Intelligent automation is having a moment because language models can finally sit in the middle of messy work. That is genuinely useful for documents, email, and classification. It is a poor reason to abandon deterministic workflows.
If a step is ‘if the invoice is over $10,000, send it to finance’, you do not need an agent. You need a rule, an audit log, and a way to change the threshold without a developer. Agents earn their keep when the input is unstructured or the next step depends on meaning, not a field.
The failure mode we see is a lattice of prompts that only the author understands. Six months later, a vendor API changes, a prompt drifts, and nobody can say what the process is supposed to do. That is not modern operations. That is a new kind of tribal knowledge.
Our bias is simple. Encode the happy path. Route exceptions. Use models to extract, summarise, and draft — with a human on anything that moves money, access, or clinical or legal risk. The result is less impressive in a demo and much cheaper to run.
