How it works
This is a control plane for AI agents that spend company money. Each agent is given a budget and a written set of limits; every purchase it attempts is checked against those limits before any money moves. Spend inside the limits clears automatically, exceptions go to a named approver, prohibited spend is refused — and the whole chain is recorded as it happens. Written for finance and compliance teams.
Sandbox only. No custody, no customer funds and no live payment execution.
What this is
The pilot exists to prove the control model on one rail, validate it with a design partner against real operating decisions, and inform a focused venture build on regulated euro infrastructure. Nothing in this environment is production, licensed or customer-facing.
Payments are simulated
There are no wallets, no custody, no card issuance and no payment integrations. When a case is approved, the control plane writes a simulated transaction with a SIM- reference and records it against the mandate's monthly authority. Nothing leaves the sandbox.
Fixed rules, not AI judgement
No language model decides whether a purchase is allowed. Each intent is evaluated by ordinary application logic in a fixed order, and every check returns a readable reason.
- 01Category restrictions — blocked or outside the permitted listblocked
- 02Per-transaction ceilingblocked
- 03Remaining monthly authorityblocked
- 04Human approval thresholdrouted
- 05New-vendor rulerouted
The Policies screen is the same idea in reverse: a plain-language clause compiles into explicit conditions you can inspect before they apply.
Who approves what
- Within mandatePolicy engine
- Deterministic evaluation. No human review required.
- Commercial above thresholdDaniel Osei — Business Sponsor
- Commercial fit, partnership context, and market data, tooling or growth spend above the agent threshold.
- Compliance & recurring regulatory dataLena Fischer — Compliance Approver
- Every recurring subscription and all regulatory, KYC/KYB or legal sources.
- Scope change or exceptional limitAlex Vance — Executive Sponsor
- Resources, go/no-go, mandate expansion, new categories, or an exception to a published limit.
- Internal execution & platform deliveryDana Reeve — Head of Platform
- Agent runtimes, integrations, control-plane configuration, project cadence and coordination across tech, business and compliance.
- Framework architecture & pilot substancePriya Raman — Authority Framework Architect
- Product thesis, mandate and control-model design, pilot design, initial product deliverables, external partner strategy and communication.
Governance
- Executive Sponsor
- Alex VanceFounder & Chief Executive Officer (CEO)
- Head of Platform
- Dana ReeveInternal execution, platform delivery and pilot cadence
- Authority Framework Architect
- Priya RamanFramework architecture, pilot substance, partner strategy and communication
- Business Sponsor
- Daniel OseiCo-Founder & Chief Business Officer (CBO)
- Compliance Approver
- Lena FischerCompliance & internal control
Demonstration governance for the Northstar Systems workspace. Every name below is fictional sample data used to show how authority, approval and revocation are attributed.
The audit trail
Every case produces the same ordered chain: principal instruction → agent intent → policy decision → approval → simulated payment → result. It is written server-side as decisions are made, not reconstructed afterwards, and read back in the audit trail.
What we store
Organizations, member profiles, agents (each carrying its mandate and authority), policy clauses, purchase intents, simulated transactions and audit events. Access is scoped to the signed-in user's workspace at the database level.
The example setup
Northstar Systems runs three mandates: a Research Agent (€1,000/month, €50 per action, research APIs, datasets and analyst material, approval above €100 or for any new vendor), a Procurement Agent #03 (€5,000/month, €250 per action, cloud, software and developer tooling from approved vendors, new beneficiaries blocked) and an Operations Agent (€10,000/month, €500 per action, approved operations tooling, approval above €250, fund transfers and digital assets blocked). “Run demo request” on the Control screen has the Research Agent buy a low-value euro-area dataset — inside its mandate, so it auto-approves and writes the full chain.
What is not included
Live payment rails, custody, vendor integrations, per-user roles beyond a single operator role, notifications, and clause versioning. No claim of production readiness or regulatory approval is made for this pilot.