ArticleIssue #192 | AI Is A Power Tool
If you spent any time online (or watching the news) this week, you’d be forgiven for thinking the robots finally showed up. Barclays, American Express, Netflix & eBay all announced significant job cuts within days of each other, a decade-old flaw in the language behind most of the web turned into “Black October,” and every headline said some version of the same thing: this is it. This is the moment the “AI is taking all the jobs” narrative moves from deluded techbro fantasy to main street reality. On the surface, that’s exactly how it looks – powerful new models drop, then some of the biggest names in banking, payments and tech start cutting jobs.
I don’t think that’s the story.
If you dig a bit deeper, you’ll see things are not exactly what they seem.
The standard rebuttal is the “anti-Ultron” line: AI today is not a wanton superintelligence. It has no goals, it doesn’t decide anything, it does what people tell it with varying levels of competence. That’s all true, but entirely beside the point. The serious bear case is duller, but a lot more dangerous: AI makes a company 4x more productive, the company keeps output flat (or maybe up a bit) and the difference flows to the bottom line. If you can get the same output for 1/4 (or even ½) the cost, that’s pretty tempting.
Whether that happens comes down to 2 numbers: (1) how much more productive the tools actually make a given team/organization/whatever; (2) how much more of its output the market can absorb once said output gets cheaper.
No one talks about either one. Most of the people writing their 2027 budgets & hiring plans right now are – at best – guessing and – at worst – not thinking about either one.
What The Headlines Actually Say
Let’s start with the gloom-and-doom part of the AI story (because why not): the layoffs.
Barclays expects to cut ~5,000 roles and cited AI-driven efficiency as one of the reasons… but ~1,000 of job cuts were already gone in the first half of the year, so part of this week’s headline is a restatement. American Express is cutting ~3,500. Its CFO described the goal as getting leaner and preparing for AI. Netflix is cutting ~1,000, mostly in product, tech, legal and HR. eBay is cutting ~9% of its workforce to simplify operations and move resources toward its fastest-growing areas, with no mention of AI at all.
If you put those four big announcements one after the next, it creates the appearance of a wave. That’s exactly what the media has done, and it’s probably why I’ve gotten a half-dozen calls from reporters, conference organizers and industry analysts asking about what AI is doing to our hiring + staffing plans.
But if you count which ones specifically cited AI, the answer is only 2 of 4, and one of those 2 was partly announcing cuts it had already made.
The aggregate data does have a real spike in it. Challenger, Gray & Christmas (the staffing + restructuring firm) counts 120,136 announced U.S. cuts attributed to AI through September – which is ~21% of the year’s total. That makes AI the single-most-common reason companies have provided for reductions in force. In May alone, AI accounted for 38,579 of 97,006 cuts, the highest monthly total since Challenger started tracking it in 2023. In fact, AI has been the largest reason for RIFs every month from March through July. But even that has faded, from 1st in July to 5th in September (3,961 of 43,281, or ~9%).
For something that’s supposed to be ushering in a jobpocalyse, that’s awfully inconsistent. If anything, you’d expect to see somewhat linear growth – companies following one another to reduce headcount and rely more on AI should increase with time.
Of course, the Challenger data only counts the reasons companies give, not the reasons they have. And – at the risk of being more blunt than I usually am, “AI efficiency” plays a lot better on an earnings call than “we over-hired” or “costs are up” – and guess which 2 of this week’s 4 RIF announcements arrived attached to quarterly results? Hmmm.
You know what they say about coincidences, right? They’re just plans other people make and don’t tell you about.
Let’s Talk About Power Tools
I absolutely love power tools. I think they’re incredible. Putting together Ikea furniture with my DeWalt driver is infinitely more satisfying than getting blisters on my hands while the bookshelf still wobbles.
Power tools don’t just reduce the pain/blisters, either – they increase efficiency. I can put together a HEMNES bookshelf in about ~20 minutes with my trusty DeWalt companion; if I have to do it the old fashioned way using the tools in the box, it takes ~90 minutes (we recently went to Ikea, if you couldn’t tell).
