AI, to me, seems perfectly positioned to replace a lot of middle managers. Maybe not "every manager at X level", but if you've got some central tool that flows information up and down the chain, and can dig into data on your local network for backup/proof, you could make most middle managers much more efficient. Suddenly, one scrum master covering 4 teams is very feasible. They already know what they need to know, and their status has been reported long before they even started the meeting.

I'm pretty sure AI would advantageously replace high level executive and McKinsey-type consulting (or most consulting even), but middle manager i'm not too sure. Depends on your company and what their job is.

I could see a world where a manager is able to cover more teams, reducing the need for as many layers / middle managers.

Like… for year end reviews and bonuses, a manager might have a dashboard of feedback from an AI of everyone’s technical accomplishments, as AI can review their PRs and designs. At that point, an AI could even rank engineers based on contributions, so the manager just verifies that the feedback is all correct.

Not saying that’s necessarily a good thing, but I could see a world where that happens. It already does seem to with Amazon warehouse workers.

Isn't that basically the doge story once the agencies investigating Tesla were shut down? One of the people policy said they didn't think anybody should actually be using their code

The Scrum lords and managers might use LLMs to add beurocratic complexity instead of becoming efficient.

Alternatively, it could become more efficient at creating inefficiencies.

Imagine pushing your vibe-coded changes triggering a CI hook that uses an LLM to create a TPS report per commit+pipeline in a merge request, which another LLM-based bureaucrat could use to decide whether the MR should be approved, generating its own TPS reports in turn which some middle manager could use their LLM to summarize.

Think of all the shareholder value we could create by going all-in on AI!

The number of vibe coded PRs clogging up the review queues is going up. When you ask the submitter what some nonsense line of code is doing and what it's for, they have no idea. Slowing the whole process down.

It definitely takes me more time to review PRs now, and I have more PRs to review. Agent created code seems “fine” but also seems like 30% more than is needed (total SWAG)

The worst part is how it destroys your implicit trust that the author ran and understood the code. The generated code looks fine but doesn't actually work because it was never run. So to properly review PRs today you have to spend a lot longer assuming nothing works and every detail needs to be carefully tested.

Is there’s no screenshots of it working or logs from a console showing it working, I kinda presume it won’t. In a previous role that was more front end facing (tablet experience for self driving cars), I had a few coworkers who would put out PRs that I knew could not have worked and were not tested. My nudge was “oh can you add a screenshot to the PR?” And they would realize it didn’t work and get back to me in a few days. Been dealing with this since before gen AI produced code haha.

Most of these are solved by having better leadership, and unfortunately for job seekers, also solved by hiring fewer, higher-caliber people in the first place.

> ...higher-caliber people in the first place.

This is the idea. However, many of these "higher-caliber" people are going to burn out because productivity expectations, especially in software development, are rising insanely due to AI tools.

Ok but finding good leadership is hard. Using an AI api is easy

I agree on the cause of waste, but I'm skeptical AI will move the needle on a solution. The longer I work in big companies the more I realize that things happen because someone has the strength of will combined with the right levers and incentives to make it happen despite organizational inertia. AI is potentially a powerful tool when judiciously used by those who understand what they're doing, but it can also be an amplifier of noise that filibusters real progress when hands-on leadership and good judgement is in short supply.

A lot of what I'm seeing feels like LLMs are "perfect" bridge-to-nowhere enablers. They are the ultimate Yes Men and give upper management an impression that nowhere not only needs plenty more bridges but that the bridges are cheap to build and easy to use and may already have been built with magic strong enough to attract leprechauns and their pots of gold to the end of the bridges.

Too much of my career has been in the "Legacy Code" mines, and the LLMs seem like the worst pipeline for expanding "Legacy Code" faster than ever. Some of them may as well be called "Legacy Code-as-a-Service". I'm not worried that my "Legacy Code" skills are going to get out-of-date, I'm worried they are going to be in higher, uglier demand when some of these bridge-to-nowhere factories finally get shutdown or at least slowed down for some of the real maintenance costs to finally show up on companies' bottom lines.

True. A lot of work done in big companies is waste.

6. A lot of unneeded complexity is added.

Perhaps if we can only use AI to understand the wasted effort.

Also, I rather have the ICs with some slack and time for clear thinking to address the big issues than always be busy on a mountain of work that goes now where.

This is the big one for me. I recently needed to solve a fairly simple issue; in 3 days I had a solution in FastAPI and SQLAlchemy working with a React front end. Under 1,000 lines of code between the two.

Of course we can’t spend $20/month on compute for that, so it’s now a web hook jankily jammed into another app that writes data to Jira with an Airflow job that triggers Databricks to run a notebook that uses Spark to write the data to S3.

All that to replace a Lambda that would never have gotten to triple digit queries per day. No, instead we spent a month of engineering time to strap all this together and we probably spend 4x as much on compute because it is both slower and requires more resources.

And now I’m getting complaints that it’s complicated and hard to troubleshoot. Yeah, of course, we’ve duct taped together products whose combined manuals are thousands of pages. Color me shocked it’s hard to use.

2. Has me damn ready to leave the industry if I’m not forced out.

You forgot “meetings” :)