Recently, I compared notes with colleagues who work in several different Fortune 500 companies.  These enterprises share the following traits:

• They are big, multi-billion-dollar enterprises with over 75,000 employees each.  They’re all global companies.
• They are not IT companies but do use IT intensively (as all big business do).  They’re in traditional industries that have existed for a long time.
• They have margins that run from merely tight to razor-thin.  Cost containment is very important to these companies.  They’re not the Microsofts or Googles or Disneys or JP Morgans that lavish money on their operations but rather companies that count where every dollar goes.
• Each of these companies has large IT departments, which range from people who fix printers in the field to senior software architects.

In each company, they are introducing and adopting AI.  And in every case, these cost-conscious enterprises are pulling back.  And IT pros at each company tells the same story:

• They were given Copilot because the company has an enterprise deal with Microsoft.
• It’s okay…ish…for boilerplate code in Visual Studio, but what they want is the use better models and agents.
• Management has said “no, use Copilot” because of the cost of tokens.
• Management has been skeptical of deploying agents, unless a clear cost reduction can be shown.
• Since it can’t, agentic deployments have been very limited.  In one case, the only agent in use at the massive enterprise is a password reset bot.

Shouldn’t AI be radically slashing costs?

It’s hard to understand the cost benefits when you look at IT-driven enterprises like Google or Microsoft where they’re not paying market prices for AI and they are very focused on new products and new capabilities.  In the case of the above-listed enterprises, their products and services don’t change year-over-year.  Their only means to increase profit is to win new customers or reduce cost.

In economics terms, the (simplified) cost equation is:

Cost = Labor + Capital (machines, buildings, etc.)

What should happen is this:

Cost = AI * Labor + Capital (machines, buildings, etc.)

In other words, AI should act as factor on the cost of labor.  The AI salesmen would say that AI should be a number less than one (e.g., .8) so that labor is reduced.  But what actually happens is:

Cost = AI * Labor + Capital (machines, buildings, etc.) + Cost of AI

And Cost of AI is a larger additive than the AI factor on labor.  In other words, AI is an additional cost input, not a substitute for labor.  And the factor on labor in some cases is not less than 1.

Why is that?  Several reasons:

The Benefit AI Provides is Not that Significant.  How can this be?  Because humans remain in the loop – in reviewing AI-generated marketing reports, in checking over AI-generated code, etc.  Not only is there a review loop, but often a re-do/fix loop.  Imagine you work as a short-order cook in a restaurant and management hires another cook.  The second cook needs to be trained, reducing your overall output, and then once trained, you constantly have to check his work.  Sure, some things such as “butter the toast” and “put garnish on the steak plate” are things you can trust him with, but you’re still needed to make the main dishes.  That’s often what working with AI feels like.  In other words, AI creates work, too.

Reducing Minutes is Not Reducing Headcount. Microsoft has trumpeted various Copilot metrics to say that employees at such-and-such a firm have saved 46 minutes a week due to AI, or certain employees save an average of 2 hours per week.  But that doesn’t add up linearly and neatly.  Just because I can now saving a 45 minutes a week doesn’t mean that if you have 53 employees, you’ve saved 40 hours and can eliminate one.  (Even if you could, that’s a pretty paltry return on what you pay for tokens).  In very few enterprises can labor be recovered that efficiently.

It’s Called Generative AI: Sure, there’s a cost benefit in using AI to design a marketing banner instead of hiring a human artist.  But no one today is saying “lay off the accounting department, and we’ll let ChatGPT handle our books from now on.”  Shouldn’t that be the sort of thing AI is great at?  It can suck up all the arcane tax laws, all the public filing regulations, a few accounting textbooks, etc. and then handle your invoicing, billing, and accounts receivable…right?  No one thinks this.  AI in its present form is called generative AI for a reason.  It excels at things like producing a dozen different logo designs for you to choose from.  But if you ask it to handle a customer workflow the same way a dozen times in a row, it falters.

From the perspective of economic theory, the prediction that AI would increase profitability depends not merely on making individual tasks faster, but on lowering the firm’s overall cost of producing its services. If AI speeds up isolated tasks yet fails to reduce labor inputs, expand output, or remove production bottlenecks (and instead introduces a significant new variable cost in the form of token consumption) then the production function has not shifted in an economically meaningful way.

raindog308

raindog308 is a longtime community LETizen, technical writer, and self-described techno polymath. With deep roots in the *nix world, he has a passion for systems both modern and vintage, ranging from Unix, Perl, Python, and Golang to shell scripting and mainframe-era operating systems like MVS. He’s equally comfortable with relational database systems, having spent years working with Oracle, PostgreSQL, and MySQL.

As an avid user of LowEndBox providers, raindog308 runs an empire of LEBs, from tiny boxes for VPNs, to mid-sized instances for application hosting, and heavyweight servers for data storage and complex databases. He brings both technical rigor and real-world experience to every piece he writes.

Beyond the command line, raindog308 has a life-long love of German Shepherd Dogs, high-quality knives, target shooting, theology, tabletop RPGs, playing guitar, and hiking in deep, quiet forests.

His goal with every article is to help users, from beginners to seasoned sysadmins, get more value, performance, and enjoyment out of their infrastructure.

You can find him daily in the forums at LowEndTalk under the handle @raindog308.