Operations, data and finance scenarios

Staff want to ask everyday questions of structured business data, such as tables, totals and operational figures, without writing queries.

1. An employee asks a data agent.
2. The agent calls a scoped MCP server built on the business's API.
3. It receives the relevant rows or totals.
4. It calculates with the code interpreter, or uses totals calculated on the server.
5. It explains the result in plain language.

If the data is a bounded dataset that doesn't need to be live, Tables in raia Command offer exact rows and read-only agent access as an alternative.

Knowledge search returns relevant passages, not complete datasets. Counts and totals need every row and real arithmetic.

• Live database, or a periodic snapshot?
• Which tables, fields and limits may each employee see?
• Which calculations can be done on the server?
• Is write access needed? Usually it isn't.

Operations staff want to ask about the current state of machinery or live sensor readings held in a database.

1. A signed-in employee asks an operations agent.
2. The agent calls a read-only MCP server built on the telemetry API.
3. It retrieves current values, with timestamps and units.
4. It calculates in code.
5. It answers, stating how recent the data is.

Sensor data changes constantly and must be exact. It belongs in a live, read-only connection, never in knowledge.

If a common question needs three or four API calls, build one endpoint that returns the combined answer, and expose that instead. APIs designed for agents are faster and more reliable than reusing system-centric ones as they are.

• How fresh must readings be?
• How are devices identified?
• Are units and scales consistent?
• What happens when a sensor is missing or data is stale?
• Which time zone applies, and what are the alarm thresholds?

• Loading sensor data into knowledge and calculating from retrieved passages.
• Treating a suggested diagnosis as an instruction to the machine. The agent advises; it doesn't operate equipment.

Employees submit receipts. The business wants details extracted automatically and each claim checked against policy.

1. Employees photograph or upload receipts in a simple, mobile-friendly form.
2. An agent extracts the details and fills in the fields.
3. The employee confirms or corrects them.
4. A webhook starts a Loop.
5. An agent checks the claim against expense policy and looks for duplicates.
6. A finance approver, and a manager if required, decides.
7. Approved claims update the finance or expense system.

Uploading and correcting fields is easier on a purpose-built screen than in a chat. raia does the extraction, checking and workflow behind it.

• What's the expense policy, and what are the spending thresholds?
• How are duplicates detected?
• Who may see personal payment details?
• What's the approval chain, and which expense system is the target?

• Forgetting the front end. The upload form is a separate application that someone must build and maintain. Adding agents and Loops doesn't create it.
• Posting without approval. A person approves financial postings, according to policy.