For two years the conversation between regulated firms and AI vendors had a familiar shape. The firm asked where the data went, who could see it, what the agent could reach and what it would cost; the vendor pointed to a policy. This summer the answers started arriving as product. The shift is visible across the major platforms, and Anthropic's recent releases are a clear example of the pattern.
1. Data stays with the customer
Anthropic's Enterprise Frontier Safeguards, announced in early September and developed with more than a hundred customers in financial services, law, healthcare and the public sector, lets the vendor detect misuse while the customer keeps a zero-data-retention arrangement. Data is stored on the customer's own cloud rather than the vendor's systems, and any human review is by default done by the customer. For a firm bound by professional confidentiality, that sentence is the difference between a tool it can and cannot deploy on client matters.
2. Agents get a budget and a fence
Recent platform release notes add a hard spend cap per managed-agent session, allow- and block-lists for the domains an agent's web tools may reach, effort settings that trade cost against depth, and webhooks for environment and memory events so the customer's own systems can watch what an agent does. Each of these was, twelve months ago, something a firm's IT team had to build around the vendor. Now it is a parameter.
The question to a vendor has changed from "what is your policy?" to "show me the setting."
3. Administration catches up
Less glamorous but as important: administrative APIs for members, groups and custom roles have left beta, so a firm can provision, scope and revoke access from its own identity systems rather than by hand. Skills, reusable instruction sets that tell a model how a particular firm does a particular task, are now a standard part of the API rather than an experiment. The building blocks of a firm-specific, governed deployment are becoming ordinary.
4. How to read it
- It confirms the buying pattern. Model quality changes quarterly. The controls, integrations and memory a firm builds around models are the durable asset, and the vendors are now competing on them.
- It raises the bar for everyone else. Any tool a firm considers can now be asked for the same features: customer-side data, per-workflow scoping, spend caps, exportable logs. "We are working on it" is a weaker answer than it was in June.
- It does not do the work for you. A setting that exists but is not configured is a policy document with a different font. Someone in the firm has to own the configuration and read the logs.
5. The direction
The interesting thing about this summer is not any single feature. It is that the vendors have concluded the enterprise market is won on control, and that regulated professional firms are the customers who define what control means. That is a good position for those firms to be in, provided they ask for what is now available.
Hive Newsroom follows what is changing in AI and professional services. Sources are linked in the text; figures are as published at the time of writing. Comments and corrections: press@get-hive.ai. Back to the .