Key takeaways
- Your supplier is also a potential competitor. Model providers are combining applications, industry integrations and deployment teams to win customers directly.
- Being vertical is not a moat. Startups must prove their value against competing products, customers building internally and capabilities bundled into existing AI subscriptions.
- Defensibility must accumulate beyond the model. Workflow ownership, useful proprietary feedback, proven reliability and distribution create stronger reasons for customers to stay.
Your most consequential competitor might already be on your expense report. It supplies the model that powers your product, improves your margins when inference gets cheaper, and helps your team ship faster. It can also sell directly to the customers you are trying to win.
For vertical and agentic AI startups, that tension is becoming harder to ignore. OpenAI and Anthropic are packaging capabilities for specific buyers, connecting to industry data and helping enterprises implement workflows. The companies selling intelligence as infrastructure are also competing to own the work performed with it.
My view is that the pursuit of revenue makes this expansion structurally attractive. Selling model access captures one part of the value. Owning the application, customer relationship and deployment creates opportunities to capture more. A provider can benefit from a startup’s growth while also developing products that overlap with it.
The implication for founders is uncomfortable: if you’re building on the application layer, you’re renting the technology. And calling the product “vertical AI” or “agentic AI” does not create a moat for new competitors (including the LLMs themselves) to enter your space. In this post, I look into what this looks like in practice, and what can founders do to mitigate the risks.
The move into applications is already visible
Anthropic’s Claude for Financial Services, announced in July 2025, is a clear example of LLMs moving from ‘just infrastructure’ to ‘vertical AI solutions’. Its Financial Analysis Solution combines Claude with financial data connectors, enterprise capabilities and implementation support. The workflows include due diligence, financial modelling, investment memos and portfolio analysis, with connections to providers including FactSet, PitchBook and S&P Global.
I’d replace it with these two paragraphs:
The overlap becomes clearer when you name the companies and the work they sell. Rogo produces financial models, investment memos, diligence materials and pitch decks for financial institutions. Hebbia helps investment and deal teams analyse documents and financial data, bringing together filings, earnings transcripts and internal materials. Finster AI automates research, analysis and presentation workflows for finance professionals. These are specific jobs that banks and investment firms pay specialist AI companies to perform.
Now compare those offerings with Claude for Financial Services: due diligence, financial modelling, investment memos, pitch decks and portfolio analysis, connected to sources such as FactSet, PitchBook and S&P Global. My reading is that Anthropic is competing for parts of the same workflow and, potentially, the same software budget. That does not make these products interchangeable, but it changes the sales conversation. A specialist startup must explain why its accuracy, integrations, institutional knowledge or execution justify a separate purchase when the model provider is offering to do much of the same work directly.
OpenAI is pursuing smaller companies too. Its small business offering positions ChatGPT Business and ChatGPT Work as tools for creating, analysing and completing work. Its practical programmes include working with QuickBooks data and preparing Shopify workflows.
Small business is a customer segment rather than an industry vertical, but the competitive mechanism matters. A founder selling a narrow marketing, reporting or administrative tool may be competing with capabilities available inside the customer’s existing AI workspace. The buyer’s question becomes: how much additional value justifies another subscription?
Life sciences provides another industry example. Claude for Life Sciences, introduced in October 2025, added scientific connectors, specialised skills and domain support. Its stated use cases include literature reviews, protocol drafting, bioinformatics and preparing regulatory materials. Connections to Benchling, PubMed and other scientific resources bring the assistant closer to the environments where researchers already work.
Coding offers particularly direct evidence of the supplier becoming an application competitor. In the same announcement that introduced Claude Code, Anthropic highlighted how Cursor, Cognition and Replit used or evaluated Claude. It was simultaneously promoting its ecosystem and launching its own tool for delegated engineering work. OpenAI’s Codex launch similarly put a software engineering agent in developers’ hands.
These examples represent different degrees of expansion. An industry solution, a broad business workspace and a coding agent are not interchangeable. Together, however, they show why founders cannot assume a permanent boundary between the model supplier and the application vendor.
Forward deployed engineers change the competitive equation
The strongest version of this threat involves people as well as products. A startup might reasonably argue that its advantage comes from understanding the customer’s messy processes and making AI work inside them. That advantage becomes harder to defend when the model provider puts engineers alongside the same customer.
