For most of 2022 and 2023, the dominant narrative in enterprise technology was that generative AI was all promise and no delivery. Boards were mandating AI strategies. Chief Digital Officers were commissioning pilots. Consulting firms were billing enormous fees to produce AI readiness assessments. And then — largely — nothing happened. The pilots ran, generated impressive demos, and got shelved when the question of production deployment proved harder than anticipated. Enterprises were enthusiastic about AI in theory and cautious about it in practice.
2024 is when that changed. The shift has been quiet — there are no press releases announcing that a Fortune 500 company has graduated its AI applications from pilot to production — but in our conversations with enterprise buyers, operators, and the founders building AI applications for them, we have observed a clear and unmistakable transition. Enterprise AI has crossed the chasm. The question is no longer whether AI can do useful things in enterprise contexts. The question is which AI applications are winning the production deployment competition, why they are winning, and what the emerging market structure looks like.
This piece is our attempt to answer those questions through the lens of four companies — Harvey, Glean, Sierra, and Cognition — that we believe represent the leading edge of enterprise AI adoption. They are not identical businesses; they serve different functions, different buyers, and different industries. But they share a set of characteristics that, in our view, explain why they have been able to move from "interesting pilot" to "company standard" faster than their competitors, and why they command the valuations they do.
The 2022-2023 Pilot Trap and Why Enterprises Got Stuck
To understand the significance of the 2024 transition, it helps to understand why enterprises got stuck in the pilot phase for so long. The reasons are structural, not cultural — and many of them apply equally well to the next generation of AI applications, which is why understanding them matters for founders building AI products today.
The first reason was hallucination and reliability. The language models available in 2022 and most of 2023 were impressive in their breadth but unreliable in their accuracy. For most enterprise use cases — legal research, financial analysis, customer-facing communications, internal knowledge retrieval — a system that is right 90% of the time and confidently wrong 10% of the time is not useful. It is actively dangerous, because it gives users false confidence in outputs that occasionally contain material errors. Enterprise buyers who ran pilots discovered quickly that their employees could not distinguish high-confidence correct outputs from high-confidence incorrect ones, and no enterprise is willing to deploy a system in production that requires every output to be manually reviewed for accuracy.
The second reason was integration complexity. Enterprise AI applications do not operate in isolation. A legal AI system needs to integrate with the law firm's document management system, its billing system, its client portal, and its conflict-check database. An enterprise search system needs to integrate with SharePoint, Salesforce, Slack, Google Drive, Confluence, Jira, and a dozen other systems where knowledge lives. In 2022 and 2023, most AI application vendors underestimated this integration complexity and delivered products that worked well in demos but required months of custom integration work before they could operate in a real enterprise environment.
The third reason was change management and workflow fit. Enterprise employees are resistant to changing workflows that work adequately. AI applications that required users to context-switch to a new interface, learn new behaviors, or break established patterns faced significant adoption friction even when the underlying AI capability was genuinely superior to the existing manual process. The AI applications that succeeded in production were those that fit seamlessly into existing workflows rather than requiring workflows to reorganize around the AI.
The breakthrough in 2024 has come from a new generation of AI application companies that solved all three of these problems — not by waiting for AI to improve (though it has improved), but by engineering their products specifically to navigate these enterprise adoption challenges.
Harvey: Legal AI That Law Firms Actually Deploy
Harvey was founded in 2022 by Winston Weinberg and Gabe Pereyra, a former lawyer and a former Google Brain researcher. The company's core product is an AI system for legal work — contract review, legal research, document drafting, and regulatory analysis — built specifically for law firms and legal departments rather than as a general-purpose AI writing assistant with a legal marketing overlay.
Harvey raised $100 million in a Series B in February 2024, at a valuation that made it one of the most valuable legal tech companies despite being only 18 months old. The round was led by Kleiner Perkins with participation from OpenAI and a number of prominent legal industry executives who serve as both investors and customer advocates. At the time of the announcement, Harvey disclosed customer relationships with Allen & Overy (now A&O Shearman, one of the world's largest law firms), PwC, and several other major professional services firms — a set of reference customers that would have been unattainable for most two-year-old enterprise software companies.
Why did Harvey succeed where dozens of legal AI companies that preceded it failed? Several reasons stand out in our analysis. First, Harvey built reliability architecture specifically for legal contexts. Rather than applying a general-purpose language model directly to legal questions and hoping for accurate outputs, Harvey built a system that grounds its answers in specific legal documents and jurisdictional precedents, provides citations for every substantive claim, and explicitly flags areas of uncertainty rather than filling them in with confident-sounding guesses. This citation-grounded, uncertainty-explicit approach addresses the hallucination problem that made earlier legal AI systems unusable for professional work.
