A recurring work step is a good reason to consider automation. The right technology depends on the inputs, the decisions and the consequences of an error. A fixed script, a controlled workflow and an AI agent solve different tasks. Often the best solution combines them.
The first question is therefore: what must be completed reliably at the end? Only then come the tool, model and operating approach. This also makes it possible to assess whether additional AI complexity brings a clear benefit.
Three possible approaches within the same process landscape
A possible example: incoming service requests
A request contains free text and occasionally an attachment. A conventional step first checks required information and permitted file formats. AI proposes a category and summarises the request. The application checks the result against allowed categories. A responsible person decides how it will be handled.
A clear workflow is often sufficient for this process. An agent becomes interesting only when it really needs to choose different approved sources or tools for different requests. This example describes a possible architecture, not a claimed customer project.
Define the boundary between suggestion and execution
A persuasive written answer is not permission to change a system. The application itself must check who is acting, which object is involved and which action is permitted. Model outputs are treated as suggestions; the actual function validates its inputs and enforces business rules.
For sensitive changes, an approval shows the specific target state and intended action. If the draft is subsequently changed, the earlier consent may no longer apply. Which steps may run without asking again is therefore agreed for each use case.
An MCP interface can expose defined tools. It does not replace their authorisation. The official MCP security documentation (external link) covers topics including separation of access scopes and protection against tool misuse.
What should happen after an error
A timeout does not automatically mean the target action did not happen. Before retrying a write, it must be possible to determine whether the first request has already been carried out. Suitable operation identifiers, unambiguous states and a targeted result query help prevent duplicate changes.
AI-assisted processes have additional stopping conditions: missing evidence, conflicting information, a tool that is not permitted or an exhausted runtime limit. The responsible person then receives the current state and the specific issue needing clarification. A traceable handover point is a normal process outcome.
Use these cases to test a pilot
- A normal incoming request: Are the result and required follow-up work useful from a business perspective?
- An incomplete request: Does it ask a targeted question instead of adding something nobody has confirmed?
- An unauthorised access attempt: Does the function refuse access regardless of how the request is worded?
- A repeated request: Does the state remain clear, without a duplicate booking or change?
- An outage: Can people continue the task with the information available?
The comparison includes more than response time and model costs. Review work, integration, error recovery and operations per usefully completed task also matter. This shows whether a script is enough, a workflow should be extended or a restricted agent offers a demonstrable additional benefit.
Agree the first manageable step together
DAXS records the current process, clarifies interfaces and develops a suitable implementation. This includes tests, approvals, documentation and agreed operations. Existing applications and conventional automation remain as important as new AI tools.
AI and automation overview · Development and integration · Introducing AI in your business
