What does an AI governance platform have to do?

Three jobs to check for, and one question to ask about your data.

Which three jobs should you check a platform for?

  • Enforce

    Your policy allows, redacts or denies each request on the device.

  • Audit

    Each captured interaction goes on a tamper-evident record.

  • Contain

    Contain one person, or turn on the kill switch, when AI isn’t working as intended.

What happens to a request before it leaves the device?

Your policy allows, redacts or denies each request on the enrolled device. Then it goes on your record.

Allow

What your policy didLocal model
  1. StaffAllowed
    Summarize tonight’s council agenda
  2. Local modelRecorded
    Here’s the summary. The budget vote comes first, then two zoning items.
Your record, on your serversSigned rows
  1. 14:02:08PromptAllowedSummarize tonight’s council agenda
  2. 14:02:11ResponseRecordedHere’s the summary. The budget vote comes first, then two zoning items.

Redact

What your policy didCloud AI provider
  1. StaffRedacted
    Reply to the caller at 303-555-0142
    Sent:Reply to the caller at [PHONE_REDACTED]
  2. Cloud AI providerRecorded
    Here’s a draft reply.
Your record, on your serversSigned rows
  1. 14:03:20PromptRedactedReply to the caller at [PHONE_REDACTED]
  2. 14:03:24ResponseRecordedHere’s a draft reply.

Deny

What your policy didCloud AI provider
  1. Staff, in an AI agentAllowed
    Clean up the 2019 case files
  2. Tool call from the providerDenied
    Delete the 2019 case folder
    For the agent tools your policy governs, a call it denies doesn’t run.
Your record, on your serversSigned rows
  1. 14:04:02PromptAllowedClean up the 2019 case files
  2. 14:04:06Tool callDeniedDelete the 2019 case folder
  • An enrolled laptop
  • AI tools your staff use
  • Tools, on the device
  • Your policy, on the device
  • Local models
  • Cloud AI providers

Allow

What your policy did

  • Staff: Summarize tonight’s council agenda (Allowed)
  • Local model: Here’s the summary. The budget vote comes first, then two zoning items. (Recorded)

Your record, on your servers

  • 14:02:08 Prompt: Summarize tonight’s council agenda (Allowed)
  • 14:02:11 Response: Here’s the summary. The budget vote comes first, then two zoning items. (Recorded)

Redact

What your policy did

  • Staff: Reply to the caller at 303-555-0142 (Redacted)
  • Sent: Reply to the caller at [PHONE_REDACTED]
  • Cloud AI provider: Here’s a draft reply. (Recorded)

Your record, on your servers

  • 14:03:20 Prompt: Reply to the caller at [PHONE_REDACTED] (Redacted)
  • 14:03:24 Response: Here’s a draft reply. (Recorded)

Deny

What your policy did

  • Staff, in an AI agent: Clean up the 2019 case files (Allowed)
  • Tool call from the provider: Delete the 2019 case folder (Denied)
  • For the agent tools your policy governs, a call it denies doesn’t run.

Your record, on your servers

  • 14:04:02 Prompt: Clean up the 2019 case files (Allowed)
  • 14:04:06 Tool call: Delete the 2019 case folder (Denied)
A server room aisle lined with racks.

Where does your data go?

Verillian is self-hosted, so there’s no Verillian cloud to assess. The record stays inside your organization, under keys you hold. That helps you meet GOVERN 6.1 in NIST’s AI RMF 1.0, which asks for policies on the risks third parties bring.

How much staff time does it take to run?

Little, once your policy is set. Verillian works at the device, so there’s nothing to set up inside each AI tool, and the record builds as staff work. One screen shows the week’s allowed, denied and redacted requests.

Verillian console, the chart Denials and redactions over time for the last 7 days, through 15 Jul, with each day’s denied and redacted requests.

What should you read before you choose one?

All posts

What a system of record for AI use is, and where it should live.

See the three jobs run on your own servers.