IN BRIEF
Agentic AI refers to AI systems that break a task into steps, call tools and systems to carry them out and adjust their approach based on intermediate results. Unlike an assistant that answers a request, an agent acts within a process. Its value and its risk therefore depend on the permissions it receives and on where a person approves.
When to involve us.
When a process touches many systems, handles unstructured input and is slowed down by handoffs and rework, it is worth examining whether an agent can take over individual steps.
We break the process into steps and assess, step by step, whether an agent, conventional automation or a person is best suited to it. For the agent's steps, we define tool access, permissions, approval points and fallback paths, and redesign the process where needed. We support the planning and evaluation of a limited pilot using real cases. Before it starts, we agree who will handle development, integration and operations. The evaluation shows how reliably the agent performs and how much review effort remains.
What the work produces.
We agree on the scope around your goals, starting point and available capacity. The following elements illustrate what an engagement can include.
WHAT WE ADDRESS
- Process analysis and suitability assessment per step
- Autonomy and approval model
- Target process with tools, permissions and fallback paths
- Pilot planning and evaluation against success and stop criteria
POTENTIAL OUTPUTS
An assessed target workflow with a permissions and approval model, plus an evaluation of the agreed pilot. The findings show whether reliability, value and review effort support moving into operation, call for further changes or justify stopping.
A focused starting point.
A concrete process with sample cases, an overview of the systems involved and the process owners. Information security and the people responsible for access rights should be involved from the start.
The question first. Then the scope.
The first conversation clarifies the objective, deliverables and people involved. We agree whether the engagement covers design, pilot support or technical implementation, who takes on each task and how the deliverables will be assessed.
Discuss a starting point by emailHow much autonomy should an AI agent have?
As much as the process can tolerate, and no more. What matters is the impact of an error, whether it can be reversed and how quickly it is noticed. Read-only steps such as research or summarisation can usually run more freely than steps that change data, trigger payments or communicate externally. Those steps call for approval points, narrowly scoped permissions and complete logging. Autonomy can be extended step by step once evaluations over a longer period show consistent quality. It must also be clear at all times how an agent is stopped and how the process continues manually.
Last updated:
