From Demos to Deals: Insights for Building in Enterprise AI

AI has become a strategic priority for virtually every enterprise: OpenAI claims 10% of the world’s systems now use their products, and many Fortune 500 companies have adopted CEO-led mandates to integrate AI.

Builders are capitalizing on this incredible demand, but AI companies behave differently from traditional SaaS businesses, and much of the common wisdom about what worked for SaaS doesn’t seem to hold true for AI. When we meet founders, they consistently have questions around what great AI companies look like, where to expect product commoditization, and how to build a company with enduring value. Based on those conversations, we’ve synthesized a few takeaways about how enterprise AI startups are adapting, growing, and breaking out.

1. Flashy demos are easy. Substantive products are hard.

After ChatGPT launched in 2022, a common refrain was that all AI software would get commoditized because they were “GPT wrappers” — implying that these products were trivial to build and would easily be subsumed by the rapidly improving model capabilities. 

Almost three years later, it’s clear that’s not the case. Creating a flashy AI demo is relatively simple with modern tools, but the last mile of product work is exceptionally difficult. In the real world, users behave unpredictably, customer data is messy, and success depends on handling the long tail of paths a user might take. 

Demos have always been simplistic versions of a final product, but the demo-to-product gap is even wider with AI products because of the constantly evolving models and their nondeterministic outputs. Incidents like Air Canada’s support bot hallucination illustrate the challenges of deploying AI in real enterprise production environments, and the financial and reputational risks of getting it wrong. This is all the more true in domains where accuracy and trust are the top priority, such as accounting and legal. You can’t vibe code your accounting statements or your legal docs. 

As a result, AI companies have become highly sophisticated at both maximizing the capabilities of state-of-the-art models and constraining them for enterprise-grade reliability. They rigorously run evals on the latest models, orchestrate sequences of actions across different models, build substantial scaffolding on top of the base models, and set a product roadmap that threads the needle between what is possible today and what will be possible tomorrow. Teams often toggle between models based on how well they perform against a specific task, and have to make trade-offs based on model quality, cost, speed and scalability. In many cases they also fine-tune their own smaller models that live alongside the larger models in production. The result is a robust product experience that no single API call can deliver.  

In addition, for any AI product to be valuable, it needs to understand the context and logic of the business it’s serving. Models do not do that out-of-the-box. As a result, AI companies are investing meaningful engineering and implementation resources into getting their products to work within the unique policies, culture, and systems of each individual customer instance. It’s gritty, in-the-weeds customer work that needs to get done, but that the horizontal model companies don’t and likely won’t do themselves. 

Given the complexities of orchestrating across a rotating cast of models, understanding how to build toward what will soon be possible, and solving the long tail of challenges required to make AI work in real-world production environments, there’s ample surface area to build a sustainable applied AI business — one that isn’t easily commoditized by the horizontal model providers.

2. It takes more than ever to break out: 10x is the new 3x.

Hitting $1 million in ARR in the first 12 months of monetization used to be a classic benchmark for enterprise startups to raise a successful Series A. But based on the sample set of Series A AI companies we’ve seen over the last 18 months, that’s now below the median. Similarly, based on underlying payments data, Stripe has said that AI customers are hitting $5 million in ARR at dramatically faster rates than their historical SaaS counterparts. AI companies just grow faster.

The fastest growing AI software companies today are growing more than 10x year over year, and there are even outliers beyond that, with companies like Cursor becoming one of the fastest growing software companies of all time through a product-led go-to-market.

Why is this happening? First off, we’ve seen a marked shift in enterprise buying behavior. The AI value proposition is evident to enterprises and has created a large cohort of eager AI software buyers with dedicated AI budgets and mandates. That’s bringing forward sales cycles and has consequently accelerated the top-line growth for these AI-native companies. In other words, buyers are actively seeking and pulling AI software into their organizations, unlike previous generations of software that often required a sales push. 

Secondly, when enterprises do buy, they’re spending more than they spent on software in the past. This is because AI software often sells the work output itself, instead of selling software to help the people do the work. This means that AI companies are replacing labor budgets, whose size exceeds traditional software budgets, giving AI companies meaningfully larger contract sizes than their historical software counterparts. 

3. The barrier to entry has gone down: expect a flood of applications.

The cost to create is plummeting — from $30 to under $5 per million tokens in under two years. Just this month, OpenAI dropped the price of its o3 model by 80%. This “LLMflation” outstrips even the historic declines in compute costs during the PC revolution and bandwidth during the dotcom boom. 

At the same time, agentic IDEs like Cursor (which dramatically accelerate developer productivity) and text-to-app platforms like Lovable and Replit (which empower non-technical users to build apps with natural language) have ushered in product categories that are dramatically accelerating the pace of new app creation. 

These two factors have already begun collapsing the cost and complexity of software creation, unlocking the possibility to create a long tail of tools that were previously uneconomical to build, and empowering a growing population of less technical users. 

