Latest News & Resources

Enterprise AI agents are moving from answering questions to taking actions across CRM, finance, customer service, and operations. Every action requires an identity that a system can recognise and a set of permissions that it can enforce. Yet many organisations still run agents through shared service accounts or broad API keys designed for conventional software integrations. That approach blurs accountability. When an agent updates a record, launches a workflow, or retrieves a document, the business must know which agent acted, on whose behalf, under which policy, and with what approved purpose. Agent identity is therefore becoming a core architectural decision rather than a minor integration detail.

Why Shared Credentials Fail

A shared credential can appear efficient during a pilot because it removes repeated authentication steps. In production, it creates a large and poorly understood permission boundary. Several agents may inherit the same access even when their business responsibilities differ. Investigations become difficult because logs identify a technical account rather than the actual actor or request. The risk grows when a model is prompted to use a connected tool in a way its designer did not anticipate. A secure design treats every agent as a distinct workload and every delegated user request as a separate context. Access is granted for the task, not merely because a connection exists.

Design Around Least Privilege

The practical starting point is to map agents to business capabilities. A sales assistant might read opportunities and draft follow-up messages but should not approve discounts or export an entire customer database. A procurement agent might compare quotes but require a separate approval before placing an order. Define permissions by resource, action, scope, and duration. Short-lived credentials, scoped tokens, and policy checks at tool execution time reduce unnecessary exposure. This design also makes testing more meaningful. Teams can verify that expected tasks succeed while sensitive operations are rejected, producing evidence that security rules operate beyond the prompt layer.

Preserve Delegation Context

An agent often acts as an intermediary between an employee and an enterprise platform. The platform should receive enough context to distinguish a user-authorised request from an independent automated action. This may include the initiating user, requested objective, permitted workflow, and trace identifier. It should not mean forwarding unrestricted personal credentials to the model. A robust implementation separates the intelligence layer from the access broker that issues bounded permissions. Audit records then connect the employee instruction, model recommendation, tool call, approval event, and final system change. That chain is valuable for investigations, compliance reviews, and operational troubleshooting.

Build a Lifecycle for Agents

Agent identities should have owners, inventories, periodic access reviews, and retirement procedures. An agent that is no longer deployed should not retain valid credentials. When its function changes, permissions should be reassessed instead of quietly expanded. Enterprises also need controls for secrets rotation, unusual activity detection, and emergency suspension. Ownership belongs jointly to platform engineering, security, and the business process team. The design should make it easy to answer basic questions: which agents are active, what can each do, which applications can they reach, and when was access last approved?

Measure Security in Practice

A strong identity programme measures more than the count of registered agents. Useful indicators include the percentage of tools using scoped credentials, the number of excessive permissions removed, completion of access reviews, and the time required to revoke an agent. Scenario tests are particularly informative: can an agent access a neighbouring business unit, approve its own transaction, or reuse permissions after a workflow ends? Findings should feed back into policy and tool design. With explicit identity, enterprises can scale autonomous workflows without losing the accountability that conventional systems were built to provide.

Putting Identity into the Delivery Roadmap

Begin with a small register of agents and their connected tools. Select one workflow with meaningful but reversible actions, create a distinct workload identity, and remove unnecessary access. Add an approval boundary for privileged changes and run negative tests before deployment. Then connect identity and transaction traces to the operational dashboard. This pilot demonstrates how security design can improve debugging, not merely add restrictions. Extend the pattern using shared libraries and templates so every new agent does not require a custom security project.

Consider how agent identity affects collaboration between departments. An assistant may receive a legitimate request from sales but need information governed by finance. The identity architecture should preserve that distinction and deny access that would not be available to the initiating employee. Authorisation can be evaluated at the point of action using verified user context and the agent role. This is more dependable than copying permission claims into a natural-language prompt. Where workflows cross multiple systems, consistent trace identifiers allow investigators to reconstruct the complete path without granting any single tool unnecessary access.

Identity also affects the speed of delivery. Security teams can preapprove reusable access patterns for document search, draft creation, and other common capabilities. Developers then select the narrowest template and request exceptions with a stated business justification. Such a model reduces repeated negotiations and prevents permissions from drifting during successive feature releases. The result is a platform where responsible agent teams can move faster because accountability is built in from the beginning.

What Leaders Should Take Forward

The next phase of agentic AI will reward organisations that can prove what an agent was permitted to do at the moment of action. Clear identity, bounded delegation, and usable audit evidence make autonomy easier to govern and easier to extend across business systems.

Service alignment: Custom AI Models & Agents | Strategic Partnerships