Drafting a reply to a customer leaves the decision with an employee. Allow the same agent to send the reply, change a price or promise a delivery date, and you have delegated authority as well as a task.
That boundary needs to be explicit before the agent becomes part of the workflow. Which actions can it take, which require approval, and who can intervene? Those decisions are the starting point for a useful design.
From proposal to action
Chatbots became steadily better at answering, summarising and explaining. An AI agent is also given a goal and access to tools, systems and data, so it can carry out tasks itself. Within the instructions and access it has, it chooses its own next steps. For example, it can:
- process email
- compile reports
- prepare proposals and decisions
A fixed workflow follows a predefined sequence. An agent selects its own next step within a task. A system can combine both approaches. For the organisation, that difference has real consequences.
How much independence should an agent have?
The discussion about AI is often framed as autonomy or none, control or none. Independence is better seen as a scale. An agent can:
- only prepare information
- make recommendations
- execute decisions within strict boundaries
- or carry out specified tasks without approval at each step
Risk depends on the agent’s discretion, access, task and the consequences of error.
Many organisations experiment with agents without defining where on this scale they operate. That is a governance and design question as well as a technical one. As soon as a system can take independent action, it must be clear who authorises those actions and who intervenes when they go wrong.
- Who is accountable if an agent makes the wrong commitment?
- Who keeps customer communication consistent?
- Who checks whether decisions can be explained?
These questions are easily left unanswered. Build access and delegated authority into the design deliberately.
Data and dependencies
An agent is only as effective as the environment it works in. If data is unreliable, processes are fragmented and responsibilities are unclear, an agent will not make the work more efficient. It mainly speeds up the confusion.
Technological ambition can run ahead of the quality and organisation of the data. You do not have to stand still. Do think carefully about the design and about the workflows, data and systems the agent relies on. Not everything that could run independently should do so straight away.
Put the boundaries into practice
The acceptable margin for error depends on the task and its consequences. An awkward draft is different from an incorrect commitment sent to a customer. Before an agent handles customer records, contracts or financial actions, establish the required access, checks and means of intervention.
At the same time, employees are often already experimenting with AI tools outside the agreed set-up. IT tries to keep up, and security either reacts after the fact or becomes restrictive. Management sees the potential but may lack a framework. That calls for explicit decisions:
- Which tasks do we want the agent to carry out, and why?
- How much independence do we want it to have?
- Which controls do we build in?
- Which actions require an employee’s approval before execution?
- What should we record, how can actions be checked and who is accountable?
Three decisions before an agent starts work
Specify the actions the agent may take, those that require approval and who can stop it. Then check whether its permissions and the workflow enforce those decisions. Written instructions are a weak boundary if the system has broader access than the task requires.
You can then test whether the application is useful. That test may also show that a simpler solution is a better fit.