AI (Artificial Intelligence)

An AI project is easiest to justify when it starts from a workflow that already exists: a queue of documents someone reads by hand, a knowledge base nobody can search, a support team re-keying the same lookups, or a forecast kept in a spreadsheet. IKRC adds models to that workflow and to the software that runs it. Work can include document review and classification, predictive analytics, fraud and anomaly detection, and assistants that answer from company data or act on it.

Where the design effort goes

Calling a model can be a small part of the build. Whether the feature is safe to run depends on the parts around it:

• Retrieval: which documents and records the model can see, and whether results respect each user's existing permissions.
• Tool calls: an assistant that can look up an order or change a record needs the same authorization checks as any other caller, applied where the action executes and using the actual user's identity. Access to a tool does not by itself approve the business action.
• Limits: in Microsoft.Extensions.AI v10.9.0, the function-invoking chat client allows up to 40 model iterations per request by default (MaximumIterationsPerRequest, which counts the initial request). IKRC's approach is to set limits like this explicitly for each feature and to test what happens when a tool call fails.
• Diagnostics: logs that support staff can use, kept access-controlled, with sensitive arguments redacted.

Related Reading

For how an assistant can reach your platform's data and actions, read MCP Servers for Custom Software Platforms. For searchable institutional knowledge, read Keeping an AI knowledge base current and permission-aware. For narrow agents built around one job, read Why a Task-Specific Agent Beats a Chatbot Bolted Onto Your App.