Your agent should not inherit all of your authority
New endpoint telemetry and personal enterprise agents expose a hidden delegation problem: one employee request can become many machine actions. Preserve the person’s discretion while giving the agent a narrower, explicit mandate.
AI-generated voice. Short companion, not a full reading.
Read the narration transcript
An employee may have permission to do something that their AI agent should not be allowed to do.
The twenty-second of October Workforce edition examines a hidden delegation problem. One human request can expand into many machine actions, even when the employee intended a much narrower task.
BeyondTrust reports early-customer coding-agent telemetry in which roughly one hundred and thirty-two thousand prompts produced about one point two eight million shell-command requests. Requested commands were not necessarily executed successfully. The sample does not establish how common harmful actions are across workplaces.
ServiceNow and NVIDIA describe a persistent personal enterprise agent with memory, skills and runtime policy controls. Kognitos describes finance-role agents with approved procedures, limits and hand-offs. These are supplier announcements or previews, with no independent outcome comparison in the material discussed.
Charter's survey provides a different perspective: people reported greater benefits when they had a say in how AI fitted into their work. Those are self-reported associations, not proof that agency caused the benefits.
Consider the article's fictional manager asking, “Help me move Sam into the new team.” The manager may explore the move and prepare a case. That does not authorise software to infer Sam's consent, change an employment record or send an unapproved message.
The article proposes an authority fork: preserve the person's discretion while giving the agent a narrower, explicit mandate. Define the purpose, permitted data and tools, duration, stop conditions and evidence returned. Keep the agent's actions distinguishable from the employee's own activity.
This is LecXie's analysis, not an industry standard. Existing identity and workflow controls may already achieve the same result. The approach earns its place only if it reduces unauthorised actions and repair without increasing delay, abandonment or hidden workarounds.
The practical action is to split one permission. Ask the process and identity owners to compare what an employee may do with what the agent actually needs. Specify one narrower task grant, expiry or human checkpoint, then rerun the same fictional case while preserving the employee's discretion.
This is a condensed companion to the published article. Synthetic voice generated with OpenAI. Editorial direction by Dominic Chiappe, research and synthesis with OpenAI.

An employee may be allowed to do something that their AI agent should not be allowed to do.
That distinction sounds obvious until the agent works through the employee's identity, applications and accumulated context. One instruction can become a sequence of searches, messages and system actions. The person may understand the unstated boundary; the machine may only see an available path.
New evidence makes this more than a theoretical access-control problem. BeyondTrust has published endpoint telemetry showing that one prompt to a coding agent can fan out into many requested commands. ServiceNow has announced a long-running personal enterprise agent that builds memory and skills while enterprise policy follows it into the runtime. Kognitos is packaging agents as defined finance roles with approved procedures and hand-offs. Separately, Charter's new survey suggests that people benefit most when they retain agency over how AI fits into their work.
Together they point to a design rule for the mixed workforce: preserve the employee's discretion while giving the agent a narrower, explicit mandate. Do not make the person and the machine one security subject.
One request is not one action
BeyondTrust reported on 21 October that its Phantom Labs team analysed early-customer endpoint telemetry from coding-agent activity. It observed about 1.28 million shell-command requests arising from approximately 132,000 prompts over four weeks, or nearly ten requested commands per prompt on average. Among observed users with coding agents configured, 85% had at least one Model Context Protocol server connection, averaging 4.35 servers and reaching 36 for one user.
The company is selling an agent-security product, so its framing has a commercial interest. The sample is not a population estimate for all employees or agents. Requested commands were not necessarily executed successfully, and the release does not provide outcome or harm rates. The useful finding is narrower: the unit a person experiences as one request is not the unit the enterprise must govern. BeyondTrust, 21 October.
An instruction such as “prepare the release” can cause an agent to inspect repositories, open files, invoke tools, write code, run tests and contact services. The employee's intention is compact; its execution surface is not. A control attached only to the initial prompt can therefore be both technically correct and organisationally incomplete.
