The five levels of AI adoption: from copilot to autonomy
Using an AI chat does not mean that an organisation has adopted AI. This framework distinguishes five levels according to autonomy, integration, accountability and the system’s ability to learn.
Two companies can claim that they “already use AI” while describing incompatible realities. In one, a few people consult a chatbot from personal accounts. In the other, an agent accesses internal systems, executes a process, records what it did and asks for approval when it encounters an exception.
The label is the same. The capability, risk and organisational change are not.
Automated driving offers a useful analogy because it separates assistance from autonomy. Precision matters here: the reference classification used in the automotive industry has six levels, from 0 to 5. The framework I propose for organisations has five, from 0 to 4. This is not a technical equivalence; it is a way to avoid calling every system that produces a convincing answer “autonomous”.
Level 0 — Accidental use
Individuals try public tools without a shared goal, common criteria or organisational awareness. There is curiosity and, occasionally, local improvement. There are also duplicated uses, results that cannot be compared and possible exposure of information.
The leadership question is not how to stop everything, but what people are trying to solve on their own. That behaviour reveals real needs.
Level 1 — Personal assistant
AI helps draft, summarise, translate, analyse or prepare a first version. The person initiates each task, provides the context and validates the result. Work improves at an individual level, but the learning usually remains with each user.
Training, basic data rules and the ability to verify matter at this stage. Automation is not required to create value.
Level 2 — Integrated copilot
Assistance appears inside the workflow: development, support, sales, operations or documentation. It has authorised context and produces results in formats compatible with existing systems.
The unit of improvement is no longer the isolated prompt. It becomes the process: what information goes in, what criteria are applied, who reviews and what happens to the output.
Level 3 — Supervised delegation
The system completes a sequence of steps, chooses between tools and maintains state until it reaches a bounded objective. A person reviews defined checkpoints or intervenes when exceptions arise.
This is the level at which it makes sense to talk about agents. It is also where traceability, least-privilege permissions, cost limits, evaluations and a clear route for returning control become indispensable.
Level 4 — Bounded autonomy
Several agents or components execute complete processes within a well-defined domain. This does not mean unlimited freedom: autonomy exists inside explicit boundaries, with observability, accountable people and stop mechanisms.
A mature system is not one that removes people. It is one that knows what it can resolve, when it should ask for help and how to demonstrate what it has done.
The right level depends on the problem
Maturity does not mean always reaching level 4. For an infrequent, irreversible decision with a high cost of error, an assistant may be more appropriate than an agent. Delegation may make sense in a frequent, observable and reversible process.
Before moving up a level, five capabilities should grow together: context quality, integration, observability, governance and clarity about who is accountable. Increasing autonomy without increasing those capabilities is not transformation; it expands the blast radius of an error.