Most organisations treat identity as a front door: someone signs in, the system recognises them, and work begins. That model is incomplete when people, AI agents and automated services can all change business records. Authentication answers who entered; operating identity must also decide what that actor may do at the moment of each action.
The distinction matters to buyers because an apparently successful automation can still be wrong: it may use the wrong organisation, update the wrong customer, rely on stale permissions or leave no evidence of who authorised it. A convenient login does not fix those failures.
Three questions identity must answer on every action
- Who or what is acting? A named person, agent or service with an accountable identity—not merely a shared credential that happens to be valid.
- In which context? The intended organisation, application, workspace and specific record. The same actor can have different access in two organisations or two workspaces.
- With what authority? The current permission to perform this particular read or write, including applicable resource limits and any required approval.
A sign-in system can answer the first question. Reliable business operations need all three, and they need a way to check the result after the action runs.
Why single sign-on is necessary but not sufficient
Single sign-on reduces password sprawl and makes human access easier to manage. It does not, by itself, determine whether a sales agent may edit a particular opportunity, whether a contractor may access a client's project, or whether an automated workflow may send an external email.
The important controls sit behind the login: selecting the correct organisation, evaluating live application and workspace permissions, limiting access to named resources, and recording what actually changed. A claim inside a token or a successful API response should not substitute for a current authorisation decision.
Agents make the missing controls impossible to ignore
Consider an agent asked to prepare a follow-up for a genuine demo request. Reading the request, drafting a suggested reply, creating an internal review item and sending a message are four different effects. An agent may be authorised for the first three but not the fourth. The business should be able to distinguish them before execution, not after a customer receives an unexpected email.
- Scope: Confirm that the agent is acting in the correct organisation and workspace, on the intended source record.
- Permission: Inspect whether the exact action is allowed, rather than assuming that access to an application permits every operation.
- Effects: Distinguish an internal draft from a delivered message, a suggested meeting from a confirmed booking, and a configured integration from a verified handoff.
- Approval: Where a consequential effect requires human approval, enforce it at the receiving service rather than trusting a prompt to provide authorisation.
- Evidence: Read the resulting record independently; a successful request alone is not proof of a completed business outcome.
This is the difference between an agent that can operate software and an agent whose work an organisation can safely rely upon.
A practical five-step identity and automation test
You can test these principles in your current software stack without running a marketing campaign or contacting a customer. Use an authorised internal record and an appropriate test operator.
- 1. Set a real boundary. Choose one existing non-sensitive internal record, its organisation, workspace, owner and permitted actions. Document the starting state.
- 2. Inspect access before acting. Ask what this operator may read, edit and send. Confirm the decision comes from the target system's current permissions, not only a model instruction.
- 3. Perform one permitted internal update. Use a supported native action, validate its input and record the returned resource identifier.
- 4. Attempt a safe denied action. Use a read-only permission check or non-mutating plan for an out-of-scope resource. Do not try to circumvent a denial or trigger an unwanted customer effect.
- 5. Reconcile independently. Read the permitted record back, verify the exact field change and check that the disallowed path did not mutate data. Record the identity, resource, timestamp and approving actor where applicable.
A useful pass condition is not 'the agent completed five tool calls'. It is one correct authorised change, one correctly enforced boundary, and independent evidence of both. Record any missing permissions, unclear owners or manual exceptions as work still to be fixed.
Five questions to ask a platform vendor
- Can a human reviewer discover the exact applications, actions, resource scope and expected effects available to an agent?
- If an operator is removed from a workspace, how is current access re-evaluated across the web UI, API, CLI and agent surface?
- Can the product separate permission to draft from permission to publish, send, book or charge?
- Can it show an immutable or attributable record of the resulting change—not merely a chat transcript saying the task is done?
- When an action is denied or the network fails after a write, how can an operator reconcile the outcome and recover safely?
How Topolo approaches the problem
TopoloOne is designed around shared identity and organisation context across focused business applications. Its published agent-facing action contracts describe available operations, input schemas, permissions, effects and verification steps. A credential-scoped agent discovers and validates an action, acts through the owning application, and reads back the receiving record. The receiving application's actual enforcement—not the agent's words—is authoritative.
The promise is an operating model, not a claim that every potential cross-app integration or customer journey has already been delivered. Evaluate a specific workflow and require its real outcome evidence before deciding whether the platform meets your needs.
Identity is not simply how work starts. It is how the organisation keeps every subsequent action attributable, appropriately scoped and verifiable.