BeyondTrust says its new product distinguishes agent actions from human actions and can restrict the former without removing the employee's own permission. That is a supplier claim, not independent evidence of effectiveness. It nevertheless exposes the right boundary: access held by a person need not become standing authority for software acting around that person.
The personal agent is becoming an enterprise actor
ServiceNow's 21 October announcement describes Arc as a long-running personal agent that can operate across workflows, email, Teams and other productivity tools. It can run on a dedicated computer or an employee's desktop, build memory and develop skills over time. ServiceNow says enterprise policy is defined in its control plane and enforced at runtime through NVIDIA OpenShell; an AI Passport carries restrictions such as risk tolerance and permitted external services.
These are announced capabilities, not an independently evaluated deployment. Availability and field performance remain to be established. The organisational shift is still important: a personal agent is no longer merely a transient answer generator. It can become a persistent actor whose memory, tools and habits grow beside one employee while IT attempts to make institutional rules travel with it. ServiceNow and NVIDIA, 21 October.
Kognitos approaches the same problem from the other direction. Its new preview offers 61 finance and banking “AI Employee” roles backed by 515 standard operating procedures. Each role is described as having a job description, approved procedures, limits on action without human approval and defined hand-offs. A context graph stores team teaching with who supplied it and why. Again, these are vendor descriptions; the release supplies no independent comparison of reliability, productivity or audit outcomes. Kognitos, 21 October.
The labels differ, but the design problem converges. A useful agent needs continuity and context. An accountable organisation needs to know which role the agent occupies, which procedure applies, which person taught it, which action it may take and where authority returns to a human.
Human agency and agent authority are different things
Charter's AI Pulse, published on 21 October, draws on a survey of 5,002 US desk workers, managers and executives who use AI, plus 91 interviews. About two-thirds said AI improved their work at least somewhat, but fewer than half said it improved their lives and about one in ten reported that it made both work and life much better.
Charter says the strongest predictor of benefit was agency: having a say in how AI fits into the job and how it is used. More than a third said they had hidden AI use from a manager, colleague or client. The survey also reports a gap between managers' accounts of what they demonstrate and what employees observe. These are self-reported associations in a US desk-worker sample; they do not show that agency causes better outcomes or that hidden use causes burnout. Charter, 21 October.
The evidence nevertheless challenges a blunt response to agent risk. If governance is experienced as removing an employee's discretion, people may conceal use or work around the approved path. If governance merely lets software inherit the employee's access, the agent may act too widely. The institution needs to expand human agency and constrain machine authority at the same time.
That is not a contradiction. The person can choose whether, when and how to delegate. The agent can receive a task-specific grant, operate within it and return evidence. The person remains able to investigate, challenge and choose another route even when the agent cannot.
A worked case: “Help me move Sam into the new team”
Consider a fictional manager asking a personal workplace agent, “Help me move Sam into the new team next month.” This is an organisational design example, not a reported product deployment or employment-law advice.
The manager may legitimately see the team plan, discuss options with Sam, consult budget information, request HR support and initiate an internal-move process. Those permissions do not mean an agent should silently perform every available step.
| Stage | Human discretion | Agent mandate |
|---|---|---|
| Clarify | Explore whether Sam wants the move and whether the receiving manager agrees. | Ask for missing intent; do not infer consent from calendars or messages. |
| Prepare | Review the role, timing, development effect and alternatives. | Assemble authorised facts and draft options with provenance. |
| Propose | Choose whether to recommend the move and explain the judgement. | Present consequences and uncertainty; do not manufacture agreement. |
| Commit | Use the organisation's named employment and budget authorities. | Open or update the workflow only under an explicit task grant. |
| Communicate | Own the conversation and repair any misunderstanding. | Send only approved messages and preserve what was authorised. |
A borrowed-identity design is tempting because it removes friction. The agent can already reach email, HR systems and workflow tools, so it appears efficient to let it act as the manager. But the manager's access was granted to a person expected to exercise judgement in context. It was not automatically granted to every sequence of actions software can derive from the manager's words.
