Key takeaways
- Automate a task when its repeated cost and coordination burden exceed the full cost of building and owning the automation.
- A short task can be valuable to automate when it blocks several people or scales with every customer, order, or employee.
- Stable inputs and checkable outputs matter more than how annoying the task feels.
- Start in draft or observation mode when the action is hard to reverse.
- Count maintenance, exception handling, and review time in the break-even calculation.
The engineer joke is a real break-even problem
The two xkcd references capture both halves. “Is It Worth the Time?” asks how much build effort a repeated saving can repay. “Automation” adds debugging and ongoing development while the original task keeps happening.
The answer is not “never automate small work.” It is to evaluate the complete internal tool. A ten-minute task performed daily by one person is different from the same ten-minute handoff that blocks four people, creates corrections, and repeats for every customer.
Use six tests before calculating ROI
| Test | Promising signal | Warning |
|---|---|---|
| Frequency | Weekly or more, or scales with business volume | Rare, shrinking, or event-specific |
| Stability | Normal path and major exceptions repeat | People renegotiate the process every run |
| Input readiness | Required facts have named sources and owners | Someone hunts across messages and memory |
| Checkability | Outcome and external effects can be verified promptly | Success is subjective or visible months later |
| Coordination | The task creates queueing, handoffs, rework, or missed work | One person finishes it privately with no downstream wait |
| Containment | Errors are reversible or can stop at a gate | One wrong run creates immediate material harm |
Google's toil guidance recommends measuring and comparing sources of repetitive work, then ensuring time saved is at least proportional to development and maintenance effort. Irritation can help you notice a candidate; it cannot justify one.
Calculate the whole break-even threshold
For one period, estimate manual cost as runs multiplied by full handling time and loaded hourly cost. Estimate saved cost as covered runs multiplied by manual time minus residual review time. Then subtract recurring software, model, support, monitoring, exception, and maintenance costs. Treat initial discovery and build as an investment to repay from the net monthly benefit.
Use the automation ROI calculator to test the threshold with your numbers. Run conservative coverage and maintenance cases rather than choosing one optimistic forecast.
Automate part of the task when the whole task is unsafe
| Pattern | Software does | Person does |
|---|---|---|
| Prepare | Collect, validate, and summarize required facts | Judge the case |
| Draft | Create a proposed classification, message, or update | Review, edit, and authorize |
| Gate | Run settled rules and present decision evidence | Approve the consequential effect |
| Monitor | Detect missing, overdue, or inconsistent state | Resolve the novel exception |
| Record | Write the decision, actor, time, and outcome | Own the policy and correction |
Make one of four explicit decisions
- 01Keep manual and documentHuman approvalUse this when work is rare, unstable, or hard to check. Name the owner and revisit only when frequency or stability changes.
- 02Batch or simplifyCodeRemove fields, handoffs, and duplicate entry before automating. Sometimes eliminating the work beats moving it faster.
- 03Assist with a gateAI judgmentAutomate collection or drafting while a person verifies and authorizes the effect.
- 04Automate a bounded pathTriggerStart with stable, checkable, reversible cases; route every unhandled case to an owned queue.
Limitations and when not to use this
- The five-year xkcd table is a useful intuition, not a business case. It does not include labor rates, coordination, risk, software, maintenance, or opportunity cost.
- Savings estimates are not guarantees. Measure real coverage, residual review, corrections, and maintenance after release.
- Do not automate a harmful or regulated action solely because the time calculation is positive. Apply the relevant risk, security, and professional review.
Sources
- Is It Worth the Time? — xkcd Accessed 14 August 2026
- Automation — xkcd Accessed 14 August 2026
- Eliminating Toil — Google, The Site Reliability Workbook Accessed 14 August 2026
Calculate the break-even point
Use the real run frequency, full manual time, expected coverage, residual review time, and monthly ownership cost.
Calculate the break-even pointUli Prantz
Builds and operates all-agents
Uli Prantz builds all-agents, the process-automation platform this site documents. He writes about the operational side of automating recurring business work: where deterministic code beats model judgment, where it does not, and where a human still has to approve.