Tomorrow’s Titans: Vertical AI

The past two years have marked a rapid evolution in AI and the core building blocks of modern software. The zeitgeist moment of ChatGPT set off a chain reaction of innovation, competition and R&D investment in AI, and two years later, we see unprecedented growth of new horizontal applications, such as ChatGPT, Perplexity, Claude, Midjourney, Cursor, ElevenLabs, and more. As AI capabilities mature to offer generative, agentic and multi-modal functions, we see an exciting opportunity for new verticalized applications. 

The next titans of software will be vertical AI companies in specialized industries. 

In this post we share our framework for vertical AI, exploring why we believe now is the optimal time to build industry-specific AI solutions and the key considerations for creating category winners.

Services to Software Turning Point

In the cloud era, we witnessed large vertical software winners such as Procore ($11B), Toast ($21B), and ServiceTitan ($9B). [1] The winners of this first wave followed a common pattern:

1. System of record: Automate and centralize customer data 

2. Payments & billing: Streamline financial workflows and business intelligence

3. Marketplace: Bring ecosystem of industry participants onto one platform 

   Now generative and agentic AI unlocks another potential layer: 

4. Agentic system: AI actions that can significantly automate remaining labor-heavy workflows. As the marginal cost of machine intelligence continues to decline due to intense competition among foundation model providers and hundreds of billions of capex investment in GPU infrastructure, it’s easier than ever to build a large business that makes AI useful for specialized, industry-specific work. [2]

The nature of software is changing from a system of record to a system of actions. [3] In the AI era, defensibility is likely to emerge from specialization and orchestration of AI agents, lending itself to verticalized solutions. Early examples include Harvey AI and Abridge.    

With AI agents advancing beyond mere digitization to actively performing tasks on behalf of end users, software companies have the opportunity to capture significantly more total addressable market (TAM). By tapping into the $11T U.S. labor spend [4] — far exceeding the approximately $450B enterprise software market [5] — these companies can unlock unprecedented growth. 

In this post, we discuss:

• Why now is the right time to build in vertical AI 

• Where vertical AI can win over horizontal products 

• The new vertical AI playbook

• Four questions worth debating

• Ideas for industries where vertical AI is likely to emerge

The New Control Point: Agents

Today, software is designed for user initiated workflows. The largest software companies—such as Salesforce, Workday, Oracle, and SAP—function as systems of record, relying heavily on human-entered data. Consider the average sales representative, who spends six hours a week [6] updating their CRM (we feel for them!). The competitive advantage of these systems of record is rooted in three key factors: 

1. High switching costs due to their role in storing and structuring critical data 

2. Ownership of workflows because users operate within their platforms

3. A robust ecosystem of applications and integrations that creates lock-in 

However, in a world where software takes actions on behalf of users, the traditional workflow advantage of systems of record diminishes. If AI agents process data before it ever reaches these systems (and generate new data abstracted from the systems of record), the incumbents risk losing their advantage tied to data ownership and the resulting switching costs. 

The "control point"—the most mission-critical software in a customer’s stack that drives upselling opportunities—has often traditionally been the system of record. For instance, in home services, it’s Dispatch and CRM; in education, it’s the LMS and SIS. 

In the future, the control point may shift to the agentic layer, potentially reducing the importance of existing systems of record and redefining the competitive landscape.

Perfect storm of opportunity for vertical AI

While agentic AI capabilities are still in their infancy, we think there’s a confluence of factors that makes it the right time right now to build a vertical AI disruptor. 

• Getting the job done: Early signals of agents achieving higher performance in discrete jobs or specialized workflows while generalist agents struggle with consistency 

• Potential to capture larger ACVs: If vertical agents can augment human labor, they can also command higher ACVs and scale revenue faster than traditional SaaS tools. 

• Labor shortages in mission-critical trades: Productive agents are the antidote to labor shortages in the U.S. Today, ~38% of Americans have a four-year college degree, compared to 30% 10 years ago and 23% 20 years ago. [7] Education in trades and vocational upskilling lags behind. America has a labor shortage of plumbers, electricians, truck drivers, accountants, teachers, police officers, and many other critical industries.

