www.iqsource.ai Google Entered Legal and Harvey Became a Connector
Ricardo Argüello, September 5, 2026
CEO & Founder
General summary
On August 25, 2026, Google Cloud launched Gemini Enterprise for Legal in preview, with Cleary, Freshfields, Weil and Williams & Connolly as launch firms. Google now joins OpenAI and Anthropic with a dedicated legal offering. Harvey and Legora appear inside it as Model Context Protocol integrations, the second time in four months for Harvey, which had already been a connector inside Claude for Legal since May. In the week before the announcement Harvey answered twice: its own model, and a second generation built around memory.
- Google Cloud launched Gemini Enterprise for Legal on August 25, 2026, in preview
- Launch firms are Cleary, Freshfields, Weil and Williams & Connolly
- It ships legal-specific skills: brief drafting, citation verification, contract lifecycle management and regulatory scanning
- Harvey, Legora, Thomson Reuters, iManage and RelativityOne appear as Model Context Protocol integrations
- Harvey was already listed as a connector inside Claude for Legal, which Anthropic launched on May 12, 2026
- On August 18, Harvey launched Harvey II with memory, preferences and shared context at its core
- On August 20, it launched Tenet, its first in-house model, post-trained on the open-weight Kimi K3 base
For years you sold the best drill on the market. One day the toolbox manufacturer ships a box with a drill mount built in, and you show up on the compatibility list. You were not pushed out. You were moved from product to accessory, and the customer relationship now starts at the box instead of at your drill. It happened to Harvey twice in four months: first inside Claude for Legal in May, and now inside Gemini Enterprise.
AI-generated summary
Harvey is on the compatibility list.
That is the sentence to sit with. On August 25 Google Cloud launched Gemini Enterprise for Legal in preview, with Cleary, Freshfields, Weil and Williams & Connolly as launch firms, and Harvey shows up inside it as a Model Context Protocol integration next to Legora, iManage, NetDocuments, RelativityOne and Thomson Reuters.
Product to accessory, twice in four months
Harvey has raised at valuations that climbed from $3 billion to $11 billion, and reports in early August put it in market for $500 million at $15.5 billion. It is the best-known legal application of this cycle.
Google was not the first to put it on a list. On May 12 Anthropic shipped Claude for Legal, with more than twenty connectors and twelve practice-area plugins, and Harvey was already inside. Four months later the best-known application in legal AI is listed as a component inside two of its own model suppliers.
Being inside the platform your customer already runs is distribution, and free distribution at that. So this is not a defeat.
What it changes is where the conversation starts. A firm used to evaluate Harvey. Now a firm evaluates Gemini Enterprise and decides which connectors to switch on.
Selling the drill and appearing on the toolbox’s compatibility list are different businesses, with different pricing power.
Harvey answered twice in one week
On August 18, a week before the Google announcement, Harvey launched its second generation with memory as the core: preferences, shared context, learning from what the firm’s lawyers want and what their clients want.
Two days later, on August 20, it shipped Tenet, its first in-house model. Not trained from scratch: they took Kimi K3, an open-weight base, and post-trained it on legal data.
The timing reads as deliberate, and the order matters.
Tenet is the loud answer, and it defends the least. Harvey just demonstrated that a specialised model is two months of compute on top of an open base, which means Thomson Reuters can do it, and so can a large firm. Proving something is reproducible in two months is a strange way to make it your moat.
Memory is the quiet answer, and it is the one thing a connector cannot copy. Google can integrate with your document management system over an API in an afternoon. It cannot integrate with three years of knowing that this partner wants indemnity clauses formatted a particular way, that this client rejects certain language, that this jurisdiction always produces the same fight.
That accumulation does not live in a document. It is built through use, and it is lost when the tool changes.
We made the same argument in the harness is the moat: once the model is interchangeable, the defensible part is everything wrapped around it.
What this tells any software company
If you sell B2B software with AI inside, treat this as your dress rehearsal.
The platform is coming to your category. Google, Microsoft or whoever will ship the version configured for your industry, with connectors into the systems your customers already run, inside a contract they already signed.
When it happens, the question is what you hold that the platform cannot connect to.
If your advantage was calling a model well, you become a connector. If your advantage is accumulated context about how this specific customer works, you keep the relationship even when the front door changes owner.
We looked at the law firm side of that same question in Kirkland is not building a model.
Three things to check in your own product
What does your system know about each customer that is not sitting in a database they could export tomorrow?
How long would it take a customer to replace you? If the honest answer is two weeks, you have an interface, not a memory.
Are you measuring usage that compounds, or only usage that bills? Those are different metrics, and the second one does not predict retention on the day the platform shows up.
Harvey shipped memory seven days before Google entered its category, and its own model five days before. Whether that was reaction or foresight I do not know, but the order of the moves is right and worth copying.
Let’s find the part of your product an API cannot reachFrequently Asked Questions
It is a version of the Gemini Enterprise platform configured for law firms and corporate legal departments, launched in preview on August 25, 2026. It ships legal-specific skills, pre-built agents, and connectors into the document management, e-discovery and legal research systems firms already run.
Both at once. Harvey is listed as an integration inside Gemini Enterprise for Legal through the Model Context Protocol, alongside Legora, Thomson Reuters, iManage and RelativityOne. Google did not displace it from the market, but it did reposition it as a component inside a platform that now owns the customer's entry point. And Google was not the first to do it: Harvey was already a connector inside Claude for Legal, the legal offering Anthropic launched on May 12, 2026.
Harvey II is the second generation of Harvey's legal AI platform, launched August 18, 2026, built around memory. It adds preference setting, shared context, and learning from what both a firm's lawyers and their clients want, which makes it harder to swap out for a generic connector.
Tenet is Harvey's first proprietary model, announced August 20, 2026. It was not trained from scratch: Harvey started from Kimi K3, an open-weight base from Moonshot AI, and post-trained it on legal data with attorney labelling. Harvey positions it as the path for each firm to build and own a specialised model instead of licensing a general-purpose one.
Whatever the platform cannot reach through an API: accumulated memory of how that specific customer works, learned exceptions, and judgment encoded over years of use. An application whose only advantage was calling a model well becomes a connector. One that compounds customer-specific context keeps the relationship even when the platform changes hands.
Related Articles
Harvey vs. MikeOSS: the code is free, the risk isn't
MikeOSS matches most of what Harvey does, for free. But someone still has to carry the operational and contractual risk that Harvey takes off a client's plate.
The decade's most expensive rebrand: Big Five to FDE
Google joined the FDE race. Maurizio calls it the rebrand of the 2020s. It's the third label for the same job since 1990. This time the difference is Satya.
Thomson Reuters Trained Its Own Model for $450,000
Thomson Reuters built a legal model that beats GPT 5.4 and Claude Sonnet 5 on its own content. The final training run cost $450,000. The corpus is the moat.
Harvey Tenet runs on Kimi K3. The benchmark came first
Harvey post-trained an open Chinese base and placed first on contracts, second overall, on a benchmark it owns. The measurement is what made it possible.
Blackstone bet on Norm AI. The market just agreed.
Blackstone put $50M into Norm AI in November and started using it inside its own legal function. Eight months later, a Series C valued the company at $1.2B.
Five Tech Giants Back ARD. Anthropic Isn't Invited
Google, Microsoft, Salesforce, ServiceNow, and Snowflake back ARD, a new protocol for AI agent discovery that pointedly leaves out Anthropic and OpenAI.
