What makes an AI system truly agentic?
“Agentic” does not mean that the system possesses unlimited independence or human-like judgement.
Different systems operate with different degrees of autonomy.
However, a genuinely agentic system should combine several of the following capabilities:
Capability | Meaning |
Goal understanding | Translates a general purpose into concrete activities |
Planning | Determines which actions to take and in what sequence |
Tool use | Interacts with email, databases, calendars and other software |
Memory | Retains information that remains relevant during the task |
Adaptation | Revises the plan when it encounters an obstacle or receives new information |
Autonomy | Continues operating without requiring human instructions at every step |
Verification | Evaluates whether the result corresponds to the original objective |
This is also why not every solution marketed as an “AI agent” is genuinely agentic.
Some are chatbots connected to a few external tools. Others are conventional automations with a conversational interface.
Others follow predetermined sequences while creating the impression of independent action.
The risk is that organisations may purchase an integration carrying a more sophisticated name and a higher price.
Gartner has described this inflation of the term as “agent washing” and predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value and inadequate risk controls.
But the distinction between a genuine agent and a repackaged automation is only part of the problem.
A real agent can still fail
Even a technically authentic AI agent can fail when it enters an organisation that is not prepared for agency.
An agent may be able to plan and act, yet still produce poor outcomes when:
the objective is ambiguous;
the data reflects outdated assumptions or hidden biases;
decision rights have never been made explicit;
different functions pursue conflicting priorities;
no one knows when human approval is required;
responsibility for the consequences remains unclear;
efficiency is measured without considering the value that may be damaged.
Imagine an AI agent instructed to reduce operating costs.
Which costs should it reduce?
Over what period?
What must it protect?
Can it modify supplier relationships?
Can it sacrifice service quality?
Who decides whether immediate savings justify longer-term reputational damage?
The agent can optimise only according to the architecture of objectives, criteria and constraints that surrounds it.
If that architecture remains invisible, the technology may accelerate decisions without understanding what the organisation is unintentionally sacrificing.
Autonomy requires boundaries
Agentic AI should not mean unrestricted freedom.
An effective agent needs a clearly designed perimeter that may include:
defined authorisations;
controlled access to information;
spending thresholds;
escalation criteria;
human approval for sensitive actions;
traceability of decisions;
mechanisms for correction or interruption;
clear accountability for final outcomes.
The challenge is therefore not to choose between complete human control and complete machine autonomy.
It is to decide where autonomy should begin, where it should end and where human judgement remains indispensable.
This requires organisations to ask questions that go far beyond technological performance:
Which decisions can be delegated?
Which decisions must remain human?
When must the agent stop and request approval?
Which values cannot be reduced to an optimisation criterion?
Who is accountable when an autonomous action creates an unintended consequence?
What happens when the agent’s recommendation challenges established power, incentives or informal practices?
Which organisational assumptions are being embedded in the system?
These are questions of governance, leadership and organisational design.
The invisible architecture behind the agent
Every organisation already contains an architecture of decision-making.
It exists in formal reporting lines, but also in less visible elements:
assumptions, incentives, relationships, power structures, habits, fears and unspoken rules.
Agentic AI does not enter a neutral environment.
It enters this system and begins to act inside it.
It may reproduce its contradictions.
It may expose them.
It may accelerate them.
A company can delegate tasks to an agent without realising that it is also encoding a particular interpretation of what matters, what should be prioritised and what can be ignored.
This is why the adoption of agentic AI cannot be treated solely as an IT implementation.
Before delegating action, leaders need to make visible the logic already shaping their own choices.
Otherwise, the organisation risks automating not only its processes, but also its blind spots.
The real transformation
The most important question may therefore not be:
Is this technology truly agentic?
It may be:
Is the organisation ready for agency?
Many agentic AI projects may fail not because the agent is “fake”, but because the organisation has not reconsidered the architecture of decision-making required to support it.
Agentic AI marks a transition from technology that assists human action to technology that participates in action.
And once AI begins to act, leadership can no longer focus only on what the technology is capable of doing.
It must also determine what the system should be authorised to do, which human values must frame its choices and who remains responsible for the consequences.
The future of agentic AI will not depend solely on the intelligence of the agent.
It will depend on the intelligence of the system designed around it.
#AgenticAI #ArtificialIntelligence #AIGovernance #Leadership #SystemicThinking #DecisionMaking #OrganizationalTransformation #ExecutiveCoaching #FutureOfWork #InvisibleDynamics
Comments