The AI Math is Not Quite Adding Up
Companies are realizing that AI is actually kind of expensive, even compared to those pesky human employees.
When I first began using AI at work, I had a revelation that I continue to remind teams I work with: generative AI models are English majors, not Math majors. They are not great with numbers, and their outputs resemble something you’d be more likely to find in a creative writing class at a liberal arts college than in a math class at an engineering school. Just the other day, I had to remind CoPilot how to round numbers properly. Ironic, given that machine learning is how most of us interacted with artificial intelligence before GenAI came along.
That could be one of the reasons that some math-centric CFOs are having buyer’s remorse with AI. Despite the promises of massive cost savings of replacing the largest debit item on the old P&L - human labor - with bits and bobs behind the computer screen, reality is proving the inverse. And the main reason is how human labor has historically been utilized in the services industry and how that equation does not add up for artificial intelligence.
How Human Labor Is Utilized Today
In a TDNBW post about how AI might end up killing the billable hour once and for all, we dove into the FTE model that agencies have turned to in a project/AOR scope world. We touched on incentivized burnout briefly, but let’s take a slightly deeper dive.
There are two aspects to this: human labor as a fixed cost, and the way agencies approach allocation of human resources.
When someone is hired into an agency, overwhelmingly, they are hired as an exempt employee on a fixed salary with no bonus eligibility. So whether that person works 35 hours or 70 hours in a week, their cost to the agency remains the same: their salary and their benefits. Purely from a financial standpoint, it would make sense that the company would want to get as much labor out of a single employee as possible, given that every incremental hour worked is essentially free.
Now, things like employee morale and burnout significantly reduce work product quality and have to be taken into account, so the approach to employee utilization is not nearly as Machiavellian as the previous paragraph might make it out to be. But that tension exists.
Secondly, there is the allocation aspect of running an agency whose P&Ls are account-by-account. While each agency tends to approach this slightly differently, my experience across multiple holdcos saw the same general approach.
Most employees are allocated 100% to an account - this means 100% of their billable hours and 100% of their cost go to a single client. Clients like this, as they like having stability and an employee’s full attention. Employees like this for the same reasons.
But agency life isn’t perfect, and things like attrition or busy seasons or SME borrowing or pitches occur across accounts. When this happens, employees can be loaned out to other clients or house projects. Without getting into the nitty-gritty financial details, this cost is spread across the clients or house codes an employee works to ensure that everyone is paying their fair share. The cost to the agency remains the same - the employee’s salary and benefits do not change, no matter how often they are overallocated or spread across multiple clients.
Utilization and overallocation tend to go hand-in-hand. If you’re put on a pitch, you’re expected to do that work on top of your regular duties. If you’re loaned out to another client, you’re expected to maintain the same level of service to your allocated account in addition to whatever help the other client needs. The key thing to remember is that while a particular employee’s output or work product may increase drastically in the short-term, their remuneration (and thus cost to the agency) does not. Human labor is essentially a fixed-cost model for the agency.
So AI’s Gonna Make All This Faster, Right?
Not quite.
First, there’s the fact that AI is not driving efficiencies in employee work in the way originally thought. It turns out that what AI has done has, in fact, increased work for existing employees by widening their scope and expanding their remit into areas previously unknown to them.
But Matt, you ignorant slut! - you might say - this is the very efficiency we’ve dreamed of! Product managers are vibe-coding, engineers are doing creative work! We have finally realized the AI utopia of everyone being an expert in everything!
Again, not quite.
When a product manager vibe codes, you can’t just ship the resulting product. Someone with even a modicum of expertise in coding should be reviewing it. What does that do for the existing engineers at your company? It adds a lot of work to their plate to double-check the work of people who have no business writing customer-facing code. So now Jim in sales - whose chatbot sold cars for $1 - not only added work onto his own plate but also onto his peers’ plates to make sure his company didn’t go bankrupt giving away product. You can see how the efficiency gains actually cancel themselves out when (if?) quality work standards are maintained.
Of course, you could ship out a vibe-coded product without those pesky checks, but then you would simply be adding to the proliferation of code that has more security flaws than a Qatari 747. But that will be a TNDBW post for later, as we have barely begun paying the piper when it comes to these vibe-coded products into which we’re entering our personal and financial information.
But AI’s Gonna Make All This Cheaper, Right?
Not even close.
Forbes’s Jemma Green has an excellent breakdown of the strategy for companies that dove into the AI pool without first taking a gander at what might be in it.
There are the companies like Uber that encouraged their employees to begin utilizing AI, but then did not see a relationship between productivity and end-user benefit despite the increased cost.
There are companies like Microsoft, that layoff thousands with one hand, citing AI efficiency, while with the other hand instruct certain divisions to stop using it because it’s too expensive.
In short, companies incentivizing AI usage are having severe regrets about it. Some corporations even set up internal leaderboards, hoping to encourage usage. As humans are wont to do when incentives aren’t properly guardrailed, employees have taken to “tokenmaxxing” - essentially having AI run pointless queries in order to burn tokens to gain prominent positions on internal leaderboards. These leaderboards have since been taken down, and companies have backtracked on asking their people to plow into AI.
My favorite anecdote is the company (unnamed, but if they’re public I have to imagine it will show up in a future 10-Q somewhere) that forgot to set usage limits on their Claude Enterprise license and spent half a billion (with a B) dollars in a single month on AI tokens.