• Latent demand for professional services: Even without a labor shortage, industries where labor can command a premium and where there is latent demand for labor (ex: financial services, law) have immense pull for AI automation. 

Specialization drives defensibility: Customers in niche markets tend to prefer a single, purpose-designed vendor that fully understands and meets specific requirements. While the bar is high for product-market-fit, AI agents could finally make “custom” scalable. 

The new vertical AI playbook

Unlike the cloud vertical SaaS pioneers who faced minimal competition, tackling pen-and-paper workflows or outdated on-prem solutions in companies with limited engineering resources, today’s AI startups face a tougher landscape. They must contend with horizontal big-tech incumbents and AI model companies moving up the stack to the application layer. The competitive field is intense. So what does it take to win?

Elements of the new playbook: 

1. Leverage AI agents to process unstructured data: Build AI agents capable of handling unstructured customer data such as phone calls, emails, PDFs, invoices, internal documents, etc., and capture this new data set.

2. Use AI-powered wedges to land the customer: Voice agents, semantic document search, content generation are the most common AI-powered wedges. Many companies are giving some part of their wedge away for free to land the customer.

3. Integrate with existing systems of record, at first: Don’t ask the customer to rip-and-replace existing software to start using your product. 

4. Collect new data that comes through the system of actions (engagement layer / agentic workflows) and build to own the new system of record in the long run. 

5. Nail industry-specific GTM: Customers buy solutions, not AI. Know your customer. 

Examples of companies deploying this playbook

We’re observing early and highly successful implementations of this playbook across a diverse range of industries. Here are a few illustrative examples:

Voice agents

• Public safety: Prepared gives away their video and text product for free, allowing PSAPs (911 response centers) to get video and text data from 911 callers. Once they land, they expand with voice agents that ingest unstructured 911 call data, transcribe, summarize, and log the data into the CAD system (the system of record for PSAPs).

• Healthcare: Medical scribes like Abridge and Freed ingest unstructured audio data from patient conversations and log the data into EHR systems. Abridge’s cooperative approach with Epic was a large driver of their growth.

• Home services: Companies like Avoca, Revin, and Drillbit use voice agents to handle inbound and outbound sales, integrating with incumbents like ServiceTitan. For new jobs that come over the phone, data for new jobs first flows into the contact center before it ever reaches ServiceTitan for dispatch / CRM. 

Semantic search and document analysis 

• Insurance underwriting: Companies like Sixfold are using semantic document search to ingest documents and regulatory codes and compare them against an insurer’s underwriting policy to give underwriting recommendations, automating hours of human labor to process this unstructured data. Sixfold integrates with incumbent underwriting workbenches like Guidewire, sitting on top of them to log data in the system of record. 

• Construction: Trunk Tools is using semantic document search in their TrunkText product, integrating with document systems of record in Procore, Sharepoint, or Autodesk. Trunk Tools saves time and is able to reduce the amount of construction rework by combing through documents to find needle-in-a-haystack answers to questions about the job that are buried in the thousands of pages of documents.

Unstructured data processing & content generation

• Education: Companies like MagicSchool and SchoolAI offer AI-powered curriculum content creation to teachers for free, monetizing primarily by selling to schools who buy enterprise licenses. Once they’ve landed with teachers, they expand to AI-powered learning / assignments for students. The learning data they create sits ahead of existing systems of record like the SIS (system of record for grade data) and LMS (system of record for assignments / the classroom).

• Industrial designers: Backflip (NEA portfolio company) generates 3D assets based on a text description, a 2D drawing, or a photo of a real object. They collaborate with CAD incumbents like Solidworks, allowing users to start their work in Backflip and export into Solidworks.

• Life sciences: Companies like Collate are using text generation capabilities to automate the creation of documentation for drug and medical device development.

• Police: Leading body-camera company Axon built AI agents that analyze video footage from their cameras and generate police reports, which today take up 50% of officers’ time. [8] Body camera footage and the generated reports are logged in the CAD and RMS systems of record. Axon has now built their own CAD and RMS offerings, building backwards into a system of record having started with a system of action.

