If operators cannot name what a worker is allowed to do, they do not manage labor. They manage folklore.
A capability inventory says what the machine worker can read, decide, draft, send, change, escalate, delete, spend, and recommend.
Agent teams improve through patches, exceptions, integrations, and successful shortcuts. Without inventory, the company loses the map of its own labor.
Operators can add machine capacity without discovering authority only after a customer, forecast, campaign, or CRM record has already changed.
Human labor is legible because companies obsess over titles, managers, teams, territories, and budgets. Machine labor does not arrive that way. It arrives as a workflow, a script, a bot, a prompt, a queue, a background worker, or a helpful integration that slowly earns more trust.
That is how mystery labor appears. The renewal assistant starts with draft summaries, then gets permission to update opportunity notes. The pipeline worker begins by flagging stale deals, then starts changing next steps. The outbound agent begins with research, then queues buyer messages. Each change is reasonable. The total capability set becomes unknown.
A capability inventory is the operating document that stops autonomy from becoming rumor.
Which systems, records, accounts, conversations, files, product events, tickets, contracts, and analytics sources the worker can inspect.
Which classifications, scores, recommendations, routing choices, priority changes, and segment assignments the worker can make.
Which fields, messages, tickets, tasks, notes, sequences, forecasts, handoffs, dashboards, and customer artifacts it can create or modify.
Which actions become real without another gate: send, enroll, update, close, discount, escalate, suppress, archive, or trigger another worker.
Teams worry about spectacular agent mistakes. The more common failure is a boring permission that becomes powerful in volume.
A worker that can edit CRM next steps can reshape seller attention. A worker that can suppress leads can erase pipeline before a human sees it. A worker that can summarize customer health can steer renewal posture. A worker that can move tasks between queues can decide whose day gets consumed.
None of those capabilities sound dramatic in isolation. Together, they are management power. If they are not inventoried, they are unmanaged power.
Every integration, prompt improvement, or tool grant becomes part of production because the worker is already there and the immediate task feels useful.
New authority is added to the inventory only with owner, scope, control, evidence, expiration, and rollback path attached.
Name the worker, owning team, accountable human, business process, and operating window. Anonymous automation is not labor management.
List every read, decide, write, commit, spend, escalate, delete, and notify action. Verbs beat descriptions because verbs expose authority.
Record where the worker acts: CRM, email, support, billing, data warehouse, enrichment vendor, ad platform, calendar, Slack, task runner, or customer portal.
Bind each capability to a guardrail: approval, sampling, threshold, evidence gate, redaction gate, queue SLO, rollback receipt, or blast-radius budget.
Every new tool, prompt change, permission grant, exception, owner change, and deprecation must leave a timestamped receipt.
The right question is not “is the agent good?” That is too vague to manage. Ask which capability deserves promotion.
A research worker may be excellent at account enrichment and still unready to queue outreach. A support summarizer may be safe inside internal tickets and unsafe in customer-visible replies. A forecast agent may classify risk well but should not change close dates. Capability-level review lets operators promote the useful part without promoting the whole machine.
This is where AI-native org design gets real: workers do not get trust as a mood. They get scoped authority.
“Has Salesforce access” is not an inventory. The inventory must say which objects, fields, records, and writes are permitted.
If nobody is accountable for a capability, the worker has authority without management. That is not autonomy; it is drift.
If a prompt update, integration, exception, or new queue can expand authority without a receipt, the inventory is decorative.
Choose the agent that touches CRM, customer communication, forecast, lead routing, renewal posture, support escalation, or pricing context.
Write down what it can read, decide, draft, send, update, delete, escalate, suppress, spend, and trigger downstream.
Attach a guardrail to every meaningful capability. No control means the capability is not approved for production.
Compare today’s worker to the original design. Remove stale permissions, expire exceptions, and mark promotion candidates explicitly.
Create a capability inventory for one autonomous worker that already touches a revenue system.
The blunt rule: if the capability is not listed, the worker should not have it.