Google's industry agents inherit ethical walls and cite sources. Industry AI is about rules, not vocabulary

A legal AI that knows what an indemnity clause is, but not which lawyers are walled off from a matter, is not a legal product. It is a liability with good vocabulary.

That is the useful way to read Google’s industry editions of Gemini, announced at Gemini at Work on October 8. Google says the specialized Gemini is “now in preview for Financial Services and Legal, and coming soon to Government, Healthcare, and Retail”. The interesting parts of both previews are not what the model knows about the industry. They are the rules it promises to respect.

In legal, the walls live in the document system

For law firms and legal departments, Google’s headline is that Gemini “inherits matter-level permissions and ethical walls directly from document management platforms like NetDocuments and iManage”. Harvey and Onit are named as partners, and Cooley, Google says, is building an agent that redacts confidential information before filings go public.

Inheriting the walls instead of rebuilding them is the right design. Firms already enforce information barriers in their document systems, and those systems are what their risk teams audit. An AI layer with its own copy of the rules would drift from the original the first time a wall changed. It is the same reasoning we saw in SAP’s agents inside the transaction system: controls that already work are worth more than new ones.

Inheritance, though, raises questions that a preview has to answer before a firm relies on it:

• When is the wall checked? At the moment of each request, or when documents were indexed? A wall created this morning must apply this morning.
• Does it reach derived material? Summaries, drafts, agent memory and search indexes are copies of the matter’s content. If an agent remembers something from matter A and later works for a lawyer walled off from A, the permission on the original document no longer protects anything.
• Who sees the answer? A response shared in a team space reaches everyone in it, whatever the requester was allowed to read.

None of these are exotic. They are what a firm’s general counsel will ask, and the preview period is when to ask them.

In finance, the citation is the product

For financial services, Google lists more than 50 foundational skills, data from FactSet, LSEG, S&P Global, SEC filings and a firm’s own repositories, and outputs with “confidence scores, explicit methodologies, full data lineage for easy auditing, and source citations you can check”. CME Group and Deutsche Bank are named as customers.

Lineage and citations are what make an analyst’s work reviewable, so building them in is the right instinct. But a citation is a pointer, not proof. As we argued in our piece on regulated AI, evidence that an output is right has to come from something independent of the model: a reviewer who checks the source, a reconciliation against a trusted system. A confidence score produced by the system that wrote the answer is a signal, not a control.

Two practical checks belong in any evaluation. The keynote does not say whether the market data from FactSet, LSEG or S&P Global comes with the product or requires a firm’s existing licenses, and those licenses often restrict how data can be redistributed inside generated content. And “full data lineage” should be tested by tracing a sample of answers back to their sources, not taken from a slide.

What a preview is for

Preview means no general availability commitments, and for regulated work it should mean no production obligations. It is also the cheapest moment to test the claims that matter: create a wall and try to cross it, change a permission and time how long the agent takes to notice, trace a figure in an answer to its source.

Industry editions will be judged on domain knowledge in the demos. In production, they will be judged on how faithfully they enforce the rules the firm already has. Evaluate them the second way: by trying to break the rules.