Independent control layer for agent-initiated spending

Give every AI agent a defined financial mandate.

ACI is an early-stage software concept exploring a provider-independent policy layer for AI agents. It is designed to govern what each agent may spend, where it may transact, and when a person must approve — with issuer-rail enforcement as the initial integration path.

Research prototype · Simulated decisions only · No live financial integrations

Authority OverviewSIMULATION
Procurement AgentFramework: custom · Allowed merchant: Viking Direct
£500 limit£218 illustrative balance
Finance AgentFramework: n8n · Approval over £100
£1,000 limit£740 illustrative balance
Ops AgentFramework: CrewAI · Status: frozen
£250 limitAuthority revoked
Static product illustration. These are example values, not real accounts, funds or payment permissions.
The research question

Do teams need one consistent policy layer across the AI agents they deploy?

We are testing whether existing spend controls are sufficient when multiple agents, frameworks and payment routes are involved — or whether a separate authority layer would solve a real operational problem.

Who we want to speak with

  • Businesses deploying AI agents that can initiate or commit purchases or payments.
  • Builders and integrators working on purchase-capable agent systems.
  • People responsible for finance, procurement, payments or agent deployment.

What the conversation is

  • A 15-minute research conversation, not a sales pitch.
  • An honest discussion of current workflows, controls and gaps.
  • No request for confidential records, customer data or payment credentials.
The proposed architecture

Separate the decision about authority from the agent doing the work.

The intended model is deterministic policy enforcement around agent-initiated spending. The current website is explanatory; it does not connect to payment rails.

LAYER A · SIMULATED

Issuer-rail controls

Potential controls include agent identity, merchant restrictions, transaction ceilings, shared budgets, velocity rules and approval thresholds. The current demonstration is a simulation of ALLOW / DECLINE decisions using the authorisation context available to an issuer integration.

Agent request
Policy engine
Issuer decision
LAYER B · PARTNER-DEPENDENT

Exact-purchase verification

Ordinary card authorisation metadata does not usually contain basket line items or SKU detail. Checking whether a cart has changed requires trusted checkout or payment-provider data, plus an integration able to enforce the result.

Mandate
Trusted cart
Hash-bound gate

Partner-dependent — not live. ACI does not claim to inspect or block item-level substitutions across ordinary card rails today.

POLICY MODEL

Authority scoped per agent

The proposed policy model includes organisation defaults, department overrides and agent-specific mandates. Rules can include allowed merchants and categories, per-transaction limits, shared budget pools, approval thresholds, validity periods and freeze or revocation status.

EVIDENCE & AUDIT

Explain every decision

The intended record includes the mandate reference, agent, request, policy version, decision and rule outcomes. The current local HMAC approach is demonstration-grade; production key management, issuer integrations and liability terms are not implemented or established.

Watch the walkthrough

A short recorded walkthrough of the local prototype: an agent request, a deterministic ALLOW / DECLINE decision, the evidence record, and the honest Layer B response.

Watch the walkthrough

Simulation only · No live financial integration · Layer B remains partner-dependent and not live

What ACI is not

Not a bank. Not an issuer. Not universal enforcement across every payment route.

ACI is an early software concept and research prototype. It does not hold deposits, issue payment cards, connect to live financial rails or guarantee enforcement across unsupported providers. Layer A requires an issuer or payment-provider integration. Layer B additionally depends on trusted cart data and a partner integration.

Research conversations

Help test whether this problem is real.

We are looking to speak with people who have direct experience deploying or governing agents that can initiate purchases or payments. The aim is to understand existing controls and where they break down — not to sell a product.

Email to book a 15-minute call
Privacy information

Who is responsible. ACI is an early-stage research project. The data controller for personal data submitted through this site is the individual running the project, reachable at aci.research@proton.me.

What this page collects. This page is a static site with no contact form, no analytics scripts and no non-essential cookies. The hosting provider (Cloudflare Pages) may process standard technical data such as IP address, request time and browser information in server logs for security and operational purposes.

When you email the project. If you contact the project by email, your email address and the contents of your message will be used to respond and, where relevant, to arrange or conduct a research conversation. Please do not send confidential business information, customer data, payment details or credentials.

Retention. Research correspondence is kept for up to 12 months from last contact and then deleted.

Your rights. Under UK GDPR you have the right to request access to, correction of, or deletion of your personal data, to object to processing, and to lodge a complaint with the Information Commissioner's Office (ico.org.uk). To exercise any of these, email the contact address above.