Trust Architecture

Human Approval

Every action Freya takes requires a designated human to authorise it. Not as a safeguard. As a structural requirement — baked into every workflow, every draft, and every output that leaves your business.

Approve before execute
Named human accountable
Permanent approval record

Gate: Approve before execute
Accountability: Named human accountable
Evidence: Permanent approval record

Section 01

Why human approval is structural, not optional

Most AI systems describe human oversight as a feature. In Antarious AI, it is an architectural constraint — baked into every agent, every workflow, and every output that Freya produces for your business.

Freya is designed to do two things exceptionally well: prepare, and present. She monitors your operations, synthesises your data, drafts your posts and replies, and surfaces your decisions — with full context, cited sources, and a clear recommendation. But she does not act on any of it until a designated human says so.

This is not a limitation of the technology. It is a deliberate design position, rooted in three convictions:

1. Accountability cannot be delegated to software

In small business, GTM, and customer-facing operations, decisions carry consequences — legal, financial, reputational, and relational. Customer messages, published content, sales outreach, and invoice reminders all go out under your business name. Accountability must remain with the people whose names, roles, and judgement are behind every outcome.

2. Context that AI cannot access always exists

Freya has access to your connected business data. She does not have access to your customer relationships, your brand voice nuances, your competitive timing, or your ethical judgement. Humans do. Approval is the point where that knowledge enters every decision.

3. Trust requires verifiability

Stakeholders — customers, partners, boards, investors, and regulators — need to know that a human was responsible for every significant action. The approval record is that proof. When someone asks who authorised a message, a campaign, or a payment chase, you can show them.

Section 02

The approval architecture

Every workflow in Antarious AI is built around an Approve Before Execute model. Freya completes her preparation — analysis, drafting, enrichment, synthesis — and then halts. Execution does not resume until an authorised approver acts.

The four-stage approval flow

Stage 01
Freya prepares
Freya completes the task: draft, reply, outreach sequence, invoice reminder, report, alert, or action recommendation. Full context is assembled — sources, rationale, and recommended next steps.
Stage 02
Review package delivered
The designated approver receives a structured review package containing the proposed action or output, the data sources and reasoning behind it, any flagged risks or anomalies, recommended next steps, and clear approval options.
Stage 03
Human decision point
The designated approver reviews and chooses: Approve (Freya executes exactly as prepared), Approve with edits (approver modifies, Freya executes the revised version), Request revision (Freya re-drafts with new instructions), or Reject (action is cancelled and logged with reason).
Stage 04
Execution + audit record
If approved, Freya executes. The approval decision, approver identity, timestamp, and any edits are permanently recorded in the audit trail — available for internal review, board reporting, or external inspection.

Section 03

Scope of approval requirements

What requires human approval

The following categories of action require explicit human authorisation before Freya proceeds. This list reflects defaults for small businesses; you may expand approval requirements through configuration.

Action category	Examples	Default approver level
External communications	Customer DMs, emails, reply drafts delivered to inboxes or messaging channels	Owner / designated teammate
Content publishing	Social posts, scheduled captions, campaign content published to connected channels	Owner / marketing lead
Sales outreach	Lead outreach sequences, follow-up messages, proposal sends, deal reminders	Owner / sales lead
Financial actions	Invoice reminders, payment chase messages, overdue follow-ups to customers	Owner / finance authority
Campaign execution	Ad or outreach campaigns activated, multi-step sequences launched, timed publishes	Owner / marketing lead
System changes	New integrations activated, workflow configurations updated, permission changes	System administrator

What Freya does without approval

The following internal operations run by design without a human gate — they are monitoring, preparation, and intelligence functions that carry no external consequence:

Continuous data ingestion and processing from connected sources
Internal dashboard updates, activity summaries, and KPI calculations
Draft preparation and document generation (queued for review, not published)
Lead detection, pipeline organisation, and CRM enrichment inside the app
Anomaly detection and internal alert generation
Institutional memory updates and context retention
Funding profile updates from live business activity (internal view only)
Scheduled report drafting delivered for review, not published externally

Section 04

Configuring approval workflows

Antarious AI does not impose a single approval model. Every deployment is configured to reflect how your business actually works — who owns posts, who owns money, and who covers when someone is away — without removing the human gate on external actions.

Configuration options

Single-approver workflows

One designated person reviews and approves. Suitable for day-to-day posts, replies, and outreach in owner-led or streamlined teams where speed matters and stakes are routine.

Sequential multi-approver workflows

Approval passes through a defined chain — for example, teammate → manager → owner — before execution. Each approver sees the previous approver's comments and decision, so context accumulates rather than resets.

Parallel approval workflows

Multiple approvers review simultaneously. Execution triggers when a quorum or unanimous agreement is reached. Suitable for actions that need cross-functional sign-off — for example, a campaign that touches both marketing and finance.

Conditional approval routing

Approval routing is determined by the nature of the action. A social media post routes to the marketing lead; an invoice reminder routes to the finance owner. Rules are configured during setup and can be updated at any time by designated administrators.

Delegated approval

Approvers can delegate authority to a named deputy for a defined period — for example, during leave or travel — without removing the approval requirement from the workflow. The delegate's identity is recorded on every decision they make.