Four questions we consider in vertical AI

What are the foundational elements to win? 

• 10x better product for tasks that require processing unstructured data: Phone/audio, email, documents, designs, video, and more.

• Sectors with labor shortages, latent demand for labor, or heavy outsourcing/BPO spend

• $10B company opportunity: Turns out there are many more markets where you can build a $10B+ company than you’d think, especially when looking at labor budgets. For example, did you know that 17% of accountants left their job from 2019–2021 and there are 18% fewer accounting graduates in 2022 than in 2016? With an average accountant salary of ~$80K, the industry has a $24B shortage of accountants. [9] There are many more labor markets like this.

• Speed: Product and GTM velocity is essential when you’re competing against incumbents. Startups need to achieve mission-criticality before incumbents adopt AI. 

• Ability to subvert and own the system of record in the long run

• Founders who are both an industry insider and technical

• Thoughtful approach to business model as we see more output-based pricing

Where will vertical win over horizontal?

• Complex ecosystem of stakeholders: Industries with highly complex stakeholder dynamics (e.g. construction) or multiple participants in a workflow. Cloud-era examples:

  • Brightwheel built their first wedge product as a parent communication platform (the most important stakeholder) before branching out into an all-in-one platform. 

  • Procore is designed to coordinate many stakeholders across GC, subs, owner, architect, and more. Serving multiple stakeholders is incredibly sticky. 

• Unique data: While generalist foundation models are trained largely on internet data that makes them good at general purpose tasks, many industries require access to proprietary data or specialized data formats, such as: 

  • Proprietary text data: In financial services, platforms like Bloomberg have proprietary data like equity research reports, market/asset pricing data, the long tail of alternative data.

  • Gathered data: Vertical applications that have a role as the data is gathered build a data moat, which is the playbook companies like Prepared and Abridge are following with audio transcription agents.

  • New modalities: No matter how much text data is scraped from the web, GPT-5 likely won’t know how to use CAD tools. Many industries, such as architecture, require specialized representations of data and domain specific models. 

• Industry-specific workflows: While we expect generalist models to continue to improve, we believe going deep in one specific job function will enable higher levels of autonomy.

When does it make sense to train a domain-specific model?

Today, many of the most successful vertical AI companies have used off-the-shelf models, but we expect to see more companies training their own specialized models going forward. There are three reasons for that:

• Plummeting model training costs: DeepSeek V3 proved to the world that you can achieve SOTA performance spending less than $6M to train. [10] We think that any company with an inference bill above $6M should think really hard about how they can train their own model.

• Customer trust: Many customers are not comfortable with their data being sent to a model provider, and the only way to avoid this is to have your own model.

• Unique data (mentioned above): As model training costs go down and if pre-training scaling laws do not hold going forward, the bar goes down for how unique a data set must be to justify training a specialized model. 

What industries do we think a $10B company could be built (a non-exhaustive list)?

We analyzed vertical software businesses in our dataset that achieved $100M ARR, a $1B valuation, or raised $100M+ (as of 1/3/2025). Compared to how large the vertical is, we found some categories had a lot of purpose-built software and some with very little. Both can be great areas to build in. 

In industries that have already produced $10B+ winners, the potential to build a large company is evident; the challenge lies in whether AI-native companies can outperform established incumbents. In underserved industries, the key challenge is achieving intense product-market fit (PMF) and selling into complex customers.

Here are a few spaces we think are ripe for vertical AI (non-exhaustive list):

Legal

• NEA recently invested in Clio, the leader in cloud-based legal practice management. The company is layering AI capabilities into their platform with Clio Duo, their GenAI-powered  tool purpose-built for legal professionals. Clio Duo can streamline everything from case management to document reviews, helping firms operate at peak efficiency.

• With AI, we’ve seen innovation around the “business of law” (CRM, billing, time tracking, etc.) and the “practice of law” (legal research, document/contract review, document drafting, etc.). There is immense latent demand for legal services, creating an opportunity to both enhance and augment traditional legal work. 

