Every Vertical AI Company Is Going to Die

Emerging foundation models threaten the existence of vertical AI companies as they absorb specialized applications across various domains. - Vikramsinh P Ghatge

For a while now, I have had this thought. It started with the tools I was using. An AI product would enter my workflow because it did one thing remarkably well. It might help me write marketing content, analyse a document, research a topic, build a presentation or organise information.

Then, a few months later, I would notice that I was using it less.

Not necessarily because the product had become worse. ChatGPT, Claude or Gemini had simply become better.

A task that once required a separate application could now be completed inside the foundation model itself. The model could retain more context, search the web, interpret files, create documents, generate images, use software and connect with business systems.

Sometimes the original tool disappeared. More commonly, it changed what it sold.

An AI writing product became a marketing platform. A copy generator became a go-to-market workflow system. A presentation generator became a CRM. A browser company stopped actively developing the browser that had made it famous and redirected its attention towards an AI-native replacement.

Most of these companies did not announce that a foundation model had undermined their original proposition. Their public explanations focused on enterprise customers, monetisation, product-market fit, richer context or a larger market opportunity.

Those explanations matter.

A company rarely pivots or fails for one reason. Distribution, pricing, retention, capital efficiency and execution all play a part. It would be misleading to look at every strategic change in the AI market and blame ChatGPT or Claude or Gemini.

But after seeing the same movement repeatedly, I began to wonder whether something more structural was happening.

My hypothesis is deliberately provocative:

Every vertical AI company is going to die.

By vertical AI company, I mean an AI-first business built around a specialised function, profession or industry. I do not necessarily mean that every one of these companies will run out of money, shut its doors or dismiss its employees.

Some will. Others will be acquired. Some will become capabilities inside larger software platforms. A smaller number will continue growing under the same company name.

But even the successful companies will eventually have to move beyond the narrow proposition that originally made them valuable.

The AI writer must become a marketing operating system.

The legal assistant must become part of the legal workflow.

The coding assistant must become the development environment.

The healthcare scribe must become part of clinical operations.

The research tool must own trusted information, organisational context and decision-making.

The legal entity may survive.

The original product thesis will not.

Foundation models are no longer just infrastructure

The traditional explanation of the AI software market separates horizontal foundation models from vertical applications.

OpenAI, Anthropic, Google and other model companies provide broad intelligence. Specialist applications adapt that intelligence for marketing, law, healthcare, finance, sales or another domain.

That description made more sense when foundation models primarily produced responses through an API or chat interface. It is becoming increasingly incomplete.

Foundation-model companies are now building the tools, interfaces and workflows through which people use that intelligence.

OpenAI launched Deep Research in February 2025 as an agent capable of conducting multistep internet research, analysing text, images and PDFs, and producing cited reports. It later added the ability to connect Deep Research to applications and Model Context Protocol servers and to restrict searches to trusted sources.

Anthropic introduced Projects as workspaces where users could give Claude their own documents, code and knowledge. It later expanded Claude from generating text to creating and editing spreadsheets, documents, presentations and PDFs directly. Anthropic described the shift as moving Claude from answering questions towards completing entire projects.

Microsoft has made a similar move inside Office. Its Agent Mode can directly modify Word documents, Excel workbooks and PowerPoint presentations. Microsoft said earlier foundation models were not capable enough to reliably command the applications, but improvements in reasoning and instruction-following made multistep editing possible.

Google is going further inside marketing execution.

In May 2025, Google announced broader agentic capabilities for Google Ads and Google Analytics. The tools were designed to assist with campaign creation, reporting, optimisation and troubleshooting, and could implement some recommendations with the advertiser’s approval. Google also introduced Marketing Advisor, an agent intended to work across Ads, Analytics, websites and content-management systems through Chrome.

By November 2025, Google had announced Ads Advisor and Analytics Advisor. Ads Advisor could generate keywords and creative assets, diagnose performance changes and apply approved changes directly to an account. Analytics Advisor could investigate why performance had changed and recommend specific actions. In May 2026, Google introduced Ask Advisor, a cross-product agent connecting Ads, Analytics and Merchant Center. Google demonstrated it using product information to set up campaigns and combining advertising and analytics data to explain results and recommend what to do next. This is not simply a better model supporting independent marketing software. Google owns the model, the advertising platform, the analytics, much of the customer data and the environment where the recommended action is executed.

That is a very different competitive structure.

A faster cloud server does not gradually become a marketing department.

A database does not decide that it can now research, create and launch an advertising campaign.

Foundation models are different because improvements in the underlying technology directly expand the work the platform can perform.

Every improvement in reasoning, memory, multimodality, file creation, tool use and system access can remove part of the reason a specialist application exists.

