Workflow automation that replaces busywork—when not to use an LLM
Deterministic jobs beat chatty models for status updates, routing, and form grunt work
Not every automation needs a language model. Most busywork dies to rules, queues, and templates. Reach for an LLM when language itself is the hard part—not when a switch statement would do.
Busywork that should stay deterministic
Status emails, SLA timers, role-based routing, invoice number assignment, and “if field X then create ticket Y” belong in workflows with tests. They should be boring, replayable, and cheap.
LLMs add variance. Variance is expensive when the correct behavior is already known.
Pick the shape that matches the uncertainty
| Shape | Use when | Avoid when |
|---|---|---|
| Deterministic | Rules are known: routing, timers, templates, field → action | Input language is messy and rules would thrash weekly |
| LLM | Language is the hard part: drafts, messy intake, novel phrasing | Correct output is already a switch statement |
| Hybrid | Extract + validate, then run rules on structured data | You skip validation and “creatively” continue on bad extracts |
Where LLMs earn a seat
Unstructured intake, messy email classification with novel phrasing, drafting replies for a human to edit—these are language problems. Even then, keep the model in a suggestion lane until accuracy is proven.
Hybrid is usually the answer
Extract structured fields with validation schemas. Run deterministic automations on the structured result. Use a model only for the fuzzy middle. If extraction fails validation, stop—do not “creatively” continue.
Validate before automating
Loading code…Measure the replacement
Track minutes saved, error rate, and how often humans override. If overrides are common, the automation is a suggestion UI wearing a production badge—fix the contract before expanding scope.
Final thoughts
Automate with rules when the rule is known. Invite a model when language is the uncertainty. Validate structured outputs before side effects. Busywork dies to clarity—not to novelty.
Related Articles
AI for small businesses: useful automations vs expensive demos
Inbox triage, quoting, and CRM sync—when to use an LLM and when rules win
Cost control for LLM features in production
Budgets, caching, and model tiers so AI remains a product cost—not a surprise invoice
Ready to build something this considered?