Automate the deadline work, not the thinking

The filter

A task is a good automation candidate when it is repetitive, structured, deadline-driven, and cheap to verify. That last one carries most of the weight: if checking the output takes as long as producing it, automation has moved the work rather than removed it.

The two-tier pattern

The architecture that consistently works is two-tier. A cheap local model does the volume — drafting, tagging, extraction, overnight batches — at zero marginal cost. A frontier model does the judgement: the final pass, the hard reasoning, anything customer-facing. Between them sits the piece most people skip.

The quality gate is the product

Generated output at volume is worthless without an automated check, because the failure rate is never zero and one bad artefact reaching a customer costs more than the whole batch saved. A gate is a set of hard invariants, checked in code, that quarantines anything failing them.

• Required fields present and non-empty.
• No placeholder text — no unfilled template variables, no leftover brackets.
• Names, amounts, and dates match the source record exactly, checked by string comparison rather than by vibes.
• Length and format within bounds.
• Banned-phrase list for anything that would embarrass you.
• Deduplication against what has already been sent, so nobody gets the same message twice.

Everything failing the gate goes to a quarantine folder rather than out the door. Volume without a gate is not throughput, it is liability at scale.

Sequencing a real build

1. Instrument first. Count how often the task happens and how long it takes. Automating something that occurs twice a month is a hobby.
2. Write the gate before the generator. If you cannot specify what correct output looks like in code, you are not ready to generate it.
3. Run it in shadow mode. Generate, gate, and inspect — but do not send. Measure the pass rate.
4. Ship the highest-volume slice only. One workflow, end to end, beats five half-built ones.
5. Keep a human at the point of irreversibility until the pass rate has been stable for long enough that you would bet on it.

What it actually costs

Less than people expect on tokens and more than they expect on plumbing. A local model on hardware you already own is free per call, so the recurring cost of a high-volume drafting pipeline can be genuinely near zero. The real expense is the boring part: connecting the data sources, defining the invariants, handling the exceptions, and maintaining it when a format changes upstream.

Which is the honest summary of this whole field right now — the model is the cheap part.

A worked example of all of this in production is FreightDesk AI: local model drafts overnight, an automated audit quarantines anything failing its checks, a human sees only exceptions. The generation half is covered in the local LLM guide, and the retrieval half in the RAG guide.

Tools referenced in this guide

• FreightDesk AI — the pattern in production for freight brokerages.
• Local LLM guide — how to run the cheap tier on your own hardware.
• Automation Studio & store — playbooks and templates for building this yourself.