McKinsey wonders how to sell AI apps with no measurable benefits

> The incentive structure for managers (and literally everyone up the chain) is to maximize headcount. More people you managed, the more power you have within the organization

Ding ding ding!

AI can absolutely reduce headcount. It already could 2 years ago, when we were just getting started. At the time I worked at a company that did just that, succesfully automating away thousands of jobs which couldn't pre-LLMs. The reason it ""worked"" was because it was outsourced headcount, so there was very limited political incentive to keep them if they were replaceable.

The bigger and older the company, the more ossified the structures are that have a want to keep headcount equal, and ideally grow it. This is by far the biggest cause of all these "failed" AI projects. It's super obvious when you start noticing that for jobs that were being outsourced, or done by temp/contracted workers, those are much more rapidly being replaced. As well as the fact that tech startups are hiring much less than before. Not talking about YC-and-co startups here, those are global exceptions indeed affected a lot by ZIRP and what not. I'm talking about the 99.9% of startups that don't get big VC funds.

A lot of the narrative on HN that it isn't happening and AI is all a scam is IMO out of reasonable fear.

If you're still not convinced, think about it this way. Before LLMs were a thing, if I asked you what the success rate of software projects at non-tech companies was, what would you have said? 90% failure rate? To my knowledge, the numbers are indeed close. And what's the biggest reason? Almost never "this problem cannot be technically solved". You'd probably name other, more common reasons.

Why would this be any different for AI? Why would those same reasons suddenly disappear? They don't. All the politics, all the enterprise salesmen, the lack of understanding of actual needs, the personal KPIs to hit - they're all still there. And the politics are even worse than with trad. enterprise software now that the premise of headcount reduction looms larger than ever.

I am curious if the timing have impacted the inability to measure a benefit. AI is rolling out at the same time as widespread return to office campaigns. Remote work was widely studied and touted as improving efficiency, but no one is showing the drop for RTO. Is AI in part just balancing it out? There's also an ongoing massive brain drain. Many companies are either laying off their most tenured and competent employees, or they are making life miserable for them in the hopes that they quit.

All of this said, using AI in your back end takes a huge amount of time from your users and employees. You have to vary multiple prompts, you have to make the output sane, touch it up, etc. The most useful part of AI for me has been using it to learn something new, or push through a task that I otherwise couldn't do. I was able to partially rewrite a logging window to reduce CPU use significantly. It took me over two weeks of back and forth with AI to figure out a workable solution and implement it into the software. I competent programmer probably could have done it better than I did in less than an hour. There's no business benefit to a help desk person being able to spend 2 weeks writing code that an engineer would be much better suited to handling. But maybe that engineer could write it in 10 minutes instead of an hour if they used AI to understand the software first.

It's not the top IME, but the big fat middle of the org chart (company age seems to mirror physical age maybe?) where middle to senior managers can hide out, deliver little demonstratable value and ride with the tides. Some of these people are far better at surfing the waves than they are at performing the tasks of their job title, and they will outlast you, both your political skills and your tolerance for BS.

Except that the people that make robots for factories, aren't interested in making whatever that factory is making.

This is because they're trying to reduce the wrong headcount. The largest inefficiencies in corpo orgs lie in the ways they organize their knowledge and information stores, and in how they manage decision making.

The rank and file generally have a really good grasp on their subset of the domain -- they have expertise and experience, as well as local context. Small teams, their managers -- those are the ones who actually perform, and deliver value.

As you move up the hierarchy, access to information does not scale. People in the middle are generally mediocre performers, buried in process, ritual and politic. In addition to these burdens, the information systems do their best to obscure knowledge, with the usual excuses of Safe and Secure (tm) -- things are siloed, search does not work, archives are sunsetted, etc.

In some orgs tribalism also plays an outsized role, with teams acting competitive, which largely results in wasted resources and seven versions of the same failed attempt at New Shiny Thing.

Then as we look higher yet in the hierarchy, the so-called decision makers don't really do anything that cannot be described as "maximize profit" or "cut costs", all while fighting not to get pulled down by the Lord of the Flies shenanigans of their underlings. They are the most replaceable.

A successful "AI Transformation" would come in top-down, going after the most expensive headcount first. Only truly valuable contributors would remain at that level. Organizational knowledge bases would allow to search, analyze and reason about the institutional knowledge accrued in corporate archives over the years, enabling much more effective decision making. Meanwhile, the ICs would benefit from the AI boost, outsourcing some menial tasks to the machine, with the dual benefit of levelling up their roles, and feeding the machine more context about the lower-level work done across the org.