I keep relearning the same lesson. I add an AI step. It works. And I get slower.
Not broken-slower. Working-slower. The step runs, produces output, passes every check I wrote for it. Then I sit there reading what it made, deciding if it’s right, fixing the parts that aren’t, and half the time sending the version I would have written anyway. The AI didn’t remove work. It added a thing to verify.
Why does adding AI sometimes make a workflow slower?
Adding AI makes a workflow slower when the AI step adds a decision to check instead of removing one. A human now has to generate the output, read it, judge whether it’s correct, correct it, and sometimes redo it from scratch. That’s five actions where before there was one. The step “works” the whole time, which is why it hides.
Think about the two-minute email. It used to be: open, type, send. Two minutes. Now it’s: prompt, wait, read the draft, notice it’s 20% too formal, delete a sentence, add the one line that actually mattered, decide the whole thing is off, and send my original. Four minutes. The tool ran perfectly. I lost two minutes.
Same shape with meeting summaries. The summary generates fine. But the only people who can confirm it’s accurate are the people who were in the meeting. The exact people who no longer need a summary. So they read it to check it. Reading-to-check is not the same as being saved the reading.
Why does it feel faster even when it isn’t?
It feels faster because generating output feels like progress while judging output feels like nothing. The AI hands you a full draft in seconds, so the blank page is gone and your brain registers a win. But the slow part, reading and deciding and correcting, happens after the dopamine, and you don’t count it. Felt speed and wall-clock speed come apart, and felt speed wins the memory.
I trust this gap the least in myself. I have finished a task convinced I was quicker, then looked at the clock and been wrong. The confidence is real. The time is worse. If your only evidence is how it felt, assume you’re the one being fooled.
How do I actually tell if an AI step helped?
Measure wall-clock time end-to-end, including the human review and correction, not just whether the step ran. The question is never “did the AI produce something.” The question is “from the start of the task to the moment it’s actually done and correct, did the total time go down.” Time the human’s part. That’s where the cost hides.
So I do this. I pick one real task. I run it the old way and note the minutes. I run it with the AI step and note the minutes, the whole thing, including me re-reading and fixing. Then I compare. Not the demo. Not the happy path. The version where the output was slightly wrong, because that’s most days.
Most of my regrets pass the happy-path test and fail the clock. The step does exactly what I built it to do. It just moved the work from “writing” to “checking,” and checking turned out to be slower than writing for that particular task. You only see it if you’re honest about counting the second half.
When is an AI step actually worth keeping?
Keep the AI step only where it removed a decision, not where it added one to verify. It’s worth keeping when it turns a blank page into a yes/no, kills a lookup, or collapses five tabs into one answer. Rip it out when its only output is a draft you now have to sit in judgment of.
The keepers in my own setup all have the same shape. They took something with no obvious starting point and gave me a single thing to accept or reject. Yes or no is fast. Editing a paragraph is slow. A classification I can trust and move past is a decision removed. A draft I have to read word by word is a decision added, wearing a helpful costume.
Some tests I use before a step earns its place:
- Did it delete a lookup I used to do by hand, or did it hand me something to fact-check?
- Can I accept the output with a glance, or do I have to read every line to trust it?
- When it’s wrong, is fixing it faster than doing the task from zero, honestly?
- Did the number of tabs, apps, or context-switches go down?
If the answer is “it gives me a draft to review,” that’s a warning, not a feature. Review is work. Sometimes the review is worth it. But you have to call it work and count it, not file it under “the AI did that for me.”
What do I do when a step fails the clock?
Rip it out. This is the part I resist, because I built the thing and it runs and ripping it out feels like admitting I wasted the build. But a step that runs and costs time is worse than no step. It costs the compute, the wait, and my attention, and returns a draft I override. Cutting it is the win, not the loss.
I don’t rip out the whole system. I rip out the one step and keep the ones that passed. What’s left is smaller and faster than either the all-manual version or the all-AI version. That’s the actual goal. Not “AI-powered.” Just faster, with the AI only where it earns the slot.
The honest version
This isn’t one clean war story where I learned the lesson and never repeated it. It’s a pattern I keep walking back into. I add the step, it works, I feel faster, and weeks later I time it and find I’ve been losing minutes the whole time. So now the timing is the habit, not the AI.
Before you add an AI step, time the task without it. After, time it again with the human review counted. If the number didn’t drop, the step didn’t help, no matter how good the demo felt.