Authority for AI agents
Access is not authority.
Trust in an agent does not become unlimited authority.
AI agents already pay invoices, change records and prepare decisions, with the credentials they were given. Adaptive Runtime sits between agents and the tools they use. Before each action, it decides whether the action goes ahead, waits for a person or is refused, and keeps evidence of that.
- Pay invoice Papelaria Central · €250AllowedWithin the limits: goes through without intervention.
- Pay invoice ACME Fornecimentos · €4,800HeldAbove the limit: a person sees the request and decides.
- Delete customer C-1042RefusedAction not authorised for this agent.
Three requests from the same agent, decided before they reach the system. Each decision is recorded.
The problem
An agent with access to a system can do everything that access allows.
The agent holds the keys of whoever set it up.
Credentials open the whole system. They do not say up to what amount, to whom, or on whose behalf.
A document can give it orders.
An invoice, an email or a ticket can carry hidden instructions. The agent reads them as if they came from its user.
In the end, no one knows who authorised it.
Logs show what ran. They do not show who authorised it, under which rules, or whether what ran is what was approved.
What it does
The organisation's systems do not change. Authority moves to a place where it can be seen, decided and proven.
Identifies
Which agent is asking, and on behalf of which person, verified with the organisation's identity provider.
Decides
By the organisation's rules and limits, and by what its systems know at that moment.
Calls a person
When the action commits the organisation. The approver sees what will happen, and the requester does not approve.
Proves
Each decision is recorded in a way that makes any alteration detectable. For an approved action, the record confirms that what was approved was actually executed.
Examples
Seven real-world scenarios where an AI agent’s access must not amount to permission to act.
A payment beyond authorised limits
An accounts payable agent receives a €25,000 invoice and prepares the payment in the ERP system.
Read the example →CUSTOMER SERVICE AND CLAIMSA fraudulent refund or compromised account
An agent handles customer requests by email and chat through the CRM, including refunds and claims settlements.
Read the example →HUMAN RESOURCESSalary diverted through a bank-account change
An HR assistant receives an email apparently sent by an employee, asking for their salary to be paid into an overseas bank account.
Read the example →IT OPERATIONSDangerous changes to live systems
An agent investigates an incident, reviews logs, prepares a fix in GitHub and interacts with Kubernetes.
Read the example →BANKING AND INSURANCEA credit or insurance decision without accountability
An agent reviews credit applications or insurance claims, including supporting documents, and prepares a proposed decision.
Read the example →PUBLIC ADMINISTRATIONAn application decided without proper authority
An agent supports the processing of an application, from reviewing the submission to preparing the notification.
Read the example →LAW FIRMS AND LEGAL DEPARTMENTSA legal filing submitted without review
An agent reviews correspondence, drafts pleadings and documents, and can access deadline management and document-submission tools.
Read the example →Evidence for auditors and supervisors
Every decision is recorded: who requested it, on whose behalf, who approved it, and under which rules. Any alteration to the record is detectable, and each decision can be provided to an auditor together with its supporting evidence. This helps demonstrate the human oversight required by the EU AI Act (Article 14) and the controls and record-keeping required under DORA, the GDPR and NIS2 (Article 21).