• Legal services on fixed-fee billing (e.g. personal injury, immigration) are best positioned to be augmented by AI agents, as AI automation can help satisfy latent legal demand, growing a firm’s topline (and bottomline via more efficiency). 

• Building a product that automates or speeds up legal work is necessary but not sufficient. We’re looking for companies that have cracked a GTM motion and who have found ways to make their space winner-take-all, often through building a compounding data moat. But in order for legal professionals to trust AI, data security is key for adoption.

• For example, Clio Duo is set up to operate securely within Clio’s core case management platform using only its data - ensuring full data safety and privacy, and most importantly, no data is used to train LLMs.

Public Safety

• In public safety, we’ve already seen some of the most successful AI agent deployments from companies like Prepared and Axon.

• Part of the competitive advantage of companies like Axon (body camera) and Flock Safety (license plate reader) is that the data flows through their hardware before it gets to the system of record.

• Labor shortage of police officers following the nationwide protests. [11] Similarly, there is a labor shortage of 911 center operators with a 25% vacancy rate. [12]

Home Services

• The market is so large that ServiceTitan, Jobber, and Housecall Pro together only have an estimated 25% market penetration. [13]

• These businesses are primarily reliant on their website/SEO for discovery and the phone for new jobs, areas where agents are highly impactful. And businesses have the potential to get to significant scale ($10M+ revenue), so can benefit from operational improvements. 

• America faces a large labor shortage in the trades. Nearly 30% of union electricians are near retirement, there is a 100K HVAC worker shortfall by 2025 and a 550K plumber shortfall by 2027. [14]

• Private equity has also begun to roll up HVAC and other home services businesses, creating what the WSJ calls “America’s New Millionaire Class.” [15]

Construction

• Despite being 4.5% of the U.S. economy [16], there’s very little software built for construction. The only real cloud winners have been in project management software (Procore, Aconex, Plangrid, Levelset, Fieldwire, and Buildertrend, etc.).

• While Procore has made attempts and Autodesk acquired BuildingConnected, there isn’t really a software leader in pre-construction—it’s a true greenfield. But the opportunity is bigger than just pre-construction.

• Construction projects are driven by documentation and planning to make sure the design fits the jurisdictional and structural requirements while minimizing cost and keeping the project on schedule. These tasks are fundamentally unstructured data problems that AI is well suited for:

  • Design generation in the BIM (the 3D design an architect creates).

  • Image analysis of the construction site to scope the job or make sure the job is on track and of the BIM for takeoffs and procurement.

  • Semantic document search for permitting and navigating documents to supercharging project management.

P&C Insurance

• While P&C insurance has produced multiple ~$10B winners like Guidewire and CCC, we expect these systems of record to be disrupted by document search and voice agents. 

• Today, underwriters spend only 30% of their time on actual underwriting tasks, with 40% spent on admin tasks and 30% spent on sales support. [17]

• Insurers accept only a fraction of new business given time constraints and current risk models. So if P&C insurers could increase underwriting efficiency and improve risk models, they could drive substantial incremental revenue. 

Brokerages (ex: freight forwarders and some wholesalers)

• Compared to fully automated marketplaces, brokerages offer high touch, placing a human in between each transaction who is coordinating over email and phone between buyer and seller.

• Voice agents, document search, and email automation capabilities are poised to automate this human-driven touch.

• We think building for freight forwarders is a large opportunity, as freight represents ~5% of the U.S. economy and has little purpose-built software. In fact, such little software exists that the dominant approach to modernize the industry is for tech companies to start their own tech-enabled services business (digital freight forwarders like Flexport, Loadsmart, Uber Freight and others).

We’ve been fortunate to partner with AI companies like Perplexity, ElevenLabs, World Labs, Twelve Labs, Together AI, Sakana AI, and Databricks, and vertical software companies like Clio, Cvent, Tulip, D2L, Anterior, Second Front, and Farmhand. We’re motivated by ambitious and visionary founders building generational companies in these categories. If that sounds like you, please reach out to abordetsky@nea.com, tluck@nea.com and jkaplan@nea.com!