Harvey AI's Forward Deployed Engineers: How the $3B Legal AI Leader Deploys Inside BigLaw

TL;DR#

Harvey AI — the $3B legal-AI company backed by OpenAI, Sequoia, Kleiner Perkins, and GV — runs one of the most specialized forward deployed engineering functions in the AI industry, with FDEs embedded inside Allen & Overy (now A&O Shearman), PwC, Cleary Gottlieb, Macfarlanes, and dozens of other AmLaw 100 and Magic Circle firms. Harvey's forward deployed engineers do not write generic chatbot integrations. They sit inside law firm matter rooms, map partner-by-partner workflow tribalism, and rebuild firm-specific knowledge bases on top of foundation models so that the LLM speaks in the firm's house style, cites the firm's precedent vault, and respects the firm's confidentiality walls. According to Harvey's own published deployment cadence, a full firm rollout takes 6–9 months and touches three distinct populations — partners, associates, and knowledge management lawyers — each with incompatible adoption triggers. Harvey's model has become the de-facto template for any AI vendor selling into BigLaw, and it has implications well beyond law: any vertical AI company selling to regulated, partner-led, knowledge-dense buyers is now copying the Harvey playbook. The hardest part of the deployment is not the model. It is the customer-research loop — and partners, the people who sign the contract, are systematically the wrong people to interview about the workflow being automated.

What does a Harvey AI forward deployed engineer do?#

A Harvey AI forward deployed engineer embeds full-time inside a single BigLaw client for 6–9 months to translate the firm's most valuable, most idiosyncratic legal workflows — M&A diligence, securities filings, regulatory memos, litigation document review — into firm-specific LLM applications running on top of Harvey's foundation-model layer. The role blends three jobs that are usually separate at a normal enterprise software vendor: customer-discovery researcher, applied ML engineer, and onboarding lead. Harvey FDEs spend roughly a third of their time on conversation — with partners, associates, knowledge-management lawyers, and IT — a third on data engineering against the firm's document management system (typically iManage or NetDocuments), and a third on retrieval, evaluation, and fine-tuning work to make Harvey speak in the firm's voice.

The function exists because BigLaw firms cannot be sold the same SaaS product. A magic-circle firm's M&A practice and a New York white-shoe firm's M&A practice both look like "M&A" to outsiders, but the precedent libraries, mark-up conventions, and partner sign-off chains are unrecognizably different. A forward deployed engineer is the only economic way to bridge that gap.

This is the same general pattern documented in our 2026 survey of 1,500 forward deployed engineers — but Harvey's version is unusually specialized because the customer is unusually specialized.

Why BigLaw deployment demands forward-deployed staffing#

BigLaw deployment demands forward-deployed staffing because three structural features of large law firms — strict confidentiality regimes, on-prem and private-cloud data-residency requirements, and partner-by-partner adoption autonomy — make remote, self-serve SaaS rollouts mathematically impossible. You cannot ship a Harvey instance to Allen & Overy's London office the way Notion ships to a fintech startup. The technical, organizational, and cultural friction has to be absorbed by a human who lives inside the firm.

Confidentiality and the matter-room wall. Every document inside a BigLaw firm is bound by attorney-client privilege, ethical walls between matters, and conflict-of-interest screening. A Harvey FDE has to be cleared for specific matters, has to understand which partner's clients cannot see which other partner's clients, and has to architect retrieval such that an associate working on Deal X cannot accidentally pull precedent from Deal Y if the same partner is conflicted out. According to the American Bar Association's 2024 Legal Technology Survey, 79% of firms with 100+ attorneys cite client confidentiality as the #1 barrier to AI adoption — far ahead of cost or accuracy.

On-prem and private-cloud data residency. Most AmLaw 100 firms operate hybrid environments where the document management system, the time-and-billing system, and the conflicts database live on-prem or in a private VPC. Harvey FDEs are responsible for the integration plumbing into iManage, NetDocuments, Aderant, and Intapp — none of which expose clean public APIs the way Notion or Slack do. This work cannot be productized. It has to be re-done firm by firm.

Partner-by-partner adoption. A BigLaw firm is not a hierarchical organization. It is a federation of partner-owned practices, each with veto power over the tools their associates use. A Harvey deployment that "lands" at the firm level is meaningless if 40 of the 60 M&A partners refuse to use it. The FDE's real job — the one nobody puts in the deployment statement of work — is winning partners one at a time, partner by partner, demo by demo, mark-up by mark-up.

