Guardrails and governance

Governance in AINOW comes in two halves that work together. Guardrails constrain what Zoe may do before she does it. The audit trail records what she actually did, so you can check the constraint held. Both are administrator-facing and both live in the portal's left-hand navigation, under Guardrails and Audit.

Guardrails

A guardrail is a policy Zoe must follow on every channel — chat, voice, phone, email and scheduled automations alike. The Guardrails page lists the catalog grouped into five categories — Privacy, Security, Compliance, Safety and Quality — with a severity of Low, Medium or High on each, and you select the ones to enforce.

Each policy is enforced one of two ways, shown on its row as Instruction, Live scan, or Instruction + live scan. An instruction guardrail becomes a mandatory directive in Zoe's prompt. A live-scan guardrail is deterministic: it inspects messages as they happen and masks what it finds, regardless of what the model intended. The Live input/output scanning toggle controls that second layer for the whole workspace; with it off, guardrails apply by instruction only.

What the policies constrain, in practice:

• Privacy and security — masking third-party emails, phone numbers and government IDs; blocking disclosure of API keys, passwords, tokens and connection strings; masking payment card numbers to the last four digits; blocking data exfiltration to addresses the user did not legitimately provide; and a prompt-injection shield that tells Zoe to treat instructions found in documents, tool results or pasted content as data, never as commands.
• Compliance — declining financial, medical and legal questions and directing people to a qualified professional, and refusing competitor endorsements.
• Safety — profanity masking, a toxicity and hate filter, and a self-harm safeguard that responds with empathy and crisis-support resources.
• Quality — keeping Zoe on work-related topics, telling her not to invent facts, figures or sources, and holding a professional tone whatever the user's tone.

Select Save guardrails to apply your choices; the page confirms how many are active. Results appear on the Dashboard in a Guardrail compliance card, which reports how many audited conversations came back clean, how many spans were masked live, and a breakdown by channel and by guardrail.

The audit trail

The Audit page lists your organization's recorded conversations. Filter by window — 7 days, 30 days or 90 days — by user, by surface (Chat, Workflow, Email, Voice, Phone), and by whether a conversation contained approvals, consults with other agents, or guardrail hits. Each row shows when it started, who it was, the surface, an activity summary counting model calls, tool calls, approvals, consults and guardrail hits, and the state of its Trail.

The trail is tamper-evident, which is what the Trail column reports: Chained means the conversation carries an intact record, Redacted means content was erased from it, and Before the ledger marks conversations recorded before the trail existed — those are shown honestly as such rather than reconstructed after the fact.

Replaying a conversation

Open any row to replay it. The replay page shows a verification badge and a timeline grouped by turn: every model call with the prompt and tools it was offered, every tool call with its arguments and result, drafted changes, approval decisions, executions, skills loaded, consults, memory events and guardrail hits.

1. Select Verify now to re-check the trail on demand.
2. Read the badge: Verified, Verified · redacted content, Integrity check failed, or No trail.
3. Use Reveal on an individual entry, or Reveal all content for the whole conversation.

Revealing is itself an audited act. The confirmation says so plainly — every message, prompt and payload will be shown to you, and the reveal is recorded on the conversation's own trail with your identity. Metadata is open to administrators; the content behind it is not read without leaving a mark.

The Explain view

The audit page answers "what happened in this conversation". The Explain view answers "why this answer". In Zoe, an administrator sees an Explain chip under a completed reply; opening it shows Why this answer with the reply's own evidence grouped into Trigger, Model calls, Tools, Sources, Skills, Delegations, Memory, Guardrails and Approvals, plus Open full replay to jump to the whole conversation.

The same reveal rule applies: expanding content is recorded on the trail. Voice turns show a notice instead, because voice tool calls are recorded against the conversation rather than per turn — open the full replay for those.

Who can see what

• Administrators see the Guardrails page, the Audit page and the Explain chip. Explain is deliberately not shown to end users.
• Members use Zoe and see their own conversations; they do not see other people's, and they do not reach the portal.
• Reveals are attributable. Every time an administrator reveals conversation content, or views a member's saved memories from the Members card, that act is recorded with their identity.
• AINOW staff have their own trail: when AINOW support reveals content while helping you, the reveal is logged on AINOW's side as well.

Approvals — the other half of governance — are covered in Records, queue and approvals.