This creates what I call the Native Feature Clock.

The Native Feature Clock is the countdown between a company commercialising a model capability gap and that capability becoming sufficiently native to a foundation model or dominant software platform.

It is not a fixed period or a scientific metric. It is a way to describe the strategic uncertainty facing every vertical AI business. The company does not know whether the capability supporting its product will remain differentiated for four years or four months.

It only knows that the underlying platform is still improving.

AI writing companies were an early warning AI writing was one of the first major generative AI application markets. Before general-purpose chatbots became widely accessible, products such as Jasper and Copy.ai made language models easier for marketers and business users to adopt. They added templates, brand voices, interfaces and workflows for producing advertisements, blog posts, emails, social content and product descriptions. That was a real product advantage.

The underlying models could generate text, but most business users could not easily access them, prompt them consistently or turn the output into a repeatable marketing process.

In October 2022, Jasper announced that it had raised $125 million at a $1.5 billion valuation. It described itself as an AI content platform and launched a browser extension that worked across products including Google Docs, Gmail, HubSpot and Shopify.

Its proposition today is much broader.

Jasper now sells marketing agents, content pipelines, brand intelligence, governance, company knowledge and end-to-end marketing workflows. Its own website describes purpose-built agents that execute marketing processes and a governed intelligence layer containing brand context, rules and business logic. Reuters reported in 2024 that Jasper had experienced layoffs and leadership changes while shifting its focus from consumers towards enterprise customers.

Those facts establish that Jasper changed its position.

They do not prove that ChatGPT alone caused the change.

My interpretation is narrower: once high-quality general writing became broadly available, AI content generation by itself was no longer enough to carry the entire proposition. Jasper needed to move further into marketing context, governance and workflow.

The company called Jasper survived. The original AI writing thesis did not.

Copy.ai made a comparable transition. In 2024, the company formally launched what it called a go-to-market AI platform. It expanded beyond copy generation into workflows connecting marketing, sales and operations and said it had passed 15 million users.

This is not a minor product enhancement. It changes what the customer is buying. The original unit of value was generated copy. The newer unit of value is coordinated commercial work.

Both Jasper and Copy.ai demonstrate the first form of vertical AI death: strategic death.

The corporate entity continues, but the proposition that originally justified the company ceases to be sufficient.

Tome shows that massive adoption does not guarantee durability. Tome offers one of the clearest examples of the original product thesis dying while the founding team continued. The company launched in 2022 with a generative AI presentation and visual storytelling product.

By April 2024, Tome had around 20 million users. Most were not paying. The company announced a restructuring affecting approximately 20 percent of its 59 employees and redirected resources towards enterprise sales and products built for sales organisations.

Tome co-founder Keith Peiris told Semafor that the company’s real business was not the millions of people creating free presentations. It was the smaller number of sales and marketing teams willing to pay for a more sophisticated product. The company began adding customer research, personalisation, Salesforce data and information from internal sales calls. Peiris said the new product would be more focused on information processing than aesthetics.

Tome later shut the presentation product completely.

In a public post, Peiris said the company had killed a product used by more than 25 million people and built Lightfield for sales teams. He argued that producing genuinely useful sales work required rich customer context that a generic presentation tool did not possess.

Lightfield now describes itself as an AI-native CRM. It captures customer interactions, maintains context across emails and meetings, and allows companies to build agents that prospect, follow up and update pipeline information. The founder did not say that a foundation-model update killed Tome.

The available evidence points more directly towards weak monetisation, insufficiently mission-critical usage and the need for richer customer context. That distinction should not be erased to make the thesis more dramatic. But the result still supports the broader pattern. A generic AI-generated output attracted tens of millions of users and was still not considered a durable enough business.

The company moved from generating presentations to owning customer memory and sales workflow. The presentation company died. A deeper software company took its place.

Sometimes the model does not copy the product. It removes the need for it.

Chegg is not a vertical AI startup, but it offers unusually direct evidence of a general AI product reducing demand for a specialised service. Chegg had built a large subscription business around homework support, educational content and tutoring.

In 2023, it acknowledged that ChatGPT was affecting new-subscriber growth. The company introduced its own AI products, but subscriber and traffic pressure continued.

In May 2025, Chegg announced that it would reduce its workforce by approximately 22 percent. It reported that first-quarter subscribers had fallen by 31 percent to 3.2 million and revenue had declined by 30 percent to $121 million. Chegg said students were increasingly using products such as ChatGPT and that Google’s AI Overviews and movement towards Gemini were reducing traffic to specialist websites.

ChatGPT did not reproduce every part of Chegg. It did not need to.

