How much AI your company relies on and how you deal with it?

Posted by HalfSarcastic@reddit | ExperiencedDevs | View on Reddit | 16 comments

I'm out of commercial development circle for almost two years. And I am not celebrating ai as many do. I find it helpful as a much as a highly contextualized search engine that can combine information from somewhat reputable sources and generate a decent quality material with specific focus. Which still needs to examined, verified, tested and carefully used.

However I see that many use it as if it's the smartest person in the room and everything it says should be taken with the highest praise and admiration.

So now I'm quite anxious about the idea of returning to the commercial software development and prospects of having to deal with all that AI worshipping.

In general - please share how it is going in your company and how does AI push affects you and your daily routines.

Regardless if it is positive or negative experience.

I hope to see no judgment in this thread.

[-]

symbiatch@reddit

Rely? None. People are relied on. Used? Somewhat, depending on position and task and project and team and person.

I barely use it. They know if they push it too much on me they won’t like the result. And then they have people that flip from “we must move fast with AI” to “we must be really careful and make sure we use it where it works” in the same sentence.

YahenP@reddit

In short, we've devalued our work. We've devalued our experience. In management parlance, this is called increasing productivity. But no one needs it. We haven't made any more money, we haven't had any more clients, and our end users aren't thrilled either.

Goingone@reddit

Every engineer is still responsible for the code they commit.

Currently, we do not have any engineers who are comfortable with committing AI generated code without reviewing it (given they are on the hook if things go haywire).

We still see some efficiency gains, but until fully autonomous development seems reasonable, efficiency gains will always be limited.

Biggest change is the job has a much larger “code review” focus these days, and less keystrokes.

Zeragamba@reddit

how are you measuring efficiency gains? As it may just be shifting where the work is done

Nothing formal, only subjective.

And I should note, it’s at a “task” level. Certain tasks that used to take 15 minutes, might take 5 now.

Too many moving pieces to objectively quantify an exact increase in development velocity.

Impossible_Stock8885@reddit

My company recently got the Claude plan for everyone and at first it sucked.

Lots of vibe coded garbage, code comments had the AI "checkbox" icon bullshit and other weird emojis... Var namings were always single chars for some reason.

But down the road it got a lot better. I started using AI to review AI code from colleagues. I went from working 9 hours a day to 30 mins a day and started using the free time to destress.

Difficult-Vacation-5@reddit

For real?

originalchronoguy@reddit

It still requires skills. Regardless.

I am building agentic workflows people at my org never even dreamt or thought of. And those systems come with deep years of experience on how I leverage AI. I have MCP servers that run that measures API performance. Another that does multiple attack vectors that create repeatable automated scripts to hijack someone's backend service via black box means. Both were design with specific input in mind.

Having someone randomly ask a LLM "Build me an attack system that finds flaws on any URL I give it. Make no mistakes" isn't going to produce anything meaningful. I am simply "rebuilding" things I did in the past on previous projects; working with other teams and members. And giving context on how those systems are built.

The people who don't have those skills won't be able to build those things with or without AI assistance.

EyesOfAzula@reddit

AI got A LOT better in the last 2 years. I remember the copilot days in 2023, terrible.

AI is great today, just need to keep an eye on it and correct it because it will make slop if you don't keep it tight.

You can move so much faster and now most of the time you are reviewing plan mode plans the AI makes, then code reviewing and testing changes the AI makes.

Altruistic-Bat-9070@reddit

I recently moved to a company that use it significantly more but, at least in the team i am in, significantly better.

The company has a lot of procedures and policies around AI use and funnels users into using specific products (github copilot for coding, microsoft copilot for documents which now has claude in it).

My previous workplace wanted everyone using it but had no policies and paid for no licenses so everyone was just putting company data in the free tiers of various models and using it to prpduce things blindly. That definately happens where i am not but the company has enough control over the use that it will remove peoples access if they are considered to be relying on it too heavily and not using their brains or validating the output.

The company I am at now doesn’t do this all perfectly but it’s the right idea and generally i think these policies mixed with AI generally improving over time means this will naturally settle in a good place.

CW-Eight@reddit

I’m now in an AI startup, formally 20+ years FAANG.

It is just an entirely new game. I’m a bot herder now.

The platform stuff I fully understand, but use AI to help analyze, help change, help review.

Prompts I tend to hand-write V0 and then use AI to iterate: analyze, test, post-mortem evaluate, and revise.

But for the UI I just let the bots run free- code, test, and CR. Yes, it produces crap code, but it is well tested crap code, and it works. We will toss it out within a year, before all the crap piles up so high it is totally unmanageable. If we don’t move fast, we die as a company. The key here is super tight design specs and pitting the test bots against the coding bots.

NickW1343@reddit

We get a 20 dollar Claude sub and I've never heard of any of my co-workers say they got rate limited, so we have it, but don't use it a lot. Our use case outside of some exit interview AI app is mostly just brainstorming with it and doing analysis before maybe having it make a change. I like using it to generate code, but 90% of my prompts are me asking it questions I have about the codebases then maybe changing some code if needed. If there's a code change, then it's more questions about why and manual work figuring out if the methods it changed were referenced somewhere else that wasn't a problem which could cause a new bug to be shipped.

airemy_lin@reddit

We ramped up like crazy. I don’t know anyone pushing vibe code to consumer facing production envs with no human verification, but it is heavily being used by the team.

AcksYouaSyn@reddit

It is pervasive. The entire development lifecycle has been upended. We have token usage leaderboards and just laid off the bottom 80% of the engineering department. The remaining dev and qa are expected to practice “harness engineering” to be 10x more productive by using AI to close out 3-5 tickets in parallel per day.

Early_Rooster7579@reddit

AI use is very heavy at meta. Every team I know of is using it extensively.

Long_Egg_8644@reddit

In my experience the healthiest teams treat AI like a power tool, not an authority figure.

It’s great for boilerplate, debugging hints, summaries, quick research, etc. But good engineers still verify everything and own the final decisions. The worrying environments are the ones optimizing for “tokens burned” or raw output volume instead of maintainability and understanding.