Which work is worth automating with AI?
Start with repeatable work whose result is checkable, and include supervision and correction before calling it a saving.
On this page
Direct answer
Good automation candidates have recurring inputs, a relatively stable process and inspectable outputs. Delegate a segment you already understand, then expand gradually. An AI role needs duties, process, permissions and feedback; not every browser task should become automatic. Include setup, maintenance, human review and the cost of errors. One successful demonstration is not evidence for unattended operation.\n\nPractical suggestion: record remaining human time separately. Watching AI click continuously may save no work at all. It can remain a learning prototype without being described as unattended.
When to use
- You repeatedly research, prepare tables, draft mail or generate reports.
- Existing automation requires more watching and repair than doing the job yourself.
Step by step
List actual recurring work
Review recent work and record each input, action, duration, frequency and recipient. Choose a small clear task such as summarizing new orders into a review table rather than asking AI to run a company.
Separate judgment from mechanical execution
Use programs for rule-based steps and models for language understanding or synthesis. Keep exceptions affecting goals, budgets or external commitments with a person. If judgment criteria are inconsistent, establish them before automating.
Choose a reversible, measurable segment
Begin with organization, retrieval, classification or drafts. Define separate authorization for sending, deletion, charges and bulk changes. Verify complete results offline before connecting real business state.
Run once under observation
Use real inputs with appropriate sanitization, inspect step evidence and outputs, and classify failures as input, process, tool or missing acceptance. Repair that cause instead of covering instability with more prompting.
Compare full costs and decide whether to continue
Compare manual effort with setup, operation, review and rework. Increase frequency or scope only after stability. Frequent changes may make maintenance uneconomical; keep a semi-automatic version or stop when appropriate.
Write an input, action and delivery card for the first role
Specify trigger, allowed material, recipient, editable content and decisions requiring a person. A hypothetical stock assistant may only draft a purchase order: read authorized SKU inventory → apply approved replenishment rules → check suppliers and lead times → propose currency, unit price, quantity and total. Mark missing data for confirmation. A completed draft does not authorize placing an order.
Calculate weekly cost before expanding
Hypothetical calculation: 20 weekly runs at 8 manual minutes give a 160-minute baseline. Review at 2 minutes each costs 40 minutes; maintenance costs 20; setup of 120 minutes spread over four weeks adds 30. Total effort is 90 minutes, saving 70. Track model charges and error losses separately; do not subtract money from minutes. An extra 80 minutes of rework makes the week 10 minutes worse than manual work. Fix the exception or return to partial automation.
Start with a low-impact work-log reflection
Put only selected, sanitized personal records into a dedicated input directory. Test sanitization on synthetic samples, then manually inspect the actual text sent to the model. Require five outputs: evidenced observations, tentative obstacle, one smallest action, one useful question and missing evidence. Missing logs mean insufficient information, not a personality inference or psychological diagnosis. Generate locally once; add reminders and a pause control only after it proves useful.
Separate deployment from role-level acceptance
Keep evidence for account connection, reading the specified input, invoking the method, generating a file, content acceptance and delivery to the intended location. Observe multiple runs for failures, manual takeovers and actual adoption before expanding. A fixed count of runs or days is only a practice cadence. Personal use does not require publishing raw logs, profiles or repositories.
Worked example
Hypothetical exercise: weekly competitor updates
Read specified public update pages, classify changes and attach evidence; a person decides whether to act. Automatically changing prices, buying ads and publishing announcements crosses additional decision boundaries and can raise review cost.
Action checklist
Reusable templates
Automation opportunity screen
Recurring task:
Input provenance and scope:
Frequency and current manual time:
Stable steps:
Exceptions needing judgment:
Actual consequences of error:
Expected output and checking method:
Allowed automatic actions:
Setup, operation, review and maintenance cost:
Segment to test in the first round:
AI role definition and first run
Role and recipient:
Trigger:
Allowed material and time range:
Identity and allowed tools:
Process: input → judgment → draft → check → delivery
Single output and storage location:
Exceptions a person must decide:
Per-run and cumulative quantity or cost limits:
Duplicate, interruption and unknown-result handling:
First input, actual output and unfinished work:
Weekly time saving and maintenance cost:
Personal work-log reflection
Read only this sanitized file or directory:
Date range and missing periods:
Current work goal:
Exclude other people’s statements, identities, customer records and credentials.
Output:
1. Observation: what happened and which log passage supports it?
2. Obstacle: facts versus tentative explanations.
3. Smallest action: what next, and how will completion be judged?
4. Question: one question useful for action.
5. Evidence gap: what cannot be determined?
No personality diagnosis, shaming or reconstruction of sanitized content.
Which suggestion I accepted or rejected, and why:
Next run, pausing and record deletion:
Pitfalls
- Capability to operate does not establish that automation is worthwhile.
- Average success rates can conceal rare severe errors.
- Improve an undefined manual process before automating it.
- Repetition is a discovery cue, not a numerical threshold that proves automation is worthwhile. A simple job does not need every agent framework installed.
- A sanitization regex cannot detect every sensitive detail. Test actual false negatives and false positives instead of trusting a promised rule count.