Blog · July 5, 2026 · 6 min read
When to Automate (and When Not To): A Working Framework
Frequency × volume × error cost: a practical framework for deciding what to automate, what to keep manual, and how to choose between scripts, RPA, and AI agents.
The most expensive automation is the one that shouldn't exist. After building automations across email, ERP data entry, marketing ops, and research workflows, I use a simple screen before writing any code: frequency × time-per-run × error cost. If the product of those three doesn't clear the build-plus-maintenance cost within a quarter or two, don't automate it yet.
The four quadrants
- High frequency, deterministic → classic scripting or RPA. Same inputs, same steps, every day. This is where things like SAP GUI scripting shine.
- High frequency, judgment required → AI agents with human escalation. Email triage, lead qualification, first-draft anything.
- Low frequency, deterministic → a checklist, not code. Automating the quarterly thing you do in 20 minutes is hobby engineering.
- Low frequency, high judgment → keep it human. Automate the prep (gather data, pre-fill drafts), never the decision.
Error cost decides the architecture
A wrong label on an internal doc costs nothing; a wrong invoice posted to an ERP costs a reconciliation nightmare. Cheap-error tasks can run fully autonomous from day one. Expensive-error tasks need the trust-ladder approach I used on Emmy: observe, then organize, then draft, and only then act — with logging and undo at every rung.
Questions I ask before every automation project
- 01How many times per week does this actually happen? (Count it — people overestimate 3x.)
- 02What's the true cost of one bad run, including cleanup and lost trust?
- 03Is the underlying process stable, or does it change monthly? Automating a moving target doubles maintenance.
- 04Who owns the automation when it breaks at 2am — and will they know it broke?
- 05Is there an official API, or are we driving a UI? UI automation works but budget for fragility.
Start with the boring win
Every engagement I've done that succeeded started with one high-frequency, low-risk workflow automated end-to-end — visible time savings inside two weeks. That earns the trust to tackle the ambitious stuff. Start with the flashy AI demo instead, and you'll spend the quarter explaining edge cases.
If you're sizing up an automation and want a second opinion on which quadrant it lives in, reach out — happy to talk it through.
Haider Farooq is an AI engineer and data scientist based in Lahore, Pakistan — core engineer on TryCook.ai, developer at Aligno, and creator of MarkSafe.net. He builds agentic AI systems, RAG pipelines, and automation for teams worldwide. Work with him.
RELATED WORK