An agent is a software actor
An assistant waits for a request and proposes content. An agent can interpret a goal, plan steps, call tools, change a system and coordinate other agents. Once it acts in a CRM, ERP, customer account or supply chain, it uses part of the organisation’s authority. The decisive question becomes: on whose behalf does it act, with which rights, for what outcome and under whose accountability?
Identity comes before intelligence
Every production agent needs its own identity, limited permissions and a human owner. Shared accounts destroy traceability, while broad permissions turn local errors into systemic incidents. Least privilege must apply to data, tools and financial commitments. The architecture must also preserve a readable trace of the objective, steps, sources and final action.
Autonomy should be graduated
The choice is not simply human or autonomous. 8mind distinguishes four levels: observation without action, human-approved recommendation, execution with exception-based supervision, and autonomy within a bounded scope. An agent advances only when performance, incidents and rollback capability meet defined thresholds. This makes trust measurable.
Human control must remain effective
A human approval step is not enough if the reviewer lacks time, information or authority to challenge the agent. Control must be designed: context shown to the supervisor, decision time, escalation criteria, sampling and required expertise. For high-impact actions, the organisation must ensure that the human understands the recommendation rather than merely confirming it.
Plan for incidents before production
Test drift scenarios: malicious instructions inside a document, inconsistent data, unavailable tools, runaway consumption or irreversible action. Spend limits, sandboxes, logs, emergency stops and recovery procedures are part of the product. Governance must know the review frequency, warning indicators and the person who can disable the agent immediately.
8mind framework
The four levels of delegation
Observe
The agent analyses in shadow mode without changing the system.
Recommend
It proposes; a human decides and executes.
Execute
It acts within limits, under exception-based supervision.
Autonomise
It operates a reversible scope with continuous audit.
Sources & method
The figures come from the publications below. They provide international orders of magnitude; their relevance must be validated in each organisation’s context.
McKinsey / QuantumBlackThe State of AI 2025↗BCGHow Agentic AI Is Transforming Enterprise Platforms↗NISTAI Risk Management Framework↗This publication is an original 8mind analysis. It turns public data and market findings into decision implications and frameworks for leaders without reproducing source texts.