Defined job
State the specific questions, drafting task or action the assistant is expected to handle.
AI assistants and agents
AI assistants and agents with human control
Help people find, draft and act faster while keeping consequential decisions, permissions and exceptions visible.
Automation becomes risky when authority is broader than understanding.
An assistant that answers from unapproved information or an agent that can act without clear limits may save seconds and create larger problems. Useful autonomy is bounded: the sources, actions, approvals, logs and failure path are designed before release.
AI assistants and agents
Bounded help, not unexplained autonomy
The controls below are design commitments. The exact technical and commercial scope is confirmed in the written proposal.
Approved knowledge
Connect only relevant sources and respect the access rights already attached to the underlying information.
Permission limits
Separate read, draft, recommend and act permissions; keep administrative capabilities out of routine use.
Human checkpoints
Require approval before sensitive communications, record changes, financial commitments or external actions.
Observable operation
Log useful context, outputs, approvals and errors so that owners can investigate and improve behaviour.
Exception path
Tell users what the assistant cannot do and route uncertain or high-risk cases to a responsible person.
A clear next step
Build from a bounded role
Each stage has an owner, an output and an approval point so that progress remains understandable.
Define
Choose the job, authorised users, approved sources and actions that remain human-only.
Test
Evaluate representative questions, difficult exceptions, permission boundaries and failure behaviour.
Operate
Assign ownership, review logs and feedback, measure usefulness and keep a documented way to disable the capability.
Illustrative deliverable — not a client result
An operational assistant specification
- Purpose, users and out-of-scope requests
- Approved sources and source-level permissions
- Action limits, human approvals and exception routes
- Evaluation set, monitoring, ownership and disable procedure
Direct answers
Questions Swiss organisations ask before changing IT or introducing AI
Clear answers help you compare providers, question assumptions and make a more informed decision.
What is practical AI?
Practical AI starts with a specific business workflow, a measurable outcome and controlled access to data. We design a limited pilot, document the controls and only recommend expansion when the result is useful and supportable.
How do you protect confidential information?
The design begins with minimum data access, explicit permissions and a documented choice of model, vendor, data location, retention, logging and training settings. Sensitive actions can require human approval, and every pilot includes a rollback path.
Can AI integrate with our current tools without replacing them?
Often, yes. An assistant or automation can connect to approved systems and fit into an existing workflow. We first check the available interfaces, permissions, data quality and fallback process before proposing an integration.
Start with the problem, not a technology purchase.
Book a 30-minute assessmentBook a 30-minute assessment
Book a 30-minute practical assessment
Tell us about one IT or workflow problem. You will leave with an initial view of the risks, feasible options and the most sensible next step. No technical preparation is required. Fields marked with an asterisk are required.
What happens next
- Describe one problem and suggest a convenient time.
- We use the context to focus the 30-minute conversation on IT, AI or both.
- The conversation ends with an initial view of risks, feasible options and the smallest sensible next step.
+41 22 552 08 10
info@dspsa.ch