arrow_backBack to the overview
Governance & trust

An instruction is a preference. A check is a control.

Compliance belongs in the platform, not in the prompt.

A worker making outbound contact operates in a regulated environment. Every obligation below is enforced before the worker is invoked, not delegated to a model and hoped for. That distinction is what makes the difference between an AI system that is interesting and one a regulated institution will actually deploy.

Before dialregistry checks block at the orchestration layer
InjectedAI disclosure is not left to the model
Sealedmaterial commitments carry a replayable warrant
Platform enforcement

Seven controls that run whether the model cooperates or not.

block

Registry checks

Automated pre-contact check against every applicable do not call registry, blocked before the worker is ever invoked.

campaign

Disclosure injection

AI identity disclosure injected by the platform into every outbound voice call, in the contact's language, non skippable.

receipt_long

Consent ledger

A per contact, per purpose immutable record: what, when, through which channel, and on what basis.

schedule

Calling windows

No outbound contact outside the configured window for the contact's jurisdiction and timezone.

tune

Capture level selection

Automatic selection of the conversation record depth by jurisdiction and consent, with no override permitted below the minimum.

auto_delete

Retention automation

Automatic flagging for review and deletion at the end of the applicable retention period.

history

Audit trail

Every contact attempt, consent check, registry lookup, disclosure delivery and capture decision logged with the rule matched.

verified_user

Validity warrants

Material assertions and commitments sealed with a signed, replayable record of what was known and under whose authority.

Multi-jurisdiction

The rules change by market. The architecture does not.

Singapore is the first deployment jurisdiction. The framework extends to others by configuration rather than by code, which is what allows a single campaign to run correctly across several markets at once.

JurisdictionKey additional requirements the engine applies
SingaporeDo not call registry checked before every attempt, AI disclosure as a mandatory step, consent recorded before personal data is collected, purpose limited usage, and configurable calling windows even where none are statutory
United StatesAI voice treated as artificial or pre-recorded, prior express written consent for consumer contact, disclosure within the opening seconds, the national registry honoured, records retained for five years
European UnionConsent or legitimate interest, the right to be informed about AI processing, an impact assessment for large scale calling, opt out at every contact point, data minimisation
United KingdomConsent for automated marketing calls with a business to business exemption where relevant to the role, and both preference registries checked
AustraliaConsent for commercial calls, the register honoured, and an unsubscribe mechanism
IndiaPreference register honoured, calling restricted to daytime hours, disclosure as best practice
public
One campaign, several rule sets. A contact in one market with full consent yields complete transcripts. A contact in another who has not consented yields structured summaries only. A contact on a register is blocked entirely. All of it is transparent to the worker, because the compliance engine handles it at the platform level.
Recording

Audio is a capture level, not a switch.

A transcript containing a person's name and statements is personal data under virtually every jurisdiction, whether or not audio was retained. Recording extends the existing ladder rather than introducing a parallel mechanism, which means it is selected per interaction and never set globally.

Highest burdenFull transcript with audio

The retained recording, enabling prosodic and acoustic analysis. Explicit, informed and separately obtained consent, disclosed before recording begins and repeated if a third party joins.

Explicit consentFull transcript

Complete text with speaker attribution in the language spoken, plus the parallel record language rendering.

Lower burdenStructured summary

Topics, qualification elements, action items and sentiment, without verbatim quotes or personal attribution. The recommended default where full consent is difficult.

Lowest burdenAction items only

What was agreed, the next step, the follow up. Least intelligence value, but the lowest consent requirement.

do_not_disturb_on
Refusal is a first class path. A contact who declines recording gets the conversation anyway, at the next capture level down, with the refusal written to the consent ledger and honoured on every future interaction. A worker that argues with a refusal, or continues after one, is the most serious class of defect available to this platform, and it is tested as its own cell in the simulation matrix rather than covered by a policy document.

Purpose travels with the recording

"For training and review purposes" permits quality review, coaching and worker calibration. It does not by itself permit use as evidence in a commercial dispute, feeding a third party model provider, or retention beyond the stated period. Purpose is enforced at the platform, and each recording carries the basis it was collected under.

What audio makes possible

  • Sentiment as a trajectory, not a label: the start to end delta, the turn where it moved, and whether it recovered
  • Prosodic signal: pace, pauses, interruption, overtalk, hesitation. Text says a contact replied "that's fine". Only audio says how
  • Turn metrics: talk and listen ratio, response latency, silence handling
  • Cross language sentiment is calibrated per language and never used alone as the basis for a judgement about a person
Security and access

Relationship based, because the rules are relational.

Almost every meaningful rule in this system is relational rather than positional. "A seller sees their own accounts" is not a role, it is a relationship between a person and a set of accounts. Expressing that as roles produces a combinatorial mess. Expressing it as relationships produces a small, readable policy.

Resources worth protecting

Tenants, teams, accounts, contacts, conversations, recordings, worker definitions and their individual prompt layers, knowledge bases, assets, skills, transactions, coaching records, requests and approval queues.

Relationships, equally few

Member of a team, owner of an account, assigned to a worker, manager of a team, administrator of a tenant, and subject of a coaching record. Authorisation runs on Permit.io and is enforced at the API, never in the interface.

smart_toy
Workers are principals in the same model. A worker holds an identity and is subject to policy exactly as a person is. That is how the rule that a worker cannot modify its own prompt, guardrails or compliance configuration becomes a technical fact rather than an instruction, and how a supervisory worker's bounded authority is enforced rather than trusted.
key

Authentication

Federated identity with single sign on and passwordless entry, and two factor required for all operational roles.

shield_lock

Data classification

Public, internal, confidential and restricted, with access control enforcing the classification rather than describing it.

domain

Tenant isolation

Enforced at the database level rather than only in the application, with credentials and configuration encrypted and stored separately.

lock

Infrastructure

Stateless autoscaling services, secrets in a managed store scoped to the service account, encryption at rest and in transit.

Behaviour

What a worker will never do.

Guardrails are part of the worker definition and are enforced by the kernel prompt layer, which a customer configuration cannot weaken.

Next

Governance you can evidence, operations you can budget.

Telemetry, per interaction cost capture, entitlements and caps are what turn a governed system into an operable one.