The noise and controversy around AI risks and safety continues in full-flow, but there is also a nuanced shift from a shiny new toy to a focus on outcomes, cost and being an essential business tool. It’s come through the Emperors New Clothes of the hype cycle and is rapidly gaining wider adoption, although the debate around governance, sovereignty, guardrails and ethics will continue for some time.
I have two bug bears: there is still too much focus on we’ve got an AI product you need and badging something AI when it’s simply smart workflows. We need to move from return on illusion to return on investment. AI is a capability not a business model. Calling your company an AI startup is like a business in 1999 calling itself an Internet company. AI is the engine under the hood, but the business model still depends on sales, retention, and margins.
The emergence of Agentic AI has catalysed structural shifts in knowledge based industries. Consulting has adopted AI as a delivery tool, acknowledging the cannibalisation it brings to the core business model, but nowhere has this transformation been more profound, or more meticulously executed, than in legal services, traditionally slow adopters of tech and complacent in pricing for their services. Enter Harvey.
Founded in the 2022, Harvey stands as a landmark case study in vertical application layer AI and using the Human Moat as a defensibility strategy. By bridging the gap between general purpose LLMs and human created/delivered advisory services, Harvey evolved from a coding experiment into an enterprise system trusted by global law firms.
Harvey is rooted in the convergence of legal domain expertise and frontier AI research. The company was co-founded by Winston Weinberg and Gabriel Pereyra. Weinberg was a litigator at O’Melveny & Myers. His daily routine was defined by the traditional friction points: hours spent reviewing documents, sifting through precedents, cross-referencing complex contract clauses, and drafting counter-point detailed responses
Pereyra was an AI research scientist who had worked at Google DeepMind under Demis Hassabis. The pivotal moment occurred when Pereyra introduced Weinberg to OpenAI’s GPT-3. While the tech community and general public viewed GPT-3 primarily as a text generator or creative tool, Weinberg saw its potential as a domain specific reasoning tool.
The legal sector is built on structured text, reasoning, and precedents – well suited for language modeling. However, generic models were prone to hallucinations and lacked domain context, whilst posing data security and confidentiality risks. Weinberg and Pereyra set out to build a system explicitly tailored to legal standards, naming the platform Harvey after the lawyer Harvey Specter from the television series Suits. Building an AI platform capable of operating in the legal environment required an approach far more sophisticated than wrapping an API around a commercial LLM. The engineering team prioritised four fundamental architectural pillars:
- Domain specific: Harvey secured backing from OpenAI. This relationship gave access to leading foundational models to build a custom trained model fine tuned on the entirety of U.S. case law. Rather than relying solely on generic prediction, the platform was trained to mirror how lawyers analyse and synthesise legal issues.
- Custom legal embeddings: To ensure high semantic precision when querying legal text, Harvey collaborated with Voyage AI to develop a proprietary embedding model trained on over 20 billion tokens of legal text. This specialised embedding layer reduced irrelevant search results by 25% compared to off the shelf alternatives, capturing nuances in legal terminology that general models missed.
- Multi model orchestration: As the AI matured, Harvey transitioned into a multi-model orchestrator. The architecture dynamically routes sub-tasks to the best suited model. This extended reasoning to handle complex legal analysis and step-by-step logic, and custom RAG pipelines matching extracted citations against authoritative legal databases.
- Enterprise security & privacy: Law firms operate under strict client confidentiality and responsibility. Harvey was architected to ensure that client data was never used to train foundation models. The platform was built with zero data retention options, SOC2 compliance, robust encryption, and integration options with Azure, allowing enterprise clients to run Harvey in dedicated cloud environments.
Before launching commercially, Harvey conducted rigorous testing to answer a single question: Could AI-generated legal work meet the threshold of a practicing attorney? Weinberg and Pereyra designed a blind control test. They fed 100 real, complex landlord-tenant legal queries into GPT-3, which was refined using prompting and retrieval techniques.
They submitted the outputs to three practicing specialist attorneys without revealing that the text was generated by AI. The attorneys were asked a binary question: Would you feel comfortable sending this answer directly to a client without edits? 86% were approved without amendments. To eliminate the risk of hallucinations, they engineered a multi-layered verification pipeline. Generated text was broken down into discrete factual and legal assertions, cross-referenced against authoritative sources and flagged if inconsistencies exist.