This is the same buying-committee chaos described in our forward deployed engineer vs ML engineer vs solutions architect comparison, but the BigLaw flavor is more extreme: even the CIO cannot force a senior equity partner to adopt anything.

Inside a Harvey deployment: firm-discovery → workflow-mapping → knowledge-base-build#

A Harvey AI deployment runs as a three-phase loop — firm-discovery, workflow-mapping, knowledge-base-build — that the embedded FDE runs in parallel rather than sequence, looping back to discovery whenever a new practice group is brought online. Each phase produces deliverables that feed the others.

Phase 1: Firm-discovery (weeks 1–6). The FDE maps the firm's practice groups, partner-level workflow variants, document management system architecture, conflicts and ethical walls, knowledge management function, and incumbent legal tech stack. Harvey FDEs report doing 40–80 hours of qualitative interviews in this phase — partners, associates, paralegals, KM lawyers, IT, security. This is the phase where most AI-vendor deployments at law firms fail: the FDE assumes the workflow looks like the deal closing checklist the firm published on its website, and misses that the actual workflow is encoded in a senior partner's head and a paralegal's spreadsheet.

Phase 2: Workflow-mapping (weeks 4–14, overlapping). The FDE picks two or three concrete workflows — typically one transactional (M&A diligence, securities filing) and one advisory (regulatory memo, opinion drafting) — and maps every step from intake to delivery. This phase produces a workflow specification that names exactly what the LLM is being asked to do, what inputs it can rely on, what it must escalate to a human, and what "good" looks like. Without this spec, the rest of the deployment is hallucination roulette. The methodology we use for the same kind of discovery loop on conversational AI products is documented in how forward deployed engineers run customer discovery.

Phase 3: Knowledge-base-build (weeks 8–28, overlapping). The FDE ingests, cleans, and structures the firm's precedent vault — historical engagement letters, mark-ups, opinion memos, transaction documents — into retrieval-augmented generation indexes that respect matter walls, redact privileged content, and preserve provenance. This is the part of the work that genuinely cannot be done by anyone outside the firm. Harvey's product team builds the retrieval framework; the FDE makes it speak the firm's language.

Continuous loop: evaluation and partner rollout. Once the first workflow is live, the FDE switches into rolling adoption mode — running gold-standard evaluation sets against the partner's actual output, sitting with individual partners on individual deals, and tuning the retrieval and prompting for that partner. This is where the Harvey deployment differs most from a traditional enterprise SaaS rollout: there is no "launch day." There is a multi-quarter sequence of partner-by-partner conversions.

Notable deployments: Allen & Overy, PwC, Cleary Gottlieb, A&O Shearman patterns#

Harvey AI's most publicly-documented deployments — Allen & Overy (now A&O Shearman post-merger), PwC, Cleary Gottlieb, and Macfarlanes — show a consistent pattern: start with a single global practice group, win adoption through a partner champion, and expand horizontally through the firm only after the first cohort has six months of live use.

Allen & Overy / A&O Shearman. A&O was Harvey's first major enterprise customer, announced in February 2023, when 3,500 lawyers across 43 offices got access to Harvey-powered tools. The deployment was led by David Wakeling, A&O's then-head of the Markets Innovation Group, who functioned as the firm-side counterpart to Harvey's FDE team. Wakeling has publicly said that early use was concentrated in contract drafting, due diligence, and regulatory work — exactly the workflows Harvey FDEs prioritize in Phase 2. The 2024 A&O–Shearman merger gave Harvey an even larger combined practice to deploy against.

PwC. PwC's March 2023 partnership with Harvey extended Harvey beyond pure law firms into Big Four legal services arms. PwC's 4,000+ legal professionals globally got access, and Harvey's FDE team had to build new workflow templates for the tax, deals, and regulatory advisory work that PwC's legal arm specializes in — a different workflow surface area than a pure M&A shop. The PwC deployment is a useful proof point that the Harvey FDE model is workflow-flexible, not firm-flexible.

