Buy, Wait, or Build: Where the Legal AI Market Is Settling

Law firms weighing legal AI have three real options: buy a specialized vendor platform, wait for the frontier model companies to build legal tools natively, or build workflows in-house on an AI-agnostic architecture. For most small and mid-size firms, the third option is now the strongest one, and the reason is that the specialized vendors and the frontier labs are converging on the same underlying models. In that environment, the only asset a firm can hold onto is the workflow it designs and governs.

That is the short answer. Here is the longer one.

Three Pressures on the Same Market

The legal AI market in 2026 is crowded and well funded. Harvey raised $200 million at an $11 billion valuation in March. Legora closed a $600 million Series D the following month and now sits at a $5.6 billion valuation.¹ Hundreds of smaller vendors sit beneath them, and a common criticism of the smaller tier is that many are thin software layers over a public model API, priced like enterprise software.

Pricing supports the complaint. Legal AI tools range from free to more than $1,200 per seat per month, and enterprise platforms such as Harvey and CoCounsel start around $500 per seat with annual commitments.² A firm buying ten seats at that level is spending $60,000 a year before implementation, training, or integration.

The second pressure comes from the model developers themselves. In February, Anthropic released legal plugins for its Claude Cowork agent, and legal software stocks sold off in response.³ In May, it expanded that into Claude for Legal: twelve practice-area plugins covering commercial, employment, litigation, privacy, and AI governance work, plus connectors into Thomson Reuters, Harvey, Box, Everlaw, and DocuSign.⁴ Bloomberg Law's read was that Anthropic is positioning itself as a direct provider of legal technology rather than a supplier of models to other vendors.⁵

Google followed on August 25 with Gemini Enterprise for Legal, a packaged, agentic version of its Gemini Enterprise platform with legal-specific skills (brief drafting, citation verification, contract lifecycle management, regulatory scanning), MCP connectors into iManage, NetDocuments, Everlaw, RelativityOne, Harvey, and Legora, and Cleary, Freshfields, Weil, and Williams & Connolly as launch customers.¹¹ Microsoft had already shipped a Legal Agent inside Word in April.¹² Within seven months, three of the largest platform companies in the world have each released a legal product that sits directly inside the environment where lawyers work.

The specialized vendors are responding by embedding deeper into the systems lawyers already use. Legora expanded its iManage partnership in February, added a Box integration, and acquired the agentic workflow startup Walter AI in March.⁶ The strategy is to become part of the document management layer so that a firm cannot easily remove the tool once it is installed.

The third pressure is corporate. The Thomson Reuters Institute's 2026 State of the US Legal Market report warns that corporate legal departments, already running AI internally at lower cost, are questioning why their outside firms are not passing along similar efficiencies, and that some could build in-house capability rather than pay for it.⁷

So the managing partner's question is fair: when the shakeout is done, what is left standing, and where should the firm have placed its bet?

The Build Versus Buy Math Has Changed

Bloomberg Law's June survey of law firm technology budgets found that roughly two-thirds of large firms spent a small fraction of their budgets on new technology in 2025, while one in three firms with fewer than 500 lawyers spent more than 8 percent on technology subscriptions and recurring costs.⁸ Smaller firms are paying proportionally more, and paying it to vendors.

Historically that made sense. Building meant hiring engineers, owning model risk, running bias testing, and maintaining the system indefinitely. That path was reserved for firms like Kirkland & Ellis, which is putting $500 million into a proprietary AI platform.⁸

The substrate underneath that decision has shifted. Agentic frameworks, no-code workflow builders, and AI-assisted prototyping have lowered the cost of building a working legal workflow from a multi-year engineering project to something a knowledge management team can prototype in weeks. Production deployment still requires governance, testing, and supervision. But the entry cost is no longer the barrier it was two years ago.

The choice is no longer a $500-per-seat subscription or a seven-figure software build. There is a third path.

The Case for AI-Agnostic, In-House Workflows

Our position at InGlobo AI is that firms should stop evaluating AI by product and start evaluating it by workflow. A firm that ties its operating model to a single vendor is carrying unhedged platform risk. If that vendor's underlying model falls behind, if licensing costs rise, or if the vendor is acquired and the product roadmap changes, the firm is locked in.

Building workflows in-house on an AI-agnostic architecture addresses that in three ways.

1. The firm owns the logic

When a firm designs its own workflow, the firm owns the process: the intake steps, the review criteria, the escalation rules, the output format. The model running underneath is interchangeable. Harvey, Legora, Solve Intelligence, and Eve all run on Claude's underlying models, according to the firms and Anthropic's own May announcement.⁹ The specialized vendor is renting the same engine the firm could use directly. An agnostic architecture lets the firm swap that engine as pricing, performance, and security posture change.

2. The firm can measure its own unit economics

The 2026 State of the US Legal Market report found that 90 percent of legal dollars still flow through hourly billing, and describes the tension of deploying tools that accomplish in minutes what once took hours, then billing for it by the hour.⁷ A firm cannot fix that problem with a vendor's dashboard. It needs to know what a specific workflow costs to run, how much lawyer time it replaces, and what the output is worth to the client. That measurement is only possible when the firm controls the workflow end to end. Once a firm has those numbers, flat-fee and value-based pricing become a defensible business decision rather than a leap of faith.

3. The firm controls the data path

When a workflow runs inside the firm's own environment, the firm decides where client data goes, how long it is retained, and which model provider sees it under what terms. The firm is not exposed to a third-party vendor revising its terms of service, suffering a breach, or changing its training data practices. This is a Rule 1.6 question before it is a technology question: a lawyer's duty of confidentiality applies regardless of which tool is in the stack, and a firm that can document its own data controls is in a far stronger position when a client's RFP asks how AI is handled.¹⁰

What the Landscape Looks Like When It Settles

The likely end state is AI as infrastructure rather than a single dominant point-solution vendor: embedded in the document management system, the email client, and the practice management platform, and invoked through lightweight plugins or firm-built agents rather than purchased as a separate product. Anthropic's connector strategy, Google's Gemini Enterprise for Legal, and Legora's iManage integration are all moves toward that same end state from different directions. Google's own launch materials pitch the platform as a way to customize and scale "without vendor lock-in," which tells you the platform companies know exactly what buyers are worried about.¹¹

For a firm, the practical implication is that the durable asset is the workflow design and the governance around it. Contract triage, timeline extraction, deposition preparation, conflict checks: each of these can be built once, governed once, and run on whichever model is best that quarter.

At InGlobo AI, that is the work we do with firms: designing the workflows, building the governance framework that makes them defensible, and keeping the firm free to change models as the market moves.