Your agent is ready. Your business software isn’t.
A capable model can still lack a dependable way to act inside the ERP, desktop application, portal, or internal tool where the workflow lives.
Read noteGraft notes
We write about the part of agent deployment that begins after the model is chosen: reaching real workflows, setting boundaries, and proving that an action happened where it was supposed to happen.
Edition 001 · July 2026A capable model can still lack a dependable way to act inside the ERP, desktop application, portal, or internal tool where the workflow lives.
Read noteWhen an agent says “done,” the important question is not whether the interface moved. It is whether the intended business effect exists in the source system.
Read noteWhen a business action has no usable API, prepared tool, or permitted computer-use path, the deployment stops at a human handoff: the agent can decide, but it cannot do.
Read noteAI agents should be able to use existing business software through explicit actions, narrow permissions, and source-system evidence that operators can inspect.