Agent authority
Every agent acts within
the authority you delegated.
AI agents now approve loans, move cash and pay suppliers. Meilynx gives each one a written delegation of authority drawn from your bank's policies, enforces it on every action on every platform, and seals each decision as evidence.
A file for every agent, checked on every action.
Your bank keeps a file on every customer and every employee. Each AI agent needs one too: who it is, who it acts for, what it may do and when that ends.
- 1Which agentVerified on every action. An agent's description of itself never counts.
- 2On whose behalfThe customer, the employee or the business line that delegated the work.
- 3What it may doActions, systems, counterparties and amounts, and what happens above the line.
- 4Until whenAn expiry, a revocation, or a kill switch that stops the agent at once.
Delegation of authority
Verified agent
- 1Agent
- commercial-lending-assistant
- 2Acts for
- Commercial Credit · Midwest
- 3May approve
- term loans up to $250,000
- 3Needs first
- risk score · DSCR ≥ 1.25
- 3Above its limit
- held for a credit officer
- 4Authority ends
- 31 Mar 2027 or on revocation
Start from the policies your bank already wrote.
Meilynx turns each policy into rules linked to the clause they came from. Your team approves them, then sees exactly how much of the policy is enforced and where two departments disagree.
- 1Your policyPayments policy v14, credit authority matrix, treasury limits
- 2Rules, clause by clauseEach rule links back to the sentence it came from
- 3Your reviewRisk and compliance approve before anything takes effect
Coverage · payments policy v14
example55clauses
- 41Enforced on every action
- 9Sent to a person for judgment
- 3Departments disagree
- 2Needs a fact you don't send yet
Operations allows automatic payments up to $25,000; Risk wants two approvers above $5,000. Until someone settles it, the stricter rule applies.
Controls that hold whatever the agent's instructions say.
- Required facts
- An action that depends on a sanctions result, a risk score or a beneficiary check waits until that fact is present and current.
- Velocity limits
- Caps on amounts, counts and frequency per agent, per customer and per period, counted across parallel requests.
- Approval gates
- Above a threshold, a named person with the authority to decide releases the action. Requesters never approve their own.
- Kill switch
- Suspend an agent, revoke its access and stop further actions at once. The stop is recorded like any other decision.
- Untrusted content
- Instructions hidden in a document, an email or a transaction description cannot start a transfer or an export.
- Judgment, with a person behind it
- Clauses that need judgment can go to AI review or to a person. Review can only make a decision stricter.
Six places agents already act for banks.
Lending
≤ $250,000 per loanApprovals within credit authority
Bigger or exception cases go to a credit officer with the risk model's result attached.
Payments
per mandatePayments within the customer's mandate
Each payment is checked against what the customer authorized before it executes.
Agentic payments →Treasury
≤ 4 moves per daySweeps between accounts
Agents that move cash stay inside the customer's settings and your velocity limits.
Research and advice
one recordFrom conclusion to trade
The research view, its sources and the order it led to can be reviewed together.
Servicing
≤ $75 per caseRefunds and fee waivers
Credits go out only for the reasons and amounts your policy delegates.
Fintech partners
your rulesYour conditions on their agents
Apply your limits and approvals to agents partners run for your customers, and check that they held.
The same rules on every platform your agents run on.
Agents built in-house, agents on your data and cloud platforms, and agents your partners run for your customers all answer to one policy and land in one record.
- In-house agents
- Databricks
- Amazon Bedrock AgentCore
- Microsoft Copilot Studio
- Azure AI Foundry
- Fintech partners' agents
"An agent should operate only in pursuit of an approved business intent, within a defined delegation of authority."
OCC Bulletin 2026-54 lists delegation, velocity limits, monitoring and revocation among the controls examiners weigh for agents that initiate transactions.
The HKMA's guidance on responsible AI expects a kill switch for higher-risk uses and controls that go beyond the model's instructions.
Map one agent's authority with us
Bring one agent and the policy it should follow. We'll show its delegation, the coverage of your policy and a held action from start to finish.