Second, Harvey built deep workflow integration from day one. Their system integrates with the document management systems that large law firms use (iManage, NetDocuments), the research platforms that lawyers depend on (Westlaw, LexisNexis), and the billing systems that law firm management requires. This integration work is expensive and slow, which is why most legal AI startups avoided it — but it is exactly what converts a demo into a production deployment. A lawyer who can run Harvey directly from within their existing document workflow, with results surfaced in the context where they are already working, will actually use it. A lawyer who has to log into a separate system and manually transfer outputs will not.
Third, Harvey priced and packaged for law firm economics. Law firms are accustomed to paying substantial per-seat license fees for research and document management tools. Harvey's pricing reflected this expectation — it was not cheap, which paradoxically made it easier to sell. An AI system priced at $50 per user per month reads as a consumer product and raises concerns about data security and institutional commitment. An AI system priced to fit into the existing software line items on a law firm's budget reads as a professional tool and gets evaluated by the appropriate enterprise software procurement process.
Glean: Enterprise Search and Knowledge Management
Glean was founded in 2019 by Arvind Jain and a team of engineers from Google, and went through several product iterations before landing on the product that has driven their exceptional growth: an AI-powered enterprise search and knowledge management system that indexes content across all of a company's SaaS applications and internal knowledge bases and surfaces relevant information in response to natural-language queries.
The company raised $200 million in February 2024 at a $2.2 billion valuation, with investors including Altimeter Capital, Lightspeed Venture Partners, and Sequoia Capital. At that point, Glean had enterprise customers including Grammarly, Vanta, Okta, and Duolingo, as well as a substantial presence in the financial services and technology sectors. Their annual recurring revenue trajectory and net retention rates — both of which the company has discussed publicly in general terms — reflect the strongest metrics in the enterprise knowledge management category.
Glean's core insight was that the knowledge management problem in enterprises was not being solved by any existing system. Individual applications — SharePoint, Confluence, Google Drive, Notion, Salesforce — each index and search their own content, but there is no system that a new employee can go to and ask "what is our policy on customer data handling?" and get a comprehensive answer drawn from all the places where that policy might be documented. The result is that knowledge discovery at most enterprises is a social problem — people who have been at the company for years know which person to ask, which document repository to check, which Slack channel might have the answer — rather than a systematic search problem.
Glean addressed this by building the integration infrastructure to connect to all of the major enterprise SaaS applications, establishing the access control architecture to ensure that search results only surface content the querying user is authorized to see, and building the AI layer that understands the semantic meaning of queries and can surface the most relevant content from across the entire enterprise knowledge graph rather than just returning keyword matches.
What makes Glean particularly interesting from an investment perspective is the compounding value it creates as it learns about a specific enterprise's knowledge landscape. After six months of deployment, Glean's system understands which documents are most authoritative for which types of questions, which employees are subject matter experts in which areas, and how knowledge is distributed across the various repositories. This learned knowledge graph becomes a significant switching cost — a company that has used Glean for two years has built up a semantic model of their institutional knowledge that would be expensive and time-consuming to replicate with a new system.
Sierra: The Conversational AI Platform for Customer Experience
Sierra was founded in 2023 by Bret Taylor — formerly co-CEO of Salesforce and Board Chair of Twitter — and Clay Bavor, a long-time Google executive who ran Google Labs. The company builds conversational AI systems for enterprise customer experience — specifically, AI agents that handle customer support, customer onboarding, and customer service interactions for large enterprises in a manner that is both more capable than traditional scripted chatbots and more reliable and brand-consistent than general-purpose language model deployments.
Sierra raised $110 million in a Series B in February 2024, at a reported valuation of approximately $1 billion. Their early customer base included SiriusXM, Sonos, and WeightWatchers — companies with large customer service operations that were looking for ways to handle increasing contact center volume without proportional headcount growth. The investment and customer traction were remarkable given that the company was less than a year old at the time of the funding announcement.
Sierra's approach to the enterprise AI adoption problem is instructive. Rather than building a general AI assistant and selling it to customer experience teams, they built a platform specifically designed to give enterprises precise control over the behavior, tone, and capabilities of their AI agents. Enterprise buyers in the customer experience space have legitimate concerns about deploying AI agents that might say something inconsistent with brand guidelines, make commitments the company is not willing to honor, or handle sensitive customer situations (complaints, safety concerns, billing disputes) in ways that create legal or reputational risk.