The better design does not require approval for each keystroke. It separates capability from mandate. The agent may retrieve agreed records, prepare the case and create a reversible draft. A consequential transaction receives a bounded grant tied to this purpose, this person, these systems and this period. The agent's audit trail remains distinguishable from the manager's own activity.
The authority fork
ORIGINAL SYNTHESIS. When a person delegates to an agent, authority should split rather than flow unchanged.
The human path preserves broad discretion: interpreting ambiguous circumstances, changing the objective, considering information that should not be automated and accepting responsibility. The agent path receives a narrower executable mandate: defined purpose, allowed data, tools, limits, duration, evidence and stop conditions.
Call this an authority fork. It is not a new identity product or legal standard. It is a way to avoid two common errors:
- Authority laundering: the agent performs a consequential action because the employee technically could, even though nobody explicitly delegated that action to software.
- Discretion collapse: the organisation restricts the employee to the agent's safe path, removing the ability to recognise an exception or choose a better route.
The fork should be visible in identity, logs and operating design. A record needs to show who asked, what the agent was allowed to do, which actions the agent requested, which actually executed, where a person intervened and who owns remedy.
What if we are right?
Enterprises could give agents useful scope without treating employee access as an all-or-nothing inheritance. People could delegate preparation and reversible execution while keeping the freedom to reframe the work. Security teams could govern the machine's action surface without turning every employee into a less capable version of the agent.
The constructive possibility is a simpler working relationship. The employee states an objective once. The agent receives an intelligible, temporary mandate, completes the safe work and returns evidence. The organisation spends less time on blanket approvals because the permission is attached to the delegated task rather than to a vague category such as “trusted agent”.
What would prove us wrong?
Modern identity and workflow systems may already distinguish human and machine actions adequately. Agents may remain so tightly supervised that separate mandates add little value. Permission mirroring could also be safer and simpler in low-consequence work, especially when every agent action is reversible and immediately visible.
The strongest competing explanation is that this is familiar least-privilege security applied to a new software client. Much of it is. The Workforce claim earns its place only if the human side matters too: delegated work fails when controls either let the agent borrow too much authority or remove the employee discretion needed to handle exceptions.
Test it comparatively. Does a separate agent mandate reduce unauthorised actions, unexplained tool calls and repair effort without increasing task abandonment, delay or workarounds? Can employees still change the objective and resolve exceptions? If existing controls achieve those outcomes, the authority fork adds terminology rather than value.
Revisited: inherited digital capability needs an inherited mandate
On 30 September, Workforce argued that an employee-built agent can move from personal accelerator to role capability, process automation and enterprise capability. It proposed inherited digital capabilities that survive a role change. The new announcements make that prospect more concrete: personal agents can accumulate memory, role agents can carry approved procedures, and policy can travel into the runtime.
The update is that inherited capability also needs inherited authority boundaries. Reassigning an agent should not mean transferring the creator's access, habits and private context wholesale. The new incumbent, role owner and process owner need to know what the agent may do independently and which mandate must be renewed.
The 20 October edition separated knowledge, conduct and authority. Today's evidence adds execution granularity: even a valid human request can expand into several machine actions, each of which needs its own defensible scope. The earlier judgement is strengthened, not overturned.
A plausible risk: control inverts
HYPOTHESIS. Organisations may make agents easy to govern by making employees easier to constrain.
The causal path is plausible: an agent needs predictable permissions; the enterprise standardises the work around those permissions; exceptions become expensive; managers are encouraged to use the approved agent path; employee discretion narrows; hidden workarounds grow because reality still exceeds the model.
