AI agents and business automation

Not all tasks are equal. The ones that work have three properties: an input that arrives regularly, an output that can be checked, and an error cost that is either bearable or detectable. A supplier invoice to enter meets those conditions; a credit approval decision does not.

So we start by sorting. Some requests come back with a reasoned refusal, and that is a useful result: it avoids spending six months on an automation that would never have survived production.

• Entry and reconciliation of invoices, purchase orders, statements
• Classification and routing of incoming messages
• Product copy written under verifiable constraints
• Catalogue translation with markup preserved
• Monitoring and summarisation of sources you designate

An agent without limits ends up calling a paid service twenty times inside a loop it cannot leave. We set a budget per task, a maximum number of steps, tools whose permissions are checked at execution, and a log readable by someone who did not write the system.

Decisions with financial or legal consequences stay approved by a human. That is not window dressing: it is the condition for the whole thing to remain auditable.

• Budget per task and a capped number of steps
• Explicit tools, permissions checked at execution
• Complete action log, retained and readable
• A clean line between proposing and executing
• Immediate manual override if things drift

Automation you do not measure is a belief. From day one we count volume processed, human rework rate, cost per unit and expensive errors. If one of those numbers does not improve, we say so.

• Volume processed and human rework rate
• Cost per unit processed, against the original cost
• Expensive error rate, tracked over time
• A decision to stop when the gain fails to appear