It satisfied enough of the underlying user intent that some students no longer needed to visit or pay for the specialist service.

That is another form of absorption. A foundation model does not always have to copy the product feature by feature. Sometimes it makes the product less necessary. The same principle applies to simpler AI applications. A foundation model does not need to become the best dedicated PDF tool in the market. It needs to analyse the user’s document well enough that installing and paying for another product feels unnecessary.

It does not need to recreate every presentation application. It needs to generate a sufficiently usable deck inside the environment where the user is already working.

It does not need to replace every research platform. It needs to answer enough researched questions with acceptable sourcing that the user stops opening the separate tool.

The threshold is not perfect replacement. The threshold is reduced willingness to pay.

Acquisition is another form of death. A company being acquired is not a business failure.

The founders, employees and investors may achieve an excellent financial outcome.

But acquisition can still mark the death of the company as an independent vertical product.

Thomson Reuters agreed to acquire legal AI company Casetext for $650 million in 2023. Casetext’s technology became part of CoCounsel and the wider Thomson Reuters professional product portfolio.

Docusign acquired AI contract-management company Lexion for approximately $165 million in 2024. Docusign said Lexion’s technology would strengthen its Intelligent Agreement Management platform through capabilities such as contract analysis, automated review and agreement question answering.

The specialist capabilities did not disappear. They were incorporated into companies that already owned legal content, agreement data, enterprise distribution and established professional workflows.

The Browser Company went through a related transition. It became known for Arc, a browser differentiated through interface design and a different approach to navigating the web. The company later stopped active feature development on Arc and concentrated on Dia, its AI-focused browser.

Atlassian agreed to acquire The Browser Company for $610 million in 2025, with Dia positioned as the strategic product. Arc remained available but was no longer the centre of active development.

It would be inaccurate to say that a foundation model directly killed Arc. The company also faced questions around complexity, performance and mainstream adoption.

The more defensible conclusion is that AI changed the company’s belief about what the browser needed to become.

Arc did not necessarily fail as a product. It ceased to be the company’s future.

Acquisition therefore represents a second type of vertical AI death. The capability survives.

The independent company thesis is absorbed into a larger workflow or distribution system.

The Native Feature Clock leads to the Absorb-or-Evolve Cycle

Across these examples, a recurring sequence appears.

A foundation model initially performs a task poorly, unreliably or inconveniently. A startup packages that gap into a product. It adds prompts, context, templates, a specialist interface and integrations.

Customers adopt it because it is meaningfully better than using the underlying model directly.

The model then improves. It gains stronger reasoning, longer context, memory, web access, file creation, connectors or the ability to take actions.

The foundation-model company, or an established software platform, introduces an acceptable version of the same capability inside a product the customer already uses.

The vertical company reaches a decision point.

It can remain focused on the original capability and face commoditization.

It can be acquired.

It can close the product.

Or it can move deeper into the customer’s workflow.

This is the Absorb-or-Evolve Cycle.

The Absorb-or-Evolve Cycle is the recurring process through which foundation models and dominant platforms absorb application-level capabilities, forcing vertical AI companies to build a deeper source of value or lose the market that originally created them.

This is an analytical framework, not a proven economic law.

It does not claim that every company pivot was caused by a specific model release.

It describes the strategic pressure produced by the continued expansion of the foundation layer.

The platform does not need to build a perfect substitute. It only needs to build a good-enough substitute inside a product the customer already pays for.

A marketer may prefer the specialist application. But will the organisation keep paying for it when an acceptable version is included with ChatGPT, Claude, Google Workspace, Google Ads or Microsoft 365?

That is the commercial pressure that matters.

The only escape is vertical

If the Native Feature Clock describes the problem, Vertical Escape Velocity describes the possible response.

Vertical Escape Velocity is the speed and depth with which an AI company moves into proprietary context, workflow ownership, governance and execution before its original capability becomes widely available.

A copy generator has little escape velocity.

A marketing platform that manages brand rules, approvals, campaigns, budgets, attribution and historical performance has considerably more.

A legal chatbot has little escape velocity.

A legal operating environment connected to trusted content, institutional precedents, client matters, review processes and professional accountability has more.

A medical transcription product may be replicated.

A clinical platform embedded inside hospital systems, documentation standards, coding, billing and physician workflows is considerably harder to replace.

The simplest test is:

Does a better foundation model strengthen the product, or erase the reason the customer needs it?

When a better model strengthens the product, the company may be approaching escape velocity.

Harvey illustrates this movement in legal AI.

The company has expanded beyond isolated legal questions into contract analysis, due diligence, compliance, litigation and agent-based legal workflows. In March 2026, it raised $200 million at an $11 billion valuation. Harvey said the funding would support its AI agents and legal-engineering teams.

