AI Policy

1. What the agents do

Each agent has one named job. Agents work only on your workspace's data and your knowledge base.

• Greeter - answers new enquiries instantly on WhatsApp and Instagram, only from your knowledge base; anything outside it becomes a human handoff.
• Scribe - transcribes call recordings and writes the summary, outcome, and sentiment onto the call record.
• Coach - briefs your callers before a call and reviews their calls afterwards with a score and concrete feedback.
• Chaser - runs the follow-up cadences you define - your own plan, your own words - and stops the moment a lead replies or opts out.
• Closer - drafts proposals from your product knowledge base and nudges clients about proposals awaiting an answer.
• Analyst - scans campaign and source performance and reports findings - junk sources, slow response, wasted budget - with the evidence numbers.
• Herald - writes your managers a morning digest of yesterday's leads, calls, conversions, and breaches.
• Keeper - keeps contact data clean - finds duplicates, fixes phone formats, flags stale records.
• Guardian - watches the 30-minute response SLA and escalates untouched leads before they go cold.
• Reviver - reaches out to dormant leads that have gone quiet, under strict stop conditions.
• Scout - scores and qualifies every lead with a deterministic rules engine, and explains hot leads in one line.
• Host - runs your client onboarding checklist after a conversion - your configured steps, exactly as written.
• Optimizer - recommends advertising changes based on campaign performance. Recommendations only - a human approves before anything is applied.
• Forecaster - projects lead and campaign trends from your own historical data, and labels every projection as an estimate.
• Watchdog - alerts you to ad spend anomalies and creative fatigue.
• Amplifier - builds retargeting segments from your leads - consent-gated, with identifiers hashed before any audience sync, and approval required before a sync runs.

2. Rules every agent follows

• Your rules, not the model's. Agents act within the configuration you set: which agent is on, its autonomy level, its plans, templates, thresholds, and knowledge base. Several agents (Chaser, Keeper, Guardian, Host) are fully deterministic and make no model calls at all - they execute your rules exactly.
• Knowledge-bounded. Conversational agents answer only from your resolved knowledge base. Figures such as prices or percentages that do not appear in your knowledge base are blocked, and the conversation is handed to a human instead. Agents never invent prices, discounts, or commitments.
• Consent-gated outreach. Before any agent sends outreach, consent is checked. A lead who has withdrawn consent, asked not to be contacted, or never gave consent where it is required receives no agent outreach. Consent checks are enforced by deterministic platform services that cannot be paused or overridden by agent settings.
• Scores are computed, not guessed. Lead scores come from a deterministic rules engine over recorded facts. A model may explain a score in one line, but it can never set or move the number.

3. Human approval for consequential actions

Autonomy is yours to set per agent: suggest (the agent records a proposal and does nothing), approve (actions wait for a human decision), or auto (actions execute inside the rules you set). Some actions are recommendation-first regardless of settings: advertising changes from the marketing agents are proposals that a human applies; the Optimizer cannot modify an ad account. Audience syncs by Amplifier are approval-gated. Every proposed action, its approval or rejection, and its result are recorded in the agent's run history and the audit log.

4. Labelling and records

Work done by an agent is labelled as agent work. Agent messages, notes, and actions carry the agent's name on the lead's timeline; agent-generated content in digests, reviews, and insights is identifiable as such; and every agent run is recorded with its trigger, the (masked) input it saw, the actions it proposed or took, and its outcome. Your team can always answer the question "why did this happen?" from the record.

5. Your data and model training

Your workspace data is yours. Vertex Loop does not train shared or foundation models on your private workspace data, and your data is not used to improve models for other tenants. Where an agent uses a cloud AI provider, the provider receives only the minimum content the task needs (for example, masked lead facts for a one-line score summary - names, phone numbers, and email addresses are masked or excluded as the agent's design specifies), and agent usage is metered per workspace against a budget you can set and monitor.

6. You can stop an agent at any time

Every agent has a lifecycle you control: it can be active, put to sleep (with a wake time or a recurring schedule), disabled, or retired - at platform level and per workspace, with the more restrictive setting winning. A per-conversation pause stops the Greeter from replying in that thread. The platform also keeps an emergency kill switch for agent activity. The only services that can never be paused are the consent ledger and consent withdrawal - they protect your leads, not the agents.

7. Limits and responsibility

AI output can be wrong. Summaries, reviews, recommendations, and forecasts are decision support: forecasts are labelled estimates, recommendations state their evidence, and nothing an agent produces is a guarantee of a business outcome. You remain responsible for the configuration you choose, for approvals you give, and for having a lawful basis - including consent - for the outreach your workspace runs, as set out in the Terms & Conditions and Privacy Policy.