AI agents and knowledge support

Answers or actions?

A knowledge assistant focuses on finding information and answering questions. An AI agent may also call tools, connect steps or take actions. Answering a question does not establish reliable order changes or stock updates.

RAG retrieves relevant material before generating an answer. A more complex agent needs clear task steps, available integrations and controlled permissions.

Start with lookup and recurring questions

Review product manuals, service rules or internal instructions: who maintains them, which version is valid, who can access them and what source should support an answer.

When information conflicts, is outdated or lacks evidence, the assistant should say it cannot confirm and refer the issue to a person. Keep important commitments and external responses subject to human review.

What does a pilot need?

Select authorized, redacted documents with known versions and representative questions. Separate questions that can be answered directly, need clarification or require human handling.

Test retrieval, citations and answers first. For actions, assess interfaces, read and write access, confirmation, records and failure handling. Agree changes to existing systems separately.

Assess evidence and behavior

Use questions checked by business staff to assess retrieval, whether citations support the answer, outdated content handling and appropriate refusal when evidence is missing. Record error types before expanding.

For actions, also test the target object, duplicate execution, recovery and human takeover. Model choice, data volume, integrations and maintenance affect cost and need separate agreement.

Knowledge and support collaboration

In the related RAG support project, I worked on knowledge preparation, retrieval configuration and testing. Illustrative answers are distinguished from project records. Tool-executing agents require a separate assessment.

UK water purification brand · RAG support ↗

Contributed to customized RAG support through knowledge base organization, retrieval configuration and testing, helping the team progress from demos to live testing and delivery.

Questions before starting

Do we have to train a model?

Not necessarily. Assess existing models, retrieval and tool configuration against the task before investing in training.

Can an agent handle customer work autonomously?

Validate a specific task first. Orders, payments, commitments and data changes need defined permissions, confirmations and exception handling before deciding the automation level.