Cleary Gottlieb. Cleary's deployment, public since late 2023, has been a quieter case study in partner-led rollouts. Cleary's complex cross-border M&A and sovereign restructuring practices are exactly the kind of high-stakes, partner-driven work where the FDE-led knowledge-base-build phase produces outsized returns: the precedent vault is enormous and the marginal value of fast retrieval is high.

Macfarlanes, Reed Smith, and others. Magic Circle and AmLaw 100 firm announcements have continued through 2024 and 2025. Each new firm requires its own FDE-led discovery and knowledge-base-build cycle — which is why Harvey has aggressively scaled its forward deployed engineering and applied AI teams. The same vertical-deployment logic plays out elsewhere too — at insurance carriers and at enterprise data platforms — and we cover the analogous patterns in our Scale AI forward deployed engineering breakdown, the Mistral AI enterprise LLM deployment playbook, and the broader Latham & Watkins generative AI rollout. The DLA Piper deployment — covered in our DLA Piper AI legal intake case study — is another useful comparison.

The customer-research challenge no one talks about: partners are not the right interview subjects#

The hardest, most under-discussed problem in BigLaw AI deployment is that the people who pay for Harvey — equity partners — are systematically the wrong people to interview about the workflow Harvey automates. Partners do not do the workflow. Associates and paralegals do. Partners review the output, sign the bill, and own the client relationship — but the moment-to-moment work of pulling a precedent, drafting a clause, marking up a redline, building a closing checklist happens two or three levels below them.

This creates a research trap that every Harvey FDE has to navigate. The partner buys the tool. The partner gets demoed. The partner gives feedback. But the feedback is filtered through years of distance from the actual keystrokes. A partner will say "Harvey is great, it saved me hours on the diligence memo." What the partner actually means is "the associate handed me a cleaner first draft." The workflow improvement happened to the associate, who never spoke to anyone on Harvey's team.

We see the exact same dynamic in other regulated verticals — insurance underwriting, healthcare clinical workflows, financial compliance — where the buyer (a VP, a chief medical officer, an SVP of underwriting) is not the user (an underwriter, a clinical assistant, a compliance analyst). The buyer-user gap is the single largest source of failed AI deployments. We've written about the analogous pattern at named insurance carriers in our AIG conversational underwriting deep-dive, our Allianz customer-research strategy breakdown, and the Zurich Insurance commercial-lines discovery playbook.

The fix is structural. Harvey's FDE team has to run two parallel research tracks: a partner-facing track (commercial framing, value justification, billing-rate impact) and an associate-and-KM-facing track (actual keystroke-by-keystroke workflow capture). The two tracks produce contradictory data — partners overstate Harvey's coverage, associates point to gaps the partner cannot see — and the FDE's job is to reconcile them into a single deployment plan. According to Thomson Reuters' 2024 Future of Professionals Report, 67% of legal professionals expect AI to be transformational to their work in the next five years, but the report also notes that associate-level expectations diverge sharply from partner-level expectations on which specific workflows will be automated first.

This is why interview infrastructure matters at the buyer level, not just the design-partner level. Generic survey tools that funnel a partner's response and an associate's response into the same dropdown miss the entire signal. Conversational AI research — interviews that adapt, probe, and capture the why — is the only way to surface the partner-associate gap at scale. Forward deployed engineers at Harvey, Scale AI, OpenAI, Mistral, and Anthropic increasingly run discovery this way. Perspective AI was built for product teams and CX teams running exactly this kind of multi-stakeholder customer research, with our interviewer agent handling the open-ended workflow-capture conversations and our concierge agent replacing the forms that flatten partner-associate distinctions into the same checkbox. If you're standing up an FDE function and need an interview engine that scales the way the rest of the deployment stack does, you can start a study without a meeting.

The deeper architectural pattern — that AI-first products cannot start with web forms because forms collapse the partner-associate-paralegal gap into a single schema — is the same one driving conversational adoption in insurance intake, legal intake, and healthcare intake more broadly. The AI legal intake playbook for personal injury firms and the AI patient intake playbook for mental health practices cover the consumer-facing version. Harvey's BigLaw deployment is the enterprise-facing version of the same insight.

