“The question I ask: If OpenAI or Anthropic launches a model tomorrow and is 10x better, does this company still have a reason to exist?”
– Jake Flomenberg, Partner at Wing Venture Capital
As we step into 2026, the talk about an AI bubble is intensifying. As many smarter folks than me have put it, we may be in an LLM bubble, not an AI bubble. There’s no doubt that AI and automation can both increase the productivity and potentially save costs for organisations.
For founders and investors alike, the narrative has moved beyond “Will AI work?” to “How does AI get to work?” The answer, overwhelmingly, is Vertical AI.
We are no longer just digitising generic processes; we are reimagining the systems of work for specific industries. Drawing on the latest data from Bessemer, A16Z and others, here is the state of Vertical AI in 2026: the good, the bad, and the ugly.
The Good: Tapping into the Labor P&L (The 10x Opportunity)
“Vertical AI isn’t competing for IT budgets; it’s competing for labor budgets. Unlike vertical SaaS, which typically captures a fraction of Fortune 500 IT spend, Vertical AI taps directly into the labor line of a P&L.”
– Bessemer Venture Partners
The traditional software playbook involves selling into a company’s IT budget. Vertical AI, however, is playing a much bigger game. The true opportunity for Vertical AI lies in automating high-cost, language-intensive tasks currently performed by humans, which means it’s competing for labor budgets, not software budgets.
The Economic Unlock
Legacy vertical SaaS was about efficiency. Vertical AI is about enablement. Business and professional services, dominated by repetitive, language-heavy tasks like legal discovery, clinical notes, and audit workflows, account for 13% of US GDP. That is roughly 10x the size of the software market today. Unlike traditional SaaS, which captured a fraction of IT spend, Vertical AI taps directly into the labor line of the P&L. This implies that software will replace these jobs, but it’s more likely going to enhance, rather than fully replace labor (in the short term).
We will continue to see co-pilots for every role in the organisation: developer, CRO, customer support, marketing etc. But at least at the first stage, these tools are about enhancing, not replacing the employees.
Workflow automation is the moat
As models become more accessible and their capabilities converge, their performance alone stops being a reliable moat. True defensibility in the AI era comes from something far more durable: deep workflow integration, as I shared in my previous post on vertical integration.
a16z calls it the “collaboration layer” that orchestrates multi-agent work. The goal isn’t a marginally better model, but a product so deeply embedded that it becomes the system of record for a critical business operation.
Multimodality
Text was just the appetiser. 2026 is the year AI went multimodal, unlocking physical industries that SaaS previously couldn’t touch.
- Vision & Voice: We are seeing AI that can “step into” video, understanding physics and time rather than just pixels. In construction, drones and robotics are interpreting blueprints and site photos to generate scope-of-work packages automatically,.
- Voice Agents: The “conversational voice stack” has matured. We aren’t just transcribing; we are fielding inbound sales calls for home services and handling patient intake without human intervention.
The Bad: The “Infrastructure Shock” and Data Entropy
While the revenue potential is massive, the engineering reality is brutal. Building a defensible Vertical AI business in 2026 requires solving problems that didn’t exist two years ago.
Agent-Native Infrastructure Crisis
The enterprise backend was built for humans clicking buttons, not agents triggering 5,000 recursive sub-tasks in milliseconds. Legacy systems view these “agent-speed” workloads as DDoS attacks. Founders need to build or buy “agent-native” infrastructure that handles thundering herd patterns, shared state management, and complex locking.
Take coding as an example, an organisation with 200 developers using Cursor can very quickly lose track of what code gets included in production. We need more tools to supervise the AI and co-pilots. (Disclaimer: Remagine Ventures is an investor in Pandorian, which addresses this problem).
Data Entropy
Enterprises are drowning in unstructured, multimodal sludge: PDFs, screenshots, and logs. This “data entropy” is the primary bottleneck for reliability. The winners in 2026 aren’t just training models; they are building platforms to clean, structure, and validate this chaos so that downstream agents don’t hallucinate.