Sierra addressed these concerns by building a sophisticated system for defining and enforcing behavioral guardrails — essentially, a way for enterprises to specify exactly what their AI agent can and cannot do, in enough detail that the system behaves predictably across millions of customer interactions. A Sierra-powered customer service agent for a consumer electronics company knows that it can offer a replacement for a defective product under warranty but cannot authorize returns for products outside the return window without escalating to a human agent; it knows the brand voice guidelines that govern how to discuss competitor products; it knows when a customer interaction has emotional signals that warrant immediate escalation to a senior human representative.
This control layer is what makes Sierra deployable at enterprise scale in customer-facing contexts. A company that interacts with millions of customers via AI cannot afford to discover, post-deployment, that its AI agent occasionally gives inaccurate information, makes unauthorized promises, or handles sensitive situations poorly. Sierra's product is designed to make that discovery impossible by giving enterprises the tools to specify and verify agent behavior before deployment.
Cognition: AI Software Engineering Agents
Cognition was founded in 2023 by Scott Wu, Steven Hao, and Walden Yan — a team with an extraordinary pedigree in competitive programming who had previously worked together at Google DeepMind. The company's first product, Devin, garnered significant attention in March 2024 when Cognition published a benchmark demonstrating that Devin could autonomously complete software engineering tasks — writing code, running tests, debugging failures, and iterating based on test results — at a level that significantly exceeded any previously published AI system.
Cognition raised $175 million in April 2024 at a $2 billion valuation, with Founders Fund leading the round. The round was extraordinary not just for its size but for the speed at which it came together — a reflection of how clearly Devin's capabilities demonstrated the potential of autonomous AI software engineering agents to address one of the most significant bottlenecks in enterprise technology: the cost and scarcity of software engineering talent.
What makes Cognition interesting as an investment case study is that it represents a category — AI software agents that autonomously execute complex knowledge work tasks — that is qualitatively different from AI assistants that augment human work. Earlier generations of AI coding tools (GitHub Copilot, Tabnine) operate as autocomplete systems that accelerate individual human developers. Devin operates as an autonomous agent that can take a task specification, break it into component subtasks, write and test code iteratively, debug failures, and deliver a working solution — without requiring a human to supervise each step in the loop.
The enterprise adoption implications of this shift are profound. An AI system that accelerates individual developers by 30% is a productivity tool. An AI system that can autonomously complete defined software engineering tasks is a labor input that can scale independently of human headcount. Enterprises that successfully deploy autonomous AI software engineering agents can address their software engineering backlogs, accelerate time-to-market for new features, and reduce the human capital required to maintain and extend existing systems — all without the hiring, onboarding, and management overhead associated with human engineers.
Whether Cognition specifically builds the dominant autonomous software engineering platform remains to be seen — the company faces competition from well-funded rivals and from the foundation model providers themselves, all of whom are investing heavily in coding capabilities. But Cognition's fundraising and early customer traction validate a category that we believe will generate multiple large companies over the coming decade.
Five Patterns That Separate Successful Enterprise AI from Failed Pilots
Looking across Harvey, Glean, Sierra, Cognition, and the broader set of enterprise AI companies that have successfully made the transition from pilot to production, we observe five consistent patterns that separate the winners from the companies still stuck in pilot purgatory.
Narrow scope with deep quality. Every successful enterprise AI deployment we have observed started with a narrow, precisely-defined use case where the AI system could achieve a level of quality — in accuracy, consistency, and reliability — that was genuinely better than the existing human process. Harvey started with contract review, not all legal work. Glean started with enterprise search, not all knowledge management. The companies that tried to build AI systems that could do everything for enterprise users found that the quality bar was too hard to maintain across a broad scope. Starting narrow and expanding from a position of demonstrated quality is the correct sequencing.
Workflow integration over standalone products. Enterprise AI applications that require users to leave their existing workflows to access AI assistance have dramatically lower adoption rates than those integrated into existing workflows. The most successful enterprise AI products operate through plugins, integrations, and APIs that surface AI capabilities in the context where users are already working — inside Slack, inside Google Workspace, inside the CRM, inside the code editor — rather than asking users to navigate to a separate AI interface.
Deterministic behavior through system design, not model instruction. Enterprise buyers do not trust that language models will reliably follow natural-language instructions at production scale. They trust systems that enforce behavioral constraints through architecture — access controls, output validation, grounding in authoritative sources, explicit escalation logic. The enterprise AI companies that have successfully deployed at scale are those that have built deterministic control layers on top of probabilistic foundation models, so that the behavior of the system as a whole is predictable even when the underlying model's behavior has statistical uncertainty.
