Discussion + Career
Source: Dev.to
AI Promised to Automate the Routine. Instead, It Added a New One.
We were told AI would free us from repetitive work. Instead, many developers now spend hours managing AI tools, reviewing AI output, and debugging AI‑generated code. The routine didn’t disappear — it shape‑shifted.
A Typical Day in 2026
| Time | Activity | Duration |
|---|---|---|
| Morning | Review overnight AI agent outputs | 30 min |
| Mid‑morning | Fix what the AI broke or missed | 45 min |
| Afternoon | Write prompts for tomorrow’s AI tasks | 30 min |
| Evening | Check monitoring dashboards for anomalies | 15 min |
That’s 2 hours/day managing AI. Before AI, I spent about the same amount of time on the manual tasks it replaced. Net productivity gain? Roughly zero for routine tasks.
Where the Real Gains Appear
Social posting
Before: 3 hours to post on 19 platforms manually.
Savings: 3 hours/day.Research & summarisation
Before: 2 days to research and summarise 29 platforms.
Savings: 12 hours/project.Service setup
Before: 2 hours to set up a new service with tests, Docker, CI.
Savings: 1.5 hours/project.
New Challenges Introduced by AI
- Subtle bugs – AI‑generated code often looks correct but contains hidden errors. Debugging someone else’s logic can take longer than writing the code yourself.
- Context‑window limits – Large projects exceed model context windows, forcing you to spend time summarising, chunking, and re‑explaining context to the AI.
- Prompt fragility – A prompt that works today may break after a model update, turning prompt maintenance into a new form of technical debt.
- Overhead – Managing AI outputs, monitoring dashboards, and building glue code adds its own workload.
Strategies for Benefiting from AI
- Automate the automation – Build systems that manage AI outputs automatically, reducing manual review.
- Accept imperfection – A slightly off image is better than no post at all.
- Know when NOT to use AI – Writing 20 lines of Python can be faster than crafting a prompt for an LLM.
Decision Checklist Before Adding AI
- Is the task truly repetitive? (Identical each time, not just boring.)
- Is the output verifiable? (Can you automatically check correctness?)
- Is the cost of failure low? (A wrong social post is fixable; a wrong database migration isn’t.)
If all three answers are yes, automate. If any is no, think twice.
Practical Recommendations
- Start with the highest‑volume, lowest‑risk tasks.
- Build monitoring before building agents.
- Prefer JSON state files over databases for agent state.
- Free‑tier LLMs are sufficient for ~80 % of tasks.
- An agent that runs reliably at 80 % quality beats one that runs perfectly only 60 % of the time.
Conclusion
The AI revolution is real, but it’s an evolution, not a replacement. The routine changed—it didn’t disappear.
Thoughts? Let’s discuss at sborka.work.