Some of this software will be personal software for a market of one, like custom vacation planners or personal health dashboards. Others will add tangible enterprise value. Enterprises will finally be able to productize the long tail of workflows and edge cases that off-the-shelf software couldn’t serve, and that consequently had to be papered over with human labor or brittle RPA solutions. With AI, there’s a way to turn that patchwork system into dedicated tools. AI is unlocking a vast set of new opportunities — creating new markets, unlocking categories that used to be dominated by human labor, and expanding markets that were historically too small. 

4. Speed matters more than ever.

Dozens of companies are trying to sell identical-sounding solutions to the same enterprise buyer, and these buyers are inundated. They want to buy AI from credible players, and there’s an incredible amount of value to being the first to establish yourself as that provider. 

To do that, speed and momentum matter. We’ve seen early momentum propel companies to become category leaders faster than ever before. Early movers have been able to scale product surface area, generate word of mouth, land major customers, achieve rapid revenue growth, and establish brand dominance — often before fast followers have had a chance to adequately respond. 

In coding, Cursor’s rapid product velocity has catapulted it into so much of a brand that it’s now used as a standalone noun, with large companies like Canva even expecting proficiency to land technical roles. Similarly, companies like Decagon, ElevenLabs, Hebbia, and Harvey have been able to leverage early momentum to establish a premier enterprise brand, compounding customer momentum in their respective markets.

This speed is also what has allowed these companies to stay ahead of the curve against legacy incumbents and the model providers. In many enterprise categories today, software incumbents and the model providers have actually launched similar offerings to the leading startups. In coding, for instance, Microsoft has GitHub Copilot and OpenAI has Codex, but that hasn’t stalled Cursor’s momentum. Incumbents often split their focus across their core legacy business or their many competing research and product priorities, and as a result, AI-native startups that stay laser-focused on shipping the best product still gain wide adoption.

5. To sustain that early advantage, moats still matter.

Pure execution speed and shipping velocity is one way to break out, but companies need to sustain that advantage to close the door to fast followers and legacy incumbents. AI itself is not a moat: it is a way to deliver value to customers. 

In the fiercely competitive AI market, companies still need to be building enduring products. Moats still matter. They can do so in a few ways.

Become the system of record: The most classic moat in enterprise software is becoming an organization’s core system of record, or the source of truth for critical data. AI has opened up exciting new opportunities in different markets (especially vertical markets), because it’s a great way to get customer velocity and demonstrate clear value in a short timeframe. But ultimately the system of record is still a dominant business model to ensure enduring value. 

That doesn’t mean it’s not worth pursuing AI wedges. Several AI companies, such as Eve, Salient, and Toma are using AI wedges to capture data at the point of creation (through voice calls or ingesting unstructured data) in ways that traditional software couldn’t. They are then building downstream workflows from that point of creation, with the ultimate goal of maturing into the core system of record within their industries. 

Create workflow lock-in: Once a company is able to get users to embed a product into their daily workflows, it creates operational and behavioral muscle memory that makes switching psychologically and culturally disruptive. While it’s common to say that AI software is doing the work autonomously and therefore lacks classic GUI-based user interfaces, there’s still a heavy human-to-AI interaction loop, as humans are still often overseeing and auditing the work being executed by AI today. Decagon, for example, builds AI agents that autonomously deflects support tickets, but still has robust product workflows for humans to monitor, tweak, or analyze the work these AI agents are doing, and clever ways to escalate to human agents when needed. Getting users comfortable with their product’s UI/UX has created powerful workflow moats, making customers unlikely to switch to new tools.

Build deep vertical integrations: Many enterprise customers live in and depend on a complicated web of business software systems, many of which have limited APIs and don’t speak to each other. As a result, AI companies hoping to serve these enterprises must often deeply connect and integrate into these bespoke systems to drive value. Investing in these integrations as a first-class effort helps build durability in the customer base by embedding the product into a customer’s core operational workflow, making it difficult to rip out without disrupting other systems.

In healthcare, for instance, Tennr has invested heavily into integrating with a long tail of legacy fax and healthcare systems to streamline pre-patient referrals across providers. In logistics, HappyRobot builds connections into homegrown TMS (trucking management system) platforms to deploy its AI voice agents for freight call operations. And Glean drives much of its value from critical integrations across core enterprise tools. These deep, often customer-specific integrations are key to making AI solutions work in the real world and can serve as powerful competitive moats.

Entrench customer relationships: Enterprise buyers are still human. Trusted relationships often matter more than any specific feature or vendor price, especially as AI vendors increasingly act as strategic AI thought partners to the customers they serve. Many AI companies now have the buyer’s ear in ways traditional software vendors rarely did, helping shape customer roadmaps and AI strategy, not just tool purchases.

*  *  *

This is one of the most exciting times to be a builder, as AI unlocks new markets and expands existing opportunities. These observations reflect what we’re seeing across AI startups today and are meant to guide founders aiming to build enduring companies.