An AI agent should not receive enterprise power merely because it can authenticate. A credential can establish which workload is calling. It does not prove that the workload may approve a refund, change a limit, deploy code or release customer data. My thesis is that useful agentic AI needs three separate proofs: identity, mandate and evidence.
This distinction is moving into standards work. NIST launched its AI Agent Standards Initiative in February 2026 around secure interoperability, open protocols, agent security and identity. The initiative matters because agents create value only when they can cross system boundaries—and every boundary introduces another party that must decide whether to trust the action.[1]
The risk is no longer theoretical. Reuters reported on 13 November that a US task force discussed AI-powered cyberattacks and incident reporting after a series of agentic security failures. The exact policy response is still developing, but the operating lesson is already clear: autonomy without a verifiable authority boundary converts a model error into an organizational action.[8]
Identity answers only the first question
NIST’s October design tenets call for agents to have distinct, cryptographically verifiable identities tied where possible to a workload, environment and approved configuration. That is a stronger foundation than allowing every agent to borrow a human account or a generic service credential. It improves traceability and makes revocation possible without disabling the person who delegated the work.[5]
Identity infrastructure must also survive ordinary security failures. NIST IR 8587, finalized in September, focuses on protecting tokens and assertions from forgery, theft and misuse through stronger key management, verification, lifecycle controls and continuous monitoring. These are not agent-specific luxuries. They are the plumbing beneath any trustworthy authorization decision.[3]
An agent needs a mandate, not a blank permission
A useful mandate should bind seven things: the human or organization delegating authority; the exact agent and approved configuration; the business purpose; the permitted data and tools; quantitative limits; the time window; and the conditions that force human review or immediate revocation. The mandate must travel with the action or be verifiable by the receiving system.
Consider a consumer agent asking a bank to make a payment. The bank does not operate the agent and cannot inspect its complete reasoning. It needs a signed, current statement of authority: which account, which beneficiary classes, what amount, what frequency, which authentication step and what happens when the request falls outside the mandate. A successful login is not enough.
Prompt injection becomes an authority problem
Saudi Arabia’s 2026 AI Cybersecurity Guidelines explicitly address agentic AI. The published guidance says data retrieved from external tools, memory systems or third-party APIs should be treated as untrusted and validated before entering the model context. This is important because malicious instructions can arrive disguised as ordinary data.[6]
The same guidance calls for impact-based risk assessment, formal approval before high-impact deployment and graduated autonomy with continuing human oversight. In practice, the safest response to prompt injection is not only a better filter. It is an authorization layer that prevents an untrusted instruction from expanding the agent’s mandate.[6]
Saudi implementation should classify actions, not only models
The NCA consultation organizes AI cybersecurity around governance, defense, resilience and third-party cybersecurity. That framing should move into the product backlog. Every agent action should be classified by potential impact, reversibility, data sensitivity and whether another organization must rely on it.[6]
I would use a simple enterprise ladder. Reading approved information may be autonomous. Drafting a recommendation should retain sources and uncertainty. Updating a reversible internal record needs a bounded mandate and audit event. Committing code, releasing data, approving money or changing customer rights requires stronger authentication, separation of duties and, in many cases, a human decision.
The GCC opportunity is a trust infrastructure opportunity
PwC’s recent GCC findings show the scale of the gap. Security and cybersecurity were the most frequently cited capability gaps limiting agent scale-up, at 66%. Compliance and legacy integration were each cited by 47%, while only 20% of respondents said they would allow agents full execution autonomy. These are survey findings, not forecasts for every enterprise, but they point to a market that wants automation without surrendering accountability.[7]
Saudi Arabia, the UAE and other Gulf markets should not be treated as one authorization domain. A reusable agent platform can share identity protocols and policy engines, but regulated actions, data-location requirements, approval roles and evidentiary expectations differ by jurisdiction and sector. Expansion should package local policy as configuration, not bury it in prompts.
MENA will not mature at one speed
Across the wider MENA region, identity infrastructure, cloud adoption, payment rails and enforcement capacity vary. A design that assumes every organization can issue hardware-backed credentials or maintain continuous telemetry may exclude otherwise valuable deployments. The right architecture should support progressive assurance: stronger controls for higher-impact actions while preserving a safe, reviewable mode for lower-maturity environments.
The counterargument: IAM already exists
A reasonable counterargument is that enterprises already have OAuth, service accounts, role-based access control, privileged-access management and audit logs. Creating a new security layer for agents could add complexity and another vendor category without solving the underlying discipline problem.
I agree with the first half. Existing IAM should be reused, not replaced. The missing layer is the connection between a dynamic task and a durable authorization decision. A role may allow an application to call an API; a mandate should prove why this agent may perform this action now, within this limit, on behalf of this principal. NIST’s identity project moved from concept work to a DevSecOps implementation use case after feedback from more than 600 commenters, which suggests the gap is operational rather than purely conceptual.[4][2]
The founder opportunity is bounded execution
The product opportunity is not another chat interface. It is a control plane that issues workload identity, compiles business policy into machine-enforceable limits, obtains step-up approval, protects credentials, records the mandate chain, evaluates each action and retains evidence for reconstruction. The strongest product will work across models so the customer does not lose governance when the model changes.
A commercial pilot should measure unauthorized actions blocked, legitimate actions wrongly rejected, time to approval, time to revoke, percentage of actions linked to a valid mandate, evidence completeness and time to reconstruct an incident. Model benchmark scores do not answer these operating questions.
Six decisions for boards and founders
First, name the human or legal principal behind every agent. Second, give the agent a distinct identity and lifecycle. Third, define action-level limits and expiry. Fourth, require stronger approval as impact and irreversibility rise. Fifth, make revocation and safe fallback faster than execution. Sixth, retain enough evidence for a regulator, customer or incident responder to understand what happened without reading the model’s hidden reasoning.
The next generation of enterprise AI will not be trusted because it sounds intelligent. It will be trusted when every consequential action can answer three questions: who acted, under whose authority, and where is the evidence? Identity opens the door. A mandate defines the room. Evidence proves the agent stayed inside it.
Key takeaways
- Separate agent identity from the authority to perform a specific action.
- Bind every consequential action to a scoped, expiring and revocable mandate.
- Treat external tool and memory content as untrusted data, not as new authority.
- Measure blocked unauthorized actions, false rejections, revocation speed and evidence completeness.