A copilot suggests and a person decides. An assistant executes a defined task on instruction. An agent plans and executes within boundaries, iterating until the goal is met. An agentic system is several specialised agents with shared state, orchestrated under policy controls, delivering a governed outcome. Those four levels run from human-led to AI-owned, and most of the confusion in boardroom AI conversations comes from using the words as synonyms. They are not, and the differences decide what has to be built, what it needs access to, and what governance applies.
One scenario, four levels
The exercise we run in leadership sessions takes one scenario from the client’s own business and shows it at each level. For a drives manufacturer: a customer requests a complex drive configuration and a quote.
| Level | What the AI does | In the scenario |
|---|---|---|
| Copilot | AI suggests, a person decides. A language model with retrieval surfaces context; the person approves each step | Recommends configurations, flags compliance risks, drafts the quote email |
| Assistant | AI executes, a person instructs. A model with function calling runs defined tasks on demand | Generates the export quote, applies the pricing rules, produces the final document |
| Agent | AI plans and executes within boundaries. A planning loop with tool access, iterating until the goal is achieved | Plans the quote workflow, resolves pricing and compliance, delivers an approved quote |
| Agentic system | AI delivers governed outcomes. Specialised agents with shared state, orchestrated under policy controls | Optimises price, risk and availability across coordinated agents and delivers a margin-optimised proposal |
The same table for a marketing team reads: “help me write this email” at copilot; “draft the email, segment the audience, generate variants” at assistant; “increase lead conversion by 15%, I’ll decide what to test” at agent; and at agentic system, “increase qualified lead conversion by 15% this quarter within the existing budget, with access to CRM, analytics, paid media and content tools, and alert me if the risk to target exceeds 10%”.
What changes between levels
Three things, and they are the things that matter to a business rather than to a demo.
Who decides. At copilot and assistant, a person. At agent, the system within boundaries, with escalation. At agentic system, the system, under policy, with a person informed and able to intervene.
What the system needs access to. A copilot needs retrieval over documents. An assistant needs function calls into specific systems. An agent needs tools, permissions and the ability to act in the ERP, the CRM, the helpdesk. An agentic system needs shared state across agents and orchestration. Each step up is an integration and governance decision, not just a model decision.
What happens when it is unsure. At copilot, nothing: the person was deciding anyway. From assistant upwards, the system needs a confidence score, a threshold and a handover, and the design of that escalation is most of the engineering.
The levels are not a ladder
Not every use case should climb. A copilot is the right end state for work where judgement is the point: editorial decisions, safety assessments, anything where the person’s accountability is the product. An agentic system is right where the outcome is measurable, the boundaries are clear and the cost of a wrong decision is bounded by escalation. Deciding which level each use case should reach, and stopping there, is as important as the technology, and it is a decision the leadership team should make with the vocabulary to make it.
Why it matters for governance
Under the EU AI Act, knowing which level you are building tells you which obligations apply. A customer-facing assistant carries a transparency duty. An agent making decisions that affect people carries human-oversight requirements. An agentic system touching credit, employment or safety may be high risk. The four levels are not only a way of describing capability; they are a way of classifying risk before anything is built.
What we find in practice
Most of the value in the next twelve months, in most businesses, is at the copilot and assistant levels, across the whole organisation, on governed enterprise tools. That is usually a relief to hear. Agents come next, one bounded workflow at a time: a support queue, a renewal process, a supplier-intelligence brief. Agentic systems are a programme, built from agents that have each proved themselves in production.
The four levels have their own page under our approach, and the exercise is run on your own scenario in a leadership awareness session.