The work item is the product
Varun's AI adoption critique points to the real enterprise unit: not seats, prompts, or tokens, but prepared work items with evidence, policy, routing, and receipts.
Varun's AI adoption critique points to the real enterprise unit: not seats, prompts, or tokens, but prepared work items with evidence, policy, routing, and receipts.
Single-call AI policies can block or redact one request. Contextual policies add session state, so the proxy can catch risk that only becomes clear across several turns.
Most LLM security tools conflate blocking and redacting. In a multi-turn agent session, confusing the two breaks both model reasoning and the legal audit trail.
Notes on incorporating, picking a C-Corp, and why I'm working on the infrastructure under agents instead of another agent product.