Frequently Asked Questions#

What is Harvey AI?#

Harvey AI is a legal-AI company building generative AI tools — drafting, research, document review, contract analysis — purpose-built for large law firms and corporate legal teams. Founded in 2022 by former O'Melveny & Myers litigator Winston Weinberg and DeepMind researcher Gabriel Pereyra, Harvey raised at a $3B valuation in 2024 from Sequoia, Kleiner Perkins, GV, and the OpenAI Startup Fund. Its first major customer was Allen & Overy in February 2023.

How many forward deployed engineers does Harvey AI employ?#

Harvey AI has not published an exact FDE headcount, but based on its public deployment list (Allen & Overy / A&O Shearman, PwC, Cleary Gottlieb, Macfarlanes, Reed Smith, and dozens of additional AmLaw 100 and Magic Circle firms) and its disclosed deployment cadence of 6–9 months per firm, the function is estimated at 40–80 engineers as of early 2026 — large enough that Harvey has built dedicated applied-AI and forward-deployed teams as separate reporting lines.

How does Harvey AI handle confidentiality and ethical walls?#

Harvey AI handles confidentiality and ethical walls by combining tenant-level data isolation, matter-level access controls inside the deployment, and audit logging for every retrieval and generation. The FDE-led deployment phase is where the matter walls get configured — the technical primitives exist in the product, but they have to be mapped onto the firm's specific conflicts database and partner-by-partner permission model during knowledge-base-build. According to the ABA, 79% of large firms cite confidentiality as the top AI-adoption barrier, which is why Harvey's deployment is FDE-led rather than self-serve.

Who are Harvey AI's biggest customers?#

Harvey AI's biggest publicly disclosed customers include Allen & Overy / A&O Shearman, PwC, Cleary Gottlieb, Macfarlanes, Reed Smith, Ashurst, Mishcon de Reya, and a long list of AmLaw 100 and Magic Circle firms. Harvey also serves corporate legal departments at Fortune 500 companies, though most of those engagements are less publicly disclosed than the law-firm deployments.

Why does Harvey AI need forward deployed engineers if it sells software?#

Harvey AI needs forward deployed engineers because BigLaw firms cannot consume horizontal SaaS. Every firm has its own precedent vault, its own document management system, its own conflicts and ethical walls, and a partnership structure that gives individual partners veto over tool adoption. An FDE absorbs the integration, customization, and partner-by-partner adoption work that no productized rollout could handle remotely, and bridges the gap between Harvey's foundation-model product and the firm's idiosyncratic workflows.

Is the Harvey AI deployment model relevant outside legal?#

Yes — the Harvey AI deployment model is being copied by AI vendors across every regulated, partner-led, knowledge-dense vertical, including insurance underwriting, healthcare clinical workflows, financial compliance, and Big Four advisory. Any AI company selling into a market where the buyer is not the user, the workflow is encoded in tribal knowledge, and the data lives in on-prem systems with strict residency requirements ends up building an FDE function that looks structurally similar to Harvey's. The Palantir-Anthropic-OpenAI FDE comparison covers the broader pattern.

Conclusion#

Harvey AI's forward deployed engineering function is the defining case study for vertical AI deployment in 2026. The work is half technical integration and half customer research, run inside the buyer's office over a 6–9 month cycle, with the FDE owning everything from data plumbing into iManage and NetDocuments to partner-by-partner adoption demos. The model works because BigLaw deployment cannot be productized — confidentiality regimes, on-prem data residency, and partner adoption autonomy each individually defeat self-serve SaaS, and combined they require a human inside the firm to absorb the friction. Harvey's playbook has become the template for forward deployed engineering at every vertical AI company, from Mistral's European enterprise LLM deals to Scale AI's enterprise data work to the broader applied-AI teams now standing up at Anthropic and OpenAI.

The under-discussed lesson is that the hardest part of a Harvey deployment is the customer-research loop, not the model. Partners pay for the tool but don't do the workflow. Associates do the workflow but don't sign the contract. Knowledge management lawyers see the gaps both groups miss. Generic survey tooling flattens those three populations into the same dropdowns and the same scores, which is exactly why every FDE function ends up building or buying conversational interview infrastructure. If you're running a forward deployed engineering function and want to scale the discovery side of the work the way Harvey scales the engineering side, Perspective AI gives you an AI interviewer agent that runs multi-stakeholder workflow-capture interviews at the volume an FDE function actually needs. You can browse use cases, see pricing, or start a study — no demo required.