Writing / Compliance

Where does the record of your staff's AI use live?

October 27, 2026 / 4 min read

Browser extensions, network gateways, your productivity suite and your own servers each keep the record of staff AI use in a different place. For a public agency that answers records requests, that place matters most.

Every tool for governing staff AI use keeps some record of it. They differ most in where that record lives, and for an agency that answers public records requests and audits, that matters more than any feature. The record sits in one of four places: your productivity suite, the vendor’s cloud, a gateway on your network, or servers you run. Each is a fair choice for some organizations. The question to ask is who can search it when a request arrives.

In your productivity suite

Microsoft Purview keeps AI activity with the rest of a user’s Microsoft 365 content. Its documentation says “user prompts and responses for AI apps are stored in a user’s mailbox” and “prompts and responses are captured in the unified audit log.” For AI apps outside Microsoft’s own, most features “require the Microsoft Purview browser extension and devices onboarded to Microsoft Purview.” Once the logs for those apps are turned on, “the system retains these audit logs for 180 days.”

What it means for a records officer: if your agency already searches Microsoft 365 for records requests, AI prompts land in a place your staff know. Check the retention period against your retention schedule.

In the vendor’s cloud

Many AI security products run as a service: a browser extension or endpoint agent on staff devices, and a console the vendor hosts. Many of their public pages do not say where prompt logs are stored, so ask.

What it means: if the record sits with the vendor, a records request or a legal hold runs through that vendor’s console and contract.

At a gateway on your network

Network and gateway products see AI traffic as it passes. Lasso says it enforces policies “inline at the proxy, API, or AI Gateway layer.” Netskope’s AI Gateway can also run in your own environment: “In your own deployments (VPC or on-prem), AI Gateway enables applications to access LLMs securely.” SentinelOne says the Prompt Security technology it acquired “fits into any stack, whether SaaS or self-hosted.”

What it means: a gateway you host keeps the record closer to you. Ask which traffic passes through it, since a gateway records what reaches it.

On servers you run

The last approach puts both the control and the record on your own infrastructure, under your own key, so no vendor holds a copy. This is the approach Verillian takes: the record of every captured AI interaction stays on your agency’s own servers, ready to search and export.

What it means: a records request is a search your own staff run, with no vendor to ask.

Questions to ask any vendor

  • Where is the record of each AI prompt and reply stored, and in whose account?
  • How long is it kept by default, and can we match our retention schedule?
  • Who can search it for a public records request, and how long does an export take?
  • If we leave the product, what happens to the record?

For a side by side view of the four approaches, see Approaches to governing AI, compared.

Vendor statements are quoted from each vendor’s own public pages, read on 27 October 2026. Products change; check each vendor’s current documentation.

All writing

See Verillian on a real device.

The Verillian console home for the last 30 days: 10 devices, 684 requests, 613 allowed and 71 denied, the Enforce, Audit and Contain cards with the two active holds, and the latest refused requests.The Verillian console home for the last 30 days: 10 devices, 684 requests, 613 allowed and 71 denied, the Enforce card with two suggested policies waiting, and the top of the Audit card.