Irreversible mistake
Wrong payment, contract sent without review, email shipped to the wrong customer. High-impact actions without human approval turn into labor, financial and reputational liabilities.
Synapse doesn't replace people — it amplifies decisions. Human approval where it matters, autonomy where it's safe, an auditable trail throughout. Your company adopts AI without losing control, without breaking policy, without surprises in an audit.
Mid-market companies adopting AI without a governance framework learn the cost of "speed first" early. Irreversible mistakes, leaked regulated data and audits with gaps are the three scenarios that stall roll-out — and the ones HitL governance exists to neutralize.
Wrong payment, contract sent without review, email shipped to the wrong customer. High-impact actions without human approval turn into labor, financial and reputational liabilities.
A squad ships personal data to an external LLM, the model retains it in history, the regulator finds it. Without granular RBAC and PII redaction, your team breaks LGPD silently — and the fine arrives later.
"Who approved this payment?" "The AI." "Which document was used?" "No log." Without per-agent and per-action audit trail, your company is indefensible in a compliance review.
HitL isn't "everything needs approval" or "AI decides everything alone." It's a framework that classifies each action by risk, requires a human at the critical points and allows autonomy where policy permits — with 100% auditable trail on both sides.
Every action is classified by impact and reversibility. Sending an internal summary is low. Paying a vendor is high. Deleting a record is critical. The framework sets the tier.
You configure it: "amounts above R$ 50k approved by CFO", "deletions only by director", "external customer contact goes through the account manager". The squad obeys your business policy.
When the tier calls for people, Synapse pauses the action and routes to the right approver. With full context of what is about to happen, the source backing it and the expected impact.
Every decision — human or AI — becomes a line in an immutable log: who, when, on which source, with what result. Exportable for compliance, board or regulator, with no friction.
Serious governance isn't just "approval in the flow" — it's architecture that separates roles, isolates data, logs everything immutably and exports in a format the regulator accepts. Here is how Synapse implements each piece.
Human approval is part of the agent definition, not a plugin. Policy says "when, to whom, with what context". The agent pauses, waits, resumes — without rework and without dropping state.
Append-only log with chained hash: tampering breaks the chain, evidence preserved. Every LLM call, every prompt, every human decision — with trusted timestamp and compliance export.
RBAC with 113 controllable actions — read, write, export, delete, approval override, sensitive data access. Default-deny: whatever wasn't explicitly allowed is blocked.
Same platform, many isolated spaces. The holding consolidates — each branch sees only its own. Useful for enterprise groups, franchise networks and multi-country operations with distinct regulatory requirements.
Encryption at-rest and in-transit. Native PII detection blocks personal data before it becomes an external prompt. Configurable retention, right-to-erasure via API, auditable logs for the DPO.
Dashboard shows approval rate per agent, average human latency, reversal rate. You discover where the policy is too tight and where it's too loose — before it becomes an incident.
No Squads, Your Base or Multi-LLM decision goes around governance. It is the layer that makes Synapse safe adoption for regulated mid-market.
Every squad action consults the policy before executing. If the tier requires HitL, the squad pauses and routes. No hack, no bypass — behavior guaranteed.
Documents containing personal data are tagged at ingest. The router blocks these chunks from external models. Human approval required to make the source active.
Every chosen model gets logged with reason and cost. Compliance sees which provider processed what. Policy can force an on-prem model for sensitive data.