In its February 2026 Frontier announcement, OpenAI described a platform for building, deploying and managing enterprise agents. It explicitly pairs customers with forward deployed engineers and describes a feedback loop connecting deployment experience with its research teams.
My inference is that this proximity can become a powerful commercial advantage. Working inside an organisation helps a supplier understand which problems matter, where projects stall and what a customer will pay to resolve. Repeated across customers, those lessons can inform reusable products. This does not require training on confidential customer data; implementation experience itself has value.
The model provider can therefore compete for part of the work that a startup expected to own: identifying the use case, integrating systems and getting the application into production. “We understand the workflow” becomes a claim that needs evidence of a durable lead.
The competitive set has three fronts
Founders should assess their position against three different alternatives:
| Competitive pressure | What the customer can do | What the startup needs to prove |
|---|---|---|
| Other startups | Buy a similar application built on widely available models | A sustained advantage in outcomes, distribution or accumulated expertise |
| Internal development | Assemble a narrower solution using coding agents and existing systems | Enough value to outweigh the full cost of building and maintaining it |
| Model providers | Use a direct product, bundled capability or supported deployment | Value that survives when the core AI capability becomes readily available |
The internal development threat deserves particular attention. Coding agents make it easier to explore whether a workflow can be built internally. The customer does not need to reproduce your entire company. It needs something adequate for its own users, data and process.
For lightweight tools, that can change the build-versus-buy conversation substantially. A customer may accept a less polished internal application if it solves the immediate problem and avoids another procurement process.
However, a working prototype is only part of the cost. Permissions, security, monitoring, evaluation, support and maintenance still need owners. The relevant comparison is the lifetime cost and reliability of the workflow. A startup that demonstrably handles those responsibilities has a stronger argument than one whose principal advantage is that it wrote the code first.
Being vertical is a starting point
There is a meaningful difference between an application that generates an industry-specific document and one that manages the process surrounding it.
Consider an insurance example. Summarising a submission is a task. Supporting underwriting across intake, missing information, policy rules, approvals, audit records and exceptions involves a much broader operational commitment. A better model can improve both, but replacing the second requires much more than reproducing a good answer.
That does not make complex workflows immune to competition. It gives founders a more useful place to look for defensibility:
- Workflow ownership. The product is embedded in recurring decisions and actions, with meaningful consequences if it is removed.
- Rights to useful data and feedback. Usage creates legally usable information about outcomes and exceptions that improves the product. Merely accessing a customer’s documents is a weaker claim.
- Proven reliability. The company can demonstrate performance on the customer’s actual workload, including failures, escalation and human review.
- Distribution and trust. The business has a repeatable route to buyers and relationships that a new entrant cannot acquire simply by releasing a feature.
- Compounding implementation knowledge. Each deployment improves reusable software and reduces the effort required for the next one.
The last point matters for startups adopting the FDE model themselves. Deployment teams can accelerate learning and revenue. The test is whether that work creates a better product across customers or leaves the company supporting an expanding collection of bespoke projects.
Partnership and competition can coexist
There is also evidence against the simplistic conclusion that model providers will absorb every application. OpenAI’s Frontier launch named Abridge, Ambience, Clay, Decagon, Harvey and Sierra as partners. Anthropic’s life sciences offering connects to specialist platforms such as Benchling. Those ecosystems leave room for independent businesses, even as the platform’s scope expands.
A partnership can bring distribution, functionality and enterprise access. It is still worth asking who controls the customer relationship, which capabilities are becoming standard, and how much bargaining power remains with the application vendor.
Similarly, supporting several models can reduce dependence on a single supplier and improve cost or performance. It does not by itself protect an application from substitution. A customer buying an outcome rarely cares how many providers sit behind it.
For founders and investors, I would bring the discussion back to three questions. If the model provider bundled our headline feature tomorrow, why would customers keep paying us? If a customer built an adequate internal version, what would they still need us to operate? And after another year of deployment, what will we possess that a well-funded competitor cannot quickly reproduce?
The opportunity to build valuable AI applications remains enormous. Better models expand what small teams can deliver. But they also expand what competitors, customers and infrastructure suppliers can deliver.
The model should make your product better over time. Your business needs to accumulate reasons for customers to stay that the next model release cannot supply.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- Your AI infrastructure provider wants your customers - September 17, 2026
- Who Keeps AI Safe When Everyone Has an Incentive to Go Faster? - September 15, 2026
- Weekly Firgun Newsletter – September 11 2026 - September 11, 2026
