What do you still do manually, even though AI could do it?
AI can draft a reply, build a small app, or run a routine task. But being able to automate something does not always mean you want to hand it over.
Sometimes doing it yourself is how you notice a problem or understand what someone actually needs.
A few examples:
Replying to a frustrated customer.
Checking an app before sharing it.
Deciding what to build next.
I am building Inquir Compute for apps, databases and background jobs, so I am curious about where people draw that line.
What is one task you deliberately keep manual, and what would need to change before you trusted AI with it?
152 views
Replies
Reading angry emails from users. AI can write a poilite reply but doing it myself is the only way to actually feel where e messed up .
Do you keep the original email attached when that complaint becomes a task?
A summary can lose the exact step that frustrated someone. I would want the person fixing it to see that detail too
Deciding product roadmap stays manual because context lives in human conversations not log files.
Has a conversation ever made you drop a feature that looked important in the data?
That would be a useful example of the context a roadmap summary misses
I use AI to write complex code but i still manually format my markdown files. some habbits just feel comforting.
Is it the formatting itself or the chance to read through the file while doing it?
If it is the second one then automating it also removes a small review step
Delegating execution is fine but delegating empathy is dangerous. That is where i draw the hard line with tools like inquir compute.
Inquir is the platform where the app runs so that choice stays with whoever builds it
For a support app I would want a clear handoff to a person when someone asks for help beyond what the automation can handle. Where would you put that handoff?
download payment evidences from the bank account my accountant ask for ))
Typing the card details for every payment test. An agent could do it in a second, but doing it myself is how I notice when checkout feels off.
Honestly, yes and no. I’ve been through this with long agent runs. The output would look fine, but one step quietly assumed something the previous step never established, so reading the final answer manually wasn't enough. What I ended up doing was leaving decision points and anything external-facing for review, then giving repeat work a tight brief and an explicit owner. It makes setup slower and the review load is annoying, but it stops automation from hiding a missing decision. I'm still figuring out which checks should become automatic. What would you keep manual in the first version?
For a first version I would review anything that makes a promise to someone else or changes their data
At each handoff I would also ask for the source of any assumption the next step relies on. Missing evidence should stop the run before it reaches a polished final answer
Replying to a frustrated customer is mine too, and for the same reason Lily mentioned — you only find out where things actually broke by reading the complaint in the person's own words. What I'd want before handing it over is the original message kept right next to the reply, not a summary of it, because the exact phrasing is usually the clue. The second one for me is deciding what to build next: I'm happy to let AI sort and rank everything, but the final call still needs to come from me after I've seen the raw list myself.
The final call on a candidate after a tech interview. I've done 500+ interviews, and the score was never the part I trusted most.
What decides it is how someone behaves when they're stuck: do they say "I don't know, here's how I'd find out" or do they bluff. A transcript summary flattens exactly that. It reads the same either way.
What would need to change: an AI that says "I'm not sure about this one" as often as a good interviewer does. Right now it gives a confident verdict on every candidate, and that's the one thing I can't hand over
A developer friend tried fully automating client release notes last month and it miscommunicated a breaking change. we went straight back to manual review.
Was the breaking change missing from the source material or softened when AI rewrote it?
I would review what the user needs to change separately from the prose. A readable announcement can still leave out the action they need to take