When not to use AI: jobs a cron job does better
We sell AI work, and a fair share of our advice is "don't use AI here." The three tests that separate agent jobs from script jobs, with examples from our own stack.
We run AI agents in production and sell that work to clients. We also run a lot of cron jobs, and some of the best advice we give prospects is "this one's a script." Here is how we draw the line in our own stack.
Test 1: Is the job deterministic?
Same inputs, same steps, same output, every time? Script. Our nightly business-data sync is a scheduled job with retries and logging — no model involved, because there is no judgment involved. It has run for months without an interesting failure precisely because it is boring.
The moment a job needs reading, summarizing, drafting, or classifying ambiguous input — that's where a model earns its inference cost.
Test 2: Can you afford to verify the output?
Language models are wrong sometimes. That is a property, not a bug report. So every AI step needs a verification story: a human approver, a test gate, a reconciliation check. If verification would cost more than doing the job manually, the automation is theatre. Our deploy pipeline is deliberately AI-free — build, rsync, health check — because deploys need to be provably correct, not plausibly correct.
Test 3: Does the job justify its complexity budget?
An agent pipeline is more machinery than a script: prompts, models, gates, logs. Machinery needs maintenance. A job that runs once a quarter or takes a person four minutes rarely justifies it. The payoff curve favours jobs that are frequent, language-heavy, and judgment-adjacent — drafting, triaging, summarizing at volume.
What this looks like in practice
- Agents: drafting blog content from research, turning user signals into feature specs, analyzing reviews and photos to propose listing updates.
- Scripts: data syncs, deploys, backups, health checks, report generation, image processing.
- People: publish decisions, merge authority, pricing, anything with a customer's trust on the line.
The pattern: models where there's language and judgment, scripts where there's repetition, people where there's accountability. A vendor who wants to put a model in the scripts column is selling you inference you don't need. If you want help sorting your own list into those columns, that's the consulting service — and yes, sometimes the deliverable says "cron job."