Charter's association between employee agency and reported benefit does not prove this sequence. Nor does hidden use establish bad governance. Watch for a different pattern: employees retaining broad system access but lacking a legitimate way to deviate from the agent's recommended path, explain an exception or transfer learning back into the procedure.
This risk weakens if task-scoped agent controls coexist with high employee voice, visible override routes and frequent improvement of procedures from frontline exceptions.
Operating-model implication
Give the existing identity owner, process owner and workforce owner one shared delegation contract. Identity defines who or what acted. The process owner defines the objective, permissible actions and remedy. The workforce owner protects the employee's discretion, challenge rights and development.
| Contract element | Question | Evidence |
|---|---|---|
| Purpose | What outcome did the person delegate? | Request, clarification and expiry. |
| Reach | Which data, tools and external services may the agent use? | Task-scoped grants and connection inventory. |
| Consequence | Which actions are reversible, and which require renewed authority? | Action class, approver and rollback or remedy. |
| Attribution | Can human and agent activity be distinguished? | Separate identity, requested actions and execution trace. |
| Learning | How does an exception improve the role or procedure? | Human intervention, reason and approved update. |
In the Workforce Context Architecture, catalogue the agent against the employee or role it supports, but do not collapse them into one node. Link the delegation, purpose, policy, systems and decision rights that connect them. A new role incumbent can then inherit the capability without inheriting an invisible personal mandate.
Human control watch
Assistance: the agent drafts or retrieves while the employee acts directly. The employee sees the evidence and owns the decision. Delegated execution: the agent receives a time- and task-bounded mandate, with separate attribution and defined escalation. Autonomous service: standing authority is attached to a role or process, not borrowed from a nearby human account; monitoring, scope reduction and remedy remain available to named people.
Capability-model update
| Gaining value | Under pressure |
|---|---|
| Human–agent delegation design | One shared identity treated as convenience |
| Task-scoped authority engineering | Broad standing access inherited by software |
| Agent-action attribution | Prompt logs treated as complete execution evidence |
| Procedure learning from exceptions | Employee deviation treated only as non-compliance |
| Digital-capability handover | Personal memory and credentials transferred wholesale |
1 · ONE THING
IF I WERE TO DO ONE THING NOW
Split one permission
PERSON → REQUEST → AGENT MANDATE → EVIDENCE
Return to the fictional promise test from the previous exercise, or choose one live employee-service or manager workflow. This week, ask the process owner and identity owner to compare one permission the employee holds with the authority the agent actually needs. Create or specify one narrower task grant, expiry or human checkpoint, then run the same fictional case again. Keep the employee's own discretion intact. Do not redesign the identity architecture or add a general approval layer. Make one borrowed permission explicit and test whether the service still completes.
Mental-model update
Old: the agent works as the employee, so it receives the employee's access.
Better: the employee delegates an objective, and the agent receives only the executable authority required for that objective.
The human and agent remain part of one working relationship without becoming one identity. Human discretion stays broad enough to recognise reality; machine authority stays narrow enough to be defensible.
Questions for the executive table
- Which personal agents are acting through employee permissions that were never explicitly delegated to software?
- Can we distinguish the employee's action, the agent's requested action and the action that actually executed?
- Where have agent guardrails unintentionally narrowed legitimate employee discretion?
- When a role inherits an agent, which permissions, memories and procedures must be renewed rather than copied?
- Can a worker challenge the approved agent path without having to conceal how they complete the work?
Evidence note. BeyondTrust reports early-customer endpoint telemetry and sells the control product discussed; requested commands were not necessarily executed and the sample is not a prevalence estimate. ServiceNow, NVIDIA and Kognitos describe announced or preview capabilities without independent outcome evidence. Charter reports self-assessed associations in a US desk-worker sample; it does not establish causation. The authority fork, authority laundering, discretion collapse and control-inversion hypothesis are original AyEye analysis, not claims made by the cited sources. This Workforce edition addresses organisational delegation and does not commission or duplicate Today's separately owned independent-EA feature.
