If an agent can borrow a human identity, the control plane must show exactly whose authority it is spending.
Identity boundaries define which machine workers may act under which operational identity, what they may do there, and when the borrowed authority expires.
Agents can inherit permissions from tools, teammates, aliases, and integrations faster than managers can see which identity is being used for which promise.
Heads of Growth get machine-worker speed without confusing customers, reps, managers, or auditors about who actually took responsibility.
The fastest way to make agents useful is to let them use the systems people already use. That is also the fastest way to erase accountability. A rep account can send a note. A manager login can approve a discount. A support alias can promise a fix. A generic automation user can rewrite half the operating record.
The issue is not whether the task was correct. The issue is whether the company can reconstruct the authority chain after the task changes customer reality.
If an agent sends as a human, the message must show delegated scope, source evidence, and whether the person reviewed the action before it went out.
Messages from sales@, success@, or ops@ need a machine-worker receipt so the team can see which worker acted and which policy it followed.
Shared credentials turn accountability into archaeology. Replace them with named agent identities, scoped tokens, and visible delegation records.
The generic automation account should not become a permission landfill. Every workflow needs a named sponsor and a bounded action set.
Human orgs run on delegated authority: who may speak for sales, who may approve spend, who may promise delivery, who may edit revenue truth. Agentic orgs need the same thing, but machine-readable. The worker's identity should answer five questions before it acts: who is the machine worker, whose authority is it using, what can it change, how long is that authority valid, and what evidence will be left behind?
That turns identity from a login problem into an operating contract.
Name the actual machine worker and workflow, not just the tool or integration that executed the action.
Name the human, team, policy, or control-plane rule that delegated authority for this class of work.
Define allowed systems, audiences, fields, claims, spend limits, approval classes, and customer contexts.
Set when the delegation ends: time window, campaign end, account stage, policy version, confidence drop, or exception count.
Attach a visible record: source evidence, prompt or rule version, approval state, changed object, rollback path, and reviewer.
A prospect asks a pricing question. The agent finds an approved answer and sends from the account executive's email. Without an identity boundary, the buyer believes the AE personally committed. With one, the note is either visibly drafted for review or sent under a delegated assistant identity with the exact policy and source attached.
A growth worker grants a nurture-flow exception for a strategic account because a manager approved similar cases last week. That is not authority. The boundary forces a fresh sponsor, current scope, and expiry before the exception changes the customer path.
Count every action where a machine worker used a human name, team alias, shared integration, or inherited permission.
Measure actions where the worker completed the task but the authority source cannot be named in one sentence.
Measure how long it takes to identify, explain, and reverse an autonomous action after a customer or operator disputes it.
Most teams notice identity too late: when a customer asks who promised something, a rep disagrees with a sent message, finance challenges an approval, or security finds a shared token. By then the company is not debating automation. It is debating trust.
Good machine-worker management makes the boundary visible before the work scales. Agents can still move fast. They just cannot borrow authority invisibly.
List every human account, alias, integration user, and automation account that agents can use today.
Separate draft, send, approve, edit, delete, spend, promise, and escalate permissions.
Require every borrowed identity to name a sponsor, scope, expiry, and receipt before execution.
Pick three recent autonomous actions and prove you can explain who acted, under whose authority, and how to reverse it.
Do not let autonomous workers act through human or shared identities without a visible delegation contract.