Where does AI help in your process, and where is something simpler enough?
AI can be part of a process in which information is read, summarised or prepared for a next step.
We start with that task. Then we determine what input is needed, what a useful outcome is, and what control remains necessary before anything is sent, changed or published.
Three forms to keep apart
These three forms help as a trade-off, not as a ranking. An agent is not automatically the best solution.
- 1.Normal automation. Pre-determined rules determine the route, without a language model.
- 2.A workflow with AI. One or more defined steps use a language model, while the route and the controls are fixed in advance.
- 3.An AI agent. The model is given more room within boundaries to choose follow-up steps and tools itself. This requires explicit authorities, control and a way to stop or escalate.
For the distinction between pre-arranged workflows and steps driven by a model, we refer to Anthropic's explanation on building effective agents.
That source advises adding complexity only when the task actually requires it. A simple workflow is sufficient for many tasks and easier to control than an agent with a lot of its own space.
Concrete applications as illustration
- 1.Structuring information from an intake. Input: a freely filled-in request. AI step: recognise and organise relevant fields. Control: an employee checks the classification. Destination: the file of the request.
- 2.Signaling missing fields. Input: a partially filled-in file. AI step: check which mandatory information is missing. Control: an employee assesses whether inquiry is necessary. Destination: a task for the processor.
- 3.Preparing a draft response. Input: an incoming question and available information. AI step: draft a response. Control: an employee reads, adjusts and sends it themselves. Destination: the contact channel with the customer.
- 4.Answering an internal knowledge question. Input: a question from an employee and selected internal documents. AI step: compile an answer based on those documents. Control: the employee assesses whether the answer matches the source. Destination: the internal knowledge system.
- 5.Making a proposal for a page change. Input: existing page text and a requested adjustment. AI step: draft a version. Control: an editor assesses and approves for publication. Destination: the content management system, not yet the live page.
In each example, it is fixed what the AI step does and what it does not. Without that delimitation, it is difficult to trace afterwards where an error originated or who should have paid attention to it.
Assessment remains human work
Every application must deal with normal cases, exceptions and incomplete information. Substantively plausible output is not automatically correct. We do not set a quality percentage without this having been actually measured.
AI does not replace human assessment. Where an outcome has consequences for a customer, a decision or a publication, an employee remains responsible for the final check.
That applies just as much to a simple workflow as to an agent with more of its own space. More autonomy for the model does not mean less responsibility for the human who uses the result.
Boundaries on implementation
Reading is something else than writing. Creating a draft is something else than publishing. Drafting a budget proposal is something else than actually changing advertising expenditure.
We only give an application the authorities needed for the defined task, no more. Further-reaching authorities require a separate trade-off and extra control.
This division into authorities is reviewed again for each application. A task that only reads today can later, after sufficient experience and control, be given limited writing authority.
Data and privacy
Which data ends up where and how long it is kept is agreed with you and recorded per application. Technical choices and standards are also tailored to your situation.
Further reading
If you first want to know how we look at a process as a whole, read about automation. For the collaboration form in which an AI application is built, see software development via DaaS. The broader overview is on the page about technology.