A tool is a defined operation, not unlimited access to a platform. Its name, description and input fields tell your AI application when and how to use it.
Inspect before running#
Ask the AI to list the tools on your MCP DigLab connection and explain their purpose. Find the narrowest operation that answers the task. A search, a single-record read and an update may be separate tools with different permissions.
If the expected tool is missing, check your handoff. It may not be included or may still be waiting for account verification. Do not ask the AI to guess an undocumented API call as a replacement.
Give complete inputs#
Be precise about the account, workspace, date range, record identifier and desired output. Use an explicit timezone when a task involves scheduling or daily reporting. If two records have similar names, verify the identifier before making a change.
Use my approved CRM read tools to find the test account named
Example Company. Return the record ID, account name and status.
Do not create or update a record.This is a fictional example. It works only if your setup includes an appropriate CRM tool.
Review the result#
Ask the AI to distinguish returned facts from its interpretation. Check whether a result is complete, paginated, filtered or empty. A zero-result search does not always prove that a record does not exist; the filter, workspace or access scope may be wrong.
When evidence is missing, narrow the query or ask DigLab about the available tool. Avoid repeated broad searches through sensitive data.
When the AI answers without using a tool#
A conversational answer is not proof that a live source was checked. Ask which tool ran and what result it returned. If you need current business data, make the request explicit: use the connected source and identify the evidence.
Finish the task#
Record the final output and any item identifiers that matter. For a write task, verify the resulting state in the source system. The AI saying that it completed an action is not a substitute for an accepted result.