ROI transparency from the first meeting. Enterprise buyers in 2024 have been burned by AI pilots that generated impressive demos but unclear business value. The companies winning new enterprise deployments are those that can quantify the value they deliver in terms of hours saved, headcount reduction, error rate improvement, or revenue impact — and can point to specific, named customers who have experienced that value. Generic ROI claims ("AI can reduce your legal costs by 30%") are ineffective. Specific ROI stories ("Allen & Overy reduced contract review time by 50% for standard commercial agreements in a 90-day pilot") are compelling.
Enterprise-grade trust infrastructure. Data security, privacy compliance, access control, and audit logging are table stakes for enterprise AI deployment, and they are consistently underestimated by founders who have not sold into large enterprises before. An enterprise that stores sensitive customer data, confidential business information, or regulated financial data in its knowledge base will not connect that data to an AI system that does not meet its data handling and security requirements. Building the SOC 2 certification, GDPR compliance architecture, and enterprise security controls required to pass Fortune 500 security review is expensive and time-consuming, but it is the prerequisite for meaningful enterprise revenue.
The Investment Implications: What We Are Looking For
The enterprise AI adoption transition of 2024 has significant implications for how we think about investing in AI application companies at the seed stage. The bar has risen considerably from the frothy market of 2021-2022, when any AI-related pitch could command a generous valuation based on the market opportunity alone. Enterprise buyers are now experienced enough to distinguish between AI applications that will reach production deployment and those that will stall in the pilot phase — and they are sharing those assessments with peers in ways that can accelerate or destroy enterprise software companies' reputations faster than ever before.
At Milestone AI Ventures, we have updated our investment criteria for AI application companies in light of these market dynamics. We look for founders who have direct, first-hand knowledge of the enterprise workflow they are addressing — through prior work experience, domain expertise, or unusually deep customer discovery. We look for product architectures that are designed from the ground up for enterprise reliability requirements, not consumer AI applications with an enterprise pricing tier bolted on. We look for teams that have done the integration work required to deploy their product in a real enterprise environment, not just a demo environment — and who understand the difference.
We are particularly interested in the emerging wave of AI applications being built for vertical industries where the combination of specialized data, domain-specific workflows, and regulatory requirements creates defensible positions that horizontal AI application companies cannot easily contest. Healthcare, legal, financial services, insurance, and industrial settings are all areas where deep domain expertise creates a genuine competitive advantage for AI applications, and where the category leaders are being established now.
Looking Ahead: The Enterprise AI Market in 2025 and Beyond
The transition from experiment to deployment that is underway in 2024 is the beginning, not the end, of a multi-year adoption cycle. Most enterprises have made their first meaningful AI deployments this year, but those deployments represent a small fraction of the AI applications that will eventually be integrated into enterprise operations. The companies that are winning today — Harvey, Glean, Sierra, Cognition — are establishing category leadership in specific enterprise functions, but dozens of other enterprise functions are still waiting for AI applications that are production-ready.
We expect the pace of enterprise AI adoption to accelerate through 2025 and 2026, driven by three reinforcing trends. First, the reference customer base for enterprise AI is growing rapidly — as more Fortune 500 companies make AI deployments that demonstrably work, the social proof barrier to adoption at peer companies drops. Second, the workforce that is entering enterprise organizations is the first generation to have grown up with AI tools as a standard part of their academic and professional toolkit, creating demand from below in organizations for the AI tools they are already comfortable with. Third, competitive pressure between enterprises is beginning to manifest in AI deployment speed — in industries where AI-powered efficiency improvements are material, companies that have not deployed effective AI systems will find themselves at a structural cost and speed disadvantage relative to those that have.
For founders building AI applications for enterprise, the message of 2024 is clear: the market is real, the adoption is happening, and the companies that are winning have solved the hard problems of enterprise reliability, workflow integration, and trust infrastructure. The opportunity is large and the window for establishing category leadership is real — but so is the competition. Building a company that wins in enterprise AI requires both excellent AI technology and excellent enterprise software instincts, and teams that have both are the ones we want to back.
If you are building AI applications for enterprise — in legal, finance, customer experience, software engineering, or any other function where AI can meaningfully improve enterprise workflows — we would like to hear from you. Reach us at founders@mstone-ai.com or through the contact form on our website.
Dr. Sarah Chen is the Managing Partner and Co-Founder of Milestone AI Ventures. She previously served as Research Director at Google DeepMind and holds a Ph.D. in Machine Learning from MIT. The views expressed here are her own and do not constitute investment advice.