Businesses need deterministic outcomes from LLMs/ Agents. Businesses are learning the hard way the downside of RAGs, LLMs and Agents, which is slowing wider adoption in the enterprise and keeping a lot of the deployments in the POC/experiment stage.
The Talent Gap
Despite the hype, 61% of firms still report a lack of experience with AI governance tools. Implementation costs remain high, and there is a shortage of skilled professionals who can manage AI model risk,. Founders cannot just sell software; they often have to sell the implementation and “forward-deployed” engineering to make it stick. This is something that I assume will be solved with time.
The Ugly: The Death of Metrics & The “Wrapper” Trap
The ugly is that there’s a lot of funded AI startups that are nothing but a thin wrapper on GPT or other LLMs. While they may have attracted some early revenue or even seen momentum, I can’t imagine them being able to continue on the venture path as they get commoditised by the models themselves. Startups that are thin wrappers may therefore find themselves in a pickle in 2026 and will need to pivot, sell or shut down. To survive they will have to tackle an end-to-end workflow with “LLM magic” that was impossible for humans to do alone, protected by a moat of proprietary data and deep workflow integration.
RIP Screen Time
For 15 years, we measured software value by engagement: time on site, daily active users (DAU). In Vertical AI, screen time is a vanity metric. If a doctor uses Abridge to automate clinical notes, they spend less time looking at a screen and more time with patients. That is the value. If an investment banker uses Hebbia to draft a pitch deck, they get to sleep, not click,. Founders must transition to outcome-based pricing models, charging for the work done (e.g., per demand package generated), not the seat. This is a fundamental change that will make outcome-based AI startups look much cheaper alternatives to SaaS /seat model legacy companies.
The Vendor Sprawl Backlash
CIOs are exhausted. They are pushing back on “vendor sprawl” and cutting experimentation budgets. We are seeing a bifurcation: budgets are increasing, but they are concentrating on a few winners who can prove hard ROI, while “nice-to-have” tools are seeing revenue contract. This is good news for Google, perhaps Salesforce and other established vendors (whether it’s cloud or SaaS) that have the benefit of trust and have already done the hard part of becoming an approved vendor for enterprise clients. They will ‘land and expand’ while startups fight for adoption.
What does this all mean for Vertical AI founders in 2026?
If you are building a Vertical AI startup today, here is your roadmap:
- Don’t Replace, Delegate: Aim for “progressive delegation.” Start by automating the low-value, high-volume slice of a workflow. Keep the human in the loop for high-stakes decisions to build trust, then expand.
- Find the “Magical Feature”: You need a “miracle-like” advancement to break in. Whether it’s analyzing 1,000x more data than a human could (e.g., Rilla for sales coaching) or automating a tertiary workflow like legal research, use that wedge to earn the right to expand into core workflows.
- Go Multiplayer: Vertical work is multi-party. The next evolution is agents collaborating across organizations—buyers, sellers, regulators. The moat isn’t just the model; it’s the collaboration layer.
- Price for Outcomes: Move away from pure SaaS pricing. Consider hybrid models: a base subscription for predictability plus usage tiers based on outcomes (e.g., per resolution, per document) to capture the upside of your AI’s labor.
Vertical AI in 2026 is no longer about the novelty of the technology; it is about the integration of intelligence into the economy. The founders who win will be those who possess the “insider” empathy to understand the nuance of a workflow and the “outsider” audacity to reimagine it entirely.
For further reading I also recommend Bessemer’s ‘Vertical AI playbook‘.
I'm a former general partner at google ventures, head of Google for Entrepreneurs in Europe and founding head of Campus London, Google's first physical hub for startups.
I'm also the founder of Techbikers, a non-profit bringing together the startup ecosystem on cycling challenges in support of Room to Read. Since inception in 2012 we've built 11 schools and 50 libraries in the developing world.
- Weekly Firgun Newsletter – February 27, 2026 - February 27, 2026
- The Rise of the Personal AI Agent - February 26, 2026
- Pre-Seed Is Stronger than ever, but less startups are getting funded - February 25, 2026