Don’t Forget the Variable Cost Model
And this is where the reality of AI bumps up against the reality of late-stage capitalist labor utilization. Whereas existing labor utilization calls for pushing employees to the brink - remember those incremental hours don’t come with incremental financial costs - the way AI works is that the more you use it, the more you pay.
The old linear model of fixed labor cost no longer applies, and trying to force it onto a variable-cost AI model is going to have you asking your accountant to crunch the numbers again.
It may not cost extra to put a service industry employee on a pitch, or loan them out to a client in short-term need, or have them help as an SME on a project But assign that same project to an AI agent and it requires tokens - and thus incremental money - thanks to AI companies realizing their initial pricing model was as bad as Michael Scott Paper Company’s.
The idea that AI could wholesale replace human employees has hit such a wall that even those companies with the most financial interest in this happening are walking back their statements. OpenAI’s CEO Sam Altman - who previously said that AI would “probably replace most of the jobs people do today,” has more recently conceded, “I now think I understand more about why it hasn't, and I'm obviously grateful, but that is an area where my intuitions were just off.”
We can debate how wrong Altman’s intuitions were, and how much of the financial underpinnings of the US economy are hanging on the words that were “just off,” but again, that’s a post for another day. Energy would be better spent on rehiring all those laid-off workers so they can get back in the office, so we can drive incremental hours at no incremental cost.
The strategy for services employees during the AI revolution should remain the same: Be That Human. Enhance your ability to look into the future, your clarity of process and explanation, your creativity, and - perhaps most crucially - your emotional intelligence. These continue to be AI blind spots - and as they potentially close and others open, find them and fill the gaps in ways that only humans can. This is how you remain one step ahead of AI itself, not to mention those executives trying to cut corners by cutting jobs.
Grab Bag Section
WTF Con Ed: Shortly after I moved to NYC over two decades ago, I decided to up my college wardrobe with a shirt that used Con Edison’s famous blue lettering and font to say “Con Everyone.” I thought it was clever and I was a young smartass. Now I am an old smartass, and I still think it’s clever. Alas, I cannot find the shirt (thanks a lot, trademark laws.)
And part of what made the shirt funny was that it’s kind of true. Con Edison has some of the highest electricity rates in the nation (closing Indian Point certainly did not help), but from this newsletter’s personal experience, some of the worst service reliability in the suburbs. Growing up outside of Boston, I think I lost power two or three times over an illustrious 18 years, and never for more than 12 hours. Having lived in my Westchester home for five years now, I’ve lost power two or three times for over 12 hours. To be clear, both the property tax and the power rates are lower in the Boston homes I grew up in, despite the vastly different experiences with outages.
The takeaway - other than the fact that Massachusetts has a better cost-benefit ratio than New York (even though both are high-cost, high-reward states) - is that Con Ed (and the greater New York City area) is not ready for climate change. A single thunderstorm over the July 4 weekend wreaked havoc on the area (I have neighbors who were out more than I was, so I’m actually one of the lucky ones.) Imagine what will happen during the next - real - storm?
One benefactor of all of this? Whole house generator companies. This newsletter is getting too old to wheel the portable generator out this much to keep the fridges on. Plus, it would be nice to have heating or cooling during an outage (something my wheely gen cannot handle.) It’d be nicer to have climate-ready utilities and a sincere pivot to green energy and other carbon-neutral energy approaches, but you take what you can get.
Multimedia of the Week: I’m about to close out Adam Higginbotham’s Midnight in Chernobyl (don’t tell me how it ends), and it’s one of the better books I’ve read in the past few years. The intertwining of the rise of civil nuclear power, a post hoc view behind the Iron Curtain, and the absolute chilling effect the disaster, its cover-up, and its “cleanup” had on Western nuclear power makes for a real page-turner.
And that last bit is one of the largest takeaways for me. I think it’s obvious that the USSR was a wildly bureaucratic behemoth that continually got in its own way when it came to expansive projects like civilian nuclear power, and it’s no surprise that it (literally) blew up in its face thanks to the corners that were cut and the politics involved.
But what is more surprising about it is that Western governments looked at Chernobyl, the RBMK design (which was beyond flawed compared to Western plants), and the USSR process and cover-up and said “Yeah, we’ll stop producing ours and begin phasing out our existing ones, too.” That would be like the entire car industry shutting down thanks to the Ford Pinto fiasco.
But shut it down they did, and in doing so, they took one of the cleanest, nearly limitless energy sources away from a planet in desperate need of it. We wouldn’t even be having AI data center debates about electricity if we mimicked the French model of nuclear power. It’s as close as you can get to John Galt’s ambient static electricity motor.
And we turned our backs on it, because we were scared of a potential tragedy all due to one of the least prepared nations on earth suffering one. Instead, we doubled down on fossil fuels, and now instead of a potential disaster from a nuclear plant, we’ve essentially guaranteed one from climate change. One of history’s greatest self-owns.
Quote of Week: "Oh, meltdown. It's one of these annoying buzzwords. We prefer to call it an unrequested fission surplus." - Mr. Burns
AI Usage In This Post
I’ve been messing around with Claude to act as an editor for this post, with explicit instructions to simply point out issues as opposed to writing anything or even suggesting copy. I used it briefly on this, but really only paid attention to the typos it pointed out (which I already have AI-powered Grammarly to do, so the point is kind of moot.)
This Claude Project did help me with some research on the human vs. AI cost debate in employment, but if I’m being honest, it was surface-level, and I didn’t really push it to dig too deep.
See you next time!