creative departments. AI coding has entered workplace budgets. The category is expanding from developer assistance toward complete application creation and deployment. Vertical AI is evolving from augmentation toward substitution. Most tools still support employees, but a growing group aims to complete jobs or workflows end to end. AI-native services may capture larger markets than conventional SaaS. They can compete for labor and professional-services spending, not only software budgets. Consumer-first distribution is becoming a major enterprise strategy. Employees adopt useful AI products individually and later bring them into teams. Specialized skills are becoming more widely accessible. More employees can create, code, analyze, research, and automate without deep formal expertise. Meeting data is becoming organizational infrastructure. AI can convert spoken conversations into searchable knowledge and executable workflows.

Startups are ideal early customers. They move quickly, have fewer legacy systems, and face strong pressure to produce more with smaller teams. AI spending may replace other expenses. Products should be evaluated against labor, agencies, outsourcing, and professional services, not only software subscriptions. The winning products will own workflows, not isolated outputs. Trust and governance become more important as autonomy increases. Software is gradually becoming a purchasable form of digital labor.

Frequently Asked Questions

What was the AI Application Spending Report?

It was an analysis published by Andreessen Horowitz in collaboration with Mercury. The researchers examined AI application spending among more than 200,000 Mercury customers during June through August 2025 and identified 50 prominent AI-native application companies.

Did the report include cloud and AI infrastructure spending?

No. The analysis intentionally excluded companies primarily selling cloud services, GPUs, and infrastructure products. Its focus was the application layer, where AI is used directly in business products and workflows.

What is a horizontal AI application?

A horizontal AI application can be used across many departments, professions, and industries. General assistants, meeting tools, creative platforms, and some coding products are examples.

What is a vertical AI application?

A vertical AI application is designed for a specific profession, function, workflow, or industry, such as legal work, recruiting, accounting, customer support, or sales.

Which group represented more spending?

Horizontal companies represented approximately 60 percent of the ranking, while vertical companies represented about 40 percent.

Why were creative AI tools so prominent?

Creative AI has reduced the cost and expertise required to generate images, audio, video, advertisements, presentations, and other material. These capabilities are useful across nearly every department, not only inside design teams.

What is vibe coding?

Vibe coding generally refers to building software with substantial assistance from AI, often through natural-language instructions, iterative prompting, and generated code.

Is AI replacing employees?

Most vertical applications identified in the report were augmenting employees rather than replacing them. However, some companies were designed to perform larger workflows as AI engineers, sales workers, legal providers, or IT operators.

Why are startups purchasing so many AI applications?

Startups face pressure to move quickly, control costs, and operate with small teams. AI applications can increase output, delay hiring, automate repetitive tasks, and reduce dependence on external service providers.

Why do consumer AI products succeed in enterprise markets?

Employees can adopt these products personally, demonstrate value, and introduce them to colleagues before the company signs an enterprise agreement. This bottom-up process can be faster than traditional enterprise sales.

Will one AI assistant dominate the market?

It is possible, but current spending patterns suggest many companies use multiple assistants and models. Different systems may specialize in research, coding, documents, automation, security, or cost efficiency.

How should AI applications be priced?

Pricing may be based on seats, usage, tasks, completed workflows, output volume, performance, or business outcomes. Products acting as autonomous workers may eventually require alternatives to traditional per-user SaaS pricing.

What is the biggest opportunity for vertical AI?

The largest opportunity may be owning an entire workflow or delivering a completed service rather than providing one isolated feature.

What risks should companies consider when adopting AI applications?

Important risks include inaccurate output, confidential-data exposure, weak access controls, vendor dependence, regulatory violations, unclear accountability, intellectual-property issues, and uncontrolled software spending.

What should founders build after seeing this report?

Founders should not simply copy the highest-ranked products. They should identify expensive, repetitive, measurable workflows where AI can deliver immediate value and gradually assume more responsibility.

Conclusion

The AI Application Spending Report provides a useful snapshot of where early-stage companies were directing their AI budgets in 2025. The money was not flowing only toward foundation models and infrastructure. It was flowing toward practical applications that helped startups write, design, build software, conduct meetings, support customers, recruit employees, automate compliance, manage sales, and deliver professional work. Horizontal applications held the early advantage because they could spread quickly across departments. Creative AI and coding tools became workplace categories. General assistants developed into a new interface for company knowledge. Consumer products moved rapidly into teams. Vertical applications began advancing from employee augmentation toward end-to-end execution. The broader transformation is larger than any individual company on the list. AI is changing software from a passive tool into an active participant in work. As applications gain the ability to understand instructions, use tools, make decisions, and complete workflows, businesses will increasingly purchase not only software access but also software-produced labor. The next generation of successful AI companies will therefore need to master more than model quality. They will need to combine excellent products, distribution, workflow ownership, measurable economics, safety, governance, trust, and responsibility for outcomes.

The central question for founders is no longer simply:

What can AI generate?

The more valuable question is:

What complete and economically important work can AI reliably perform?

Relevant Articles and Resources

1. Andreessen Horowitz: The AI Application Spending Report

The original report examining AI application spending across more than 200,000 Mercury customers and ranking 50 AI-native application companies.

2. Mercury: The New Economics of Starting Up

A survey of 1,500 U.S.-based early-stage business founders and executives covering AI adoption, financial confidence, hiring, spending, funding, and operating strategies in 2025.

3. OpenAI: The State of Enterprise AI 2025

OpenAI’s report on business adoption, workflow integration, enterprise usage, and organizational patterns across its business customer base.

4. Anthropic Economic Index

Anthropic’s continuing research program analyzing how AI is used across occupations, tasks, geographies, consumer interactions, and enterprise API activity.

5. Andreessen Horowitz: Top 100 Generative AI Consumer Apps

A complementary ranking that examines consumer AI products primarily through web and mobile usage signals rather than startup transaction spending.

6. Stripe: Inside the Growth of the Top AI Companies

Stripe’s analysis of the unusually rapid revenue growth, international expansion, and monetization patterns seen among leading AI businesses.

7. Stripe 2025 Annual Letter

A broader overview of trends affecting online businesses, AI companies, international commerce, financial infrastructure, and agentic economic activity.