A specific job, not a universal bot
A useful starting task is finding answers in approved materials, preparing a draft, collecting input or helping an employee navigate documents. For a simple fixed rule, AI may be unnecessary: ordinary automation can be more predictable.
Complex decisions, promises to customers and consequential actions require a defined human role. Autonomous message sending is not enabled by default.
Quality begins with the sources
We gather materials, assign owners, identify current versions and flag contradictions. The assistant should distinguish an approved answer from an outdated rule or a lack of information.
We do not indiscriminately import an entire message history. A pilot uses anonymised examples where possible. Data scope and access permissions are defined for the specific task.
Permissions are visible from the start
- Read: agreed sources only.
- Prepare: a draft answer or structured information.
- Suggest: an action with its context.
- Execute: only within agreed boundaries and with the required confirmation.
Human handoff is a valid outcome when information is insufficient, the question is outside scope or the user asks for a team member.
A pilot tests more than successful answers
The test set includes precise questions, missing answers, conflicting documents, outdated conditions and requests for a human. Answers are compared with agreed criteria, errors are recorded and fixes are retested.
After launch, the workflow needs knowledge updates, error review and an owner. AI can make mistakes; promising complete accuracy would be misleading. Experience building my own assistant, Skoobi, informs these boundaries but does not replace testing your use case.
Before we start
Does it require personal data?
Not every use case does. Public instructions and anonymised examples can be enough to start; required data and permissions are agreed separately.
Can we begin with an internal assistant?
Yes. Draft preparation and approved-document search allow a pilot without autonomous customer communication.
How do we know it works well?
Before the pilot, we define examples, answer criteria, refusal conditions and human handoff. A successful demonstration does not replace those checks.