Tests demonstrated that this approach to the detection of hallucinations pushed accuracy to 99.8%. In head to head evaluation trials, attorneys preferred Harvey’s outputs over untuned GPT-4 responses 97% of the time. This validation proved that structured LLMs could produce outputs that met expected legal standards.
Rather than releasing the product, Harvey targeted a Minimum Viable Trust pilot strategy, deploying directly into elite global law firms. In November 2022, Allen & Overy began a confidential trial. The pilot involved over 3,500 lawyers across 43 jurisdictions, testing the platform across contract analysis, due diligence, litigation and regulatory compliance. The results of the pilot were transformative:
- Lawyers saved an estimated 2 to 3 hours per week on routine drafting and research tasks.
- Contract review times were reduced by 30%.
- Savings on complex, multi-document comparative analysis reached up to 7 hours per matter.
In February 2023, Allen & Overy announced deployment of Harvey across the firm. PwC announced a global strategic alliance with Harvey to provide the platform to its 4,000 legal professionals, cementing Harvey’s credibility. Following its successful pilot phase, Harvey rolled out an integrated enterprise suite to integrate with commercial software, including Microsoft Office, iManage, SharePoint, and Google Drive.
By combining deep legal experience with frontier AI engineering, Harvey transformed how legal work is conducted. Rather than replacing lawyers, it established a new model for AI-augmented professional services, demonstrating that when domain specific trust, robust verification, and seamless workflow integration are prioritised.
In a wider perspective, Harvey is a great case study of the Human Moat building defensibility in the age of Synthetic Intelligence.
For many years, tech startup defensibility followed a repeatable pattern: write complex code with proprietary algorithms, build switching costs around data, and then scale unit economics. AI has fundamentally disrupted this model. When a single developer using coding agents can replicate a complex application in a weekend, code is no longer a moat, it is a commodity. When foundational models process, structure, and synthesise raw inputs, feature complexity falls away as a competitive tool.
A Human Moat consists of non-replicable human dynamics including domain expertise, judgement, expert-in-the-loop, trust and high empathy. Where software can be copied, these human characteristics can’t, and will take years to compound synthetically. Creating a Human Moat should be a core and deliberate strategic process if you’re developing an AI product.
Here are the lessons from Harvey for Founders building an AI product to build human defensibility across three phases:
Step 1: Architect Expert-in-the-Loop. Do not build pure self-service software in complex domains.
- The Strategy: Design product workflows where human experts supervise, correct, and refine AI outputs in real time.
- The Moat: Every time an expert corrects a model recommendation, you capture proprietary structured data that competitors relying on web scrapes cannot access, creating a moat
Step 2: Leverage high touch human contact over algorithm based distribution. In an era of constant spam and cold email sequences, adopt high empathy, high trust relationship channels.
- The Strategy: Build small, exclusive communities, and co-design solutions alongside them.
- The Moat: Enterprise buyers purchase from people they trust to solve their problems, not from faceless mass marketing.
Step 3: Own the AI governance layer. Position your firm as a fully managed service or accountable platform rather than an unmonitored AI tool.
- The Strategy: Guarantee outputs by pairing model speed with human verification. Provide indemnification of human sign offs on critical deliverables.
- The Moat: Competitors offering pure solutions cannot match your peace of mind or regulatory compliance guarantees.
Summary: As AI makes technical execution more abundant and cheaper, human-centric assets become more valuable. Ironically in the most rapid and widespread technological revolution, the most durable competitive advantage for AI startups is the Human Moat. However a Human Moat is not a permanent shield, it requires continuous maintenance, but I suspect the Human Moat will only be temporary as AI advances.
Julien Bek‘s argument in Services: The New Software is that the next trillion dollar valued AI product company won’t sell software to do part of the job quicker and cheaper, it will sell completed work – outcomes. Harvey is on this journey, autopilot territory is coming. Once judgment is affirmed, and as AI assimilates data and learns what good judgment looks like, today’s judgment becomes tomorrow’s intelligence.
A lawyer does not just need a better answer, they need the right sources, and context, and the judgment to know when an answer should not be trusted. Gen AI gave us access to intelligence, but Vertical AI turns that intelligence into professional capability. Law is one example, but the same approach can extend across any professional advisory service – consulting. finance, medicine, engineering, insurance – in fact every knowledge based profession.
This creates a marker for AI product founders: the model becomes the foundation, not the moat. The real advantage will come from everything built around it: proprietary knowledge, domain understanding, structured data, workflow integration, organisational context, trust and ultimately, delivering the work itself and outcomes than humans once used to deliver.