Harvey has not built a valuable company because Claude or ChatGPT cannot produce legal language.

It has built value around legal workflow, professional context, deployment, customer relationships and the operational requirements of legal work.

Abridge demonstrates the same movement in healthcare. The company uses AI to turn medical conversations into clinical documentation. In February 2025, it raised $250 million and said its technology had been implemented across approximately 100 US healthcare systems. The company and its investors also discussed expanding beyond documentation into the rules and complexity surrounding revenue-cycle workflows.

Norm AI focuses on legal and regulatory operations rather than general text generation. In July 2026, it raised $120 million at a $1.2 billion valuation, with its products positioned around compliance and legal workflows. These valuations are not proof of permanent defensibility.

Funding is not a moat.

These companies may still face pressure from foundation models and established software providers. What they demonstrate is the direction a vertical company must travel. The AI capability becomes only one part of a deeper system involving data, professional rules, permissions, reviews, auditability and execution.

Harvey, Abridge and Norm do not disprove the death hypothesis. They illustrate the only available alternative.

The companies are trying to become too deeply embedded to be reduced to a native feature.

Escape velocity is never permanent

Even moving deeper into a workflow does not make a vertical AI company permanently safe.

Foundation-model companies continue to expand.

Software incumbents continue adding AI.

Google owns advertising and analytics workflows. Microsoft owns Office and an enormous enterprise distribution network. Thomson Reuters owns legal and professional information. Docusign owns agreement workflows. Atlassian owns collaboration and software-development environments.

The vertical company is therefore pressured from two directions.

Foundation-model companies are moving upwards from intelligence into applications.

Established software companies are moving sideways from existing workflows into AI.

The vertical application occupies the increasingly compressed space between them.

This is why prompts, interfaces and access to a particular model are unlikely to provide durable protection.

Model orchestration may improve performance and reduce dependency on one provider, but it is increasingly becoming standard infrastructure. Connectors are useful, but standard protocols make them easier to reproduce.

The deeper advantages arise from what the company owns after the model has generated an answer:

The approved decision.

The transaction.

The institutional record.

The proprietary feedback loop.

The compliance process.

The customer relationship.

The accountability.

The system in which work is actually completed.

The strongest vertical AI company may eventually look less like an AI application and more like the next generation of industry software.

Every vertical AI company will die

The sentence sounds absolute because it is intended to challenge how we define survival.

A company is not only its legal registration, brand name, employees or investors. It is also a theory about where value exists.

• Jasper’s early theory centred on making AI content generation accessible. It now sells marketing agents, intelligence, governance and workflow infrastructure.

• Copy.ai’s early theory centred on generating copy. It now sells go-to-market orchestration.

• Tome’s theory centred on AI-generated presentations. That product was shut down and replaced by a sales platform.

• Arc’s theory centred on redesigning the browser. The company redirected its future towards an AI-native browser and was acquired by Atlassian.

• Casetext became part of Thomson Reuters.

• Lexion became part of Docusign.

The financial outcomes were different. The pattern was similar. The original proposition ceased to be enough.

Every vertical AI company begins by exploiting a gap between what a customer needs and what a general model can reliably deliver.

Every one therefore begins with a Native Feature Clock already ticking. Eventually, it enters the Absorb-or-Evolve Cycle. Companies that remain where they started risk becoming features, being bundled into larger platforms or losing demand. Companies that survive must reach Vertical Escape Velocity by moving into proprietary context, workflow control, governance, transactions and systems of record.

They may retain the same name. They may become much larger businesses. But they will no longer be the companies they started as. Some will die through failure. Some will die through acquisition. The best will die through evolution.

Research and disclosure note

This article presents “Every vertical AI company is going to die” as a strategic hypothesis, not a statistically proven universal outcome.

The research separates three types of statements:

Verified facts include documented product launches, funding rounds, acquisitions, workforce reductions and changes in company positioning.

Company-reported claims include user totals and founders’ explanations of their own strategic decisions. These are attributed to the company or founder rather than treated as independently audited facts.

Interpretations include the argument that the expansion of foundation-model capabilities contributed to a company’s need to move up the workflow. Where a company did not establish that causal relationship, the article does not present it as a company admission.

Reddit, Hacker News, X and other community forums were used to identify potential examples and understand recurring market concerns. Anonymous posts and unverified claims were not used as factual evidence.

The concepts Native Feature Clock, Vertical Escape Velocity and Absorb-or-Evolve Cycle are analytical terms introduced for this article. They are not established quantitative measures or previously validated economic frameworks.