Approval thresholds

Certain actions may require escalation based on scale. For example: routine invoice reminders require owner approval; high-value or sensitive chase sequences require a second sign-off. Campaign spends below a configured threshold require team-lead approval; above that threshold require owner or director approval. Thresholds are fully configurable.

Section 05

The approval interface

How approvers experience the review process

Approvals are delivered to designated approvers through their configured channels — the Antarious AI platform, email notification, or integrated communication tools such as Slack or Microsoft Teams.

Each approval request contains six components:

1. Action summary

A plain-language description of what Freya is proposing to do and why — enough for a busy owner to understand the ask in seconds.

2. Supporting context

The data, analysis, or reasoning that informed the recommendation. Sources are cited. Assumptions are stated. No guesswork about why Freya prepared this draft.

3. Proposed output

The exact content that will be executed — the email text, the social post, the outreach sequence, the invoice reminder — presented for review before anything is sent or published.

4. Risk flags

Any anomalies, conflicts, tone concerns, or uncertainties Freya has detected that the approver should be aware of before deciding.

5. Decision options

Approve / Approve with edits / Request revision / Reject — with a free-text field for instructions, edits, or reasoning that becomes part of the permanent record.

6. Deadline indicator

Where the action is time-sensitive, the approver is shown the window within which a decision is needed and what will happen if no decision is made. Default: the action is held, not auto-executed.

Section 06

What happens without approval

Non-approval handling

Antarious AI is designed to hold, not auto-execute, in the absence of a human decision. Silence is never treated as consent.

No response from approver: The action remains in a pending state. Freya sends a reminder after a configurable interval (default: 24 hours). If a deadline is critical, Freya escalates to a secondary approver if one is configured.
Rejection: The action is cancelled. The rejection reason is logged. Freya does not retry unless instructed to re-draft and re-submit.
Approver unavailable: If delegation is configured, approval routes to the delegated deputy. If not, the action holds until the approver is available or an administrator manually reassigns.
System failure during approval: If a technical failure interrupts the approval process, the action does not proceed. The system logs the failure and notifies administrators.

Under no circumstances does Freya self-approve an action or proceed on the assumption of approval.

Section 07

Accountability records

Every approval is permanently recorded

The approval record for every action is stored permanently in the audit trail and contains:

Who approved — Full name, role, and user ID of the approver
When they approved — Timestamp to the second (UTC)
What they approved — The exact version of the output or action at the point of approval
Any edits made — A tracked diff of changes made by the approver before authorising
Their stated reasoning — Any commentary or instructions entered during review
The outcome — Whether the action was executed, revised, or rejected

This record cannot be altered or deleted after the fact. It is available for internal review, external audit, regulatory inspection, or legal disclosure at any time.

Section 08

Compliance & regulatory relevance

Human approval as a compliance mechanism

References to ISO 27001, GDPR, NCSC, and similar frameworks on this site describe control alignment and internal self-assessment — they are not claims of third-party certification unless we say so explicitly.

For businesses operating in regulated or high-accountability environments, the Antarious AI human approval architecture provides direct support for compliance obligations across multiple frameworks.

Regulatory / framework context	How human approval supports compliance
UK GDPR / EU GDPR	Human review of any automated decision with legal or significant effect on individuals — Article 22 aligned by design
ISO 27001	Documented access control and change management processes supported by approval workflows
Business finance accountability	Full approval chain for any action with financial consequence — invoice reminders, chase sequences, campaign spends; audit evidence auto-compiled
AI Act (EU) — high-risk AI systems	Human oversight provisions supported through a mandatory approval gate on all consequential outputs

Note: Antarious AI does not provide legal advice. Organisations should confirm compliance mapping with their own legal and compliance counsel. The above reflects how the human approval architecture is designed to support — not substitute for — compliance obligations.

Section 09

Frequently asked questions

Can approval be turned off for certain workflows to increase speed?

Approval requirements for external actions cannot be disabled — this is a platform-level constraint. However, approval workflows can be streamlined: single-click approval for lower-risk actions, pre-approved template libraries, and delegated authority all reduce friction without removing the human decision point.

What if our business needs approvals from multiple people in different locations or time zones?

Antarious AI supports asynchronous approval workflows. Approvers receive notifications through configured channels and can act from any location. Parallel and sequential workflows can be combined to accommodate distributed teams.

How long does an approval typically take?

Freya prepares review packages that are designed to be actioned in under five minutes for routine approvals. The review package surfaces everything the approver needs to decide — no additional research is required from the approver.

Who configures approval workflows?

Approval workflow configuration is completed during onboarding by the Antarious AI implementation team, in consultation with your owners and operations leads. Workflows can be updated post-deployment by designated system administrators.

Is there a mobile interface for approvals?

Yes. Approvers can review and action approval requests from any device. Email-based approval is also supported for businesses where platform access is not always available.

What happens if the approver approves something incorrectly?

The audit trail records the approval with full context. Freya cannot retroactively undo an approved action, but the record of what was approved, by whom, and on what basis is permanently available. Post-approval rollback procedures depend on the action type and are configured during onboarding.

Can I edit what Freya wrote before it goes out?

Yes. Choose Approve with edits to modify the output yourself, Request revision to send Freya new instructions, or Reject entirely. Edits are tracked in the approval record.