You’ll find the same lesson everywhere else, too – a framer with a nail gun can put up a house 2x-3x faster than one with a hammer. The nail gun doesn’t read the blueprints, doesn’t know which walls are load-bearing, doesn’t pull the permit & doesn’t care whether the house is any good. It also doesn’t care if the nail goes through wood, air or your hand. It just shoots nails.
That description perfectly captures what we have in AI tools today.
AI today compresses execution – a 1st draft of ad copy in seconds, a dozen variants of an ad in a minute, a search term report pulled apart, a 40-page RFP summarized before the Keurig finishes brewing your morning coffee. Of course, that leaves deciding which of those ~12 variants fits the brand, the client, the legal constraints (see last week’s issue) and the budget to someone else (aka you). That’s the condensed argument from The Scarcest Input You Own.
What it didn’t answer is the question the readers of this week’s headlines are actually asking: what happens to us (the people who do the jobs) when execution is 4x faster?
Where The Saved Hours Go
Economics – while generally questionable as a science – does have some useful models that I’ve used in thinking about the impacts of AI.
One of those is the theory of output. At its most fundamental level, a team’s headcount is the output it’s asked to produce divided by the output each individual person can produce (that’s obviously inexact, as you have managers + entire classes of workers like HR who do not directly contribute whatsoever to the core output of the business, but certainly provide support – but you get the idea).
Now, if we assume something increases the productivity of workers (k), then obviously the number of people required is reduced. For instance, if AI makes each person 4x as productive, k = 4, and the same work needs 1/4 of the headcount (at least, in theory).
That’s exactly the model the AI doomers are using. And, to be blunt, it’s correct PROVIDED the output remains fixed. But, as William Jevons discovered way back in the 1800s, the output almost never stays fixed when it becomes cheaper. With our new productivity increase, each unit of output now costs 1/k of what it did, and how much more of it is purchased by the market depends on the price elasticity of demand for that output (ε).
Put 2 and 2 together, and you’ll see that demand rises by k^ε while each individual worker produces k times as much, so:
Headcount multiplier = k^(ε − 1)
If you look at any row, you’ll see a trend emerge. At a 4x gain, the same output requires anywhere from 50% the team to 2x the team; the variable deciding which is ε.
Similarly, if you read down any column, k (i.e. the impact of AI) simply amplifies whatever direction ε has already picked. At ε = 1, it makes no difference whether the tool makes your team 2x or 5x faster; the headcount holds.
That’s Jevons paradox with its conditions attached. Jevons argued in 1865 that more efficient steam engines would increase Britain’s total coal consumption, because cheaper power would find far more uses. It gets repeated as if it were a law. It’s a special case that only holds when demand is elastic – which means “cheaper output always creates more work” isn’t a universal law of technology, but rather a bet on the price elasticity of demand.
You can actually formalize this using econ 101 terminology.
If:
- labor L = Q ÷ k
- demand Q ∝ p^(−ε)
- Assuming labor = the dominant cost, p ∝ 1 ÷ k.
Then substitute and you’ll get L ∝ k^(ε − 1). It assumes the savings show up as lower cost per unit and that elasticity holds roughly constant over the range. That’s fine for a rule of thumb, but it is almost always wrong eventually.
Which Of Your Outputs Is Elastic?
Now, how does any of that help marketers or business owners?
Well, for starters, most businesses (including marketing teams) don’t actually sell their output by the unit, which makes the price internal: what the business will fund more of once each unit gets cheaper. Ask that question function by function and the answers diverge quite rapidly.
Reporting is about as inelastic as it gets. Nobody (and I seriously mean no client) wants 4 monthly reports. If AI makes a given report 4x cheaper to create, the client is still only getting 1 report. Using the terminology from above, this puts ε = 0 and the multiplier near 1 ÷ k. Of everything a marketing team does, reporting is where AI shrinks headcount most reliably…. and (honestly) where it should. (Quick aside: this assumes the quality of the report remains fixed as well – but you could make a compelling argument that if AI makes assembling the report significantly faster + cheaper, the surplus resources will simply go toward making it more comprehensive, valuable or tailored).
Creative testing, on the other hand, looks pretty elastic. And, up to a point, it is. Cheaper variants buy more angles, more personalization & more formats in the same unit of time. But a media budget can only support so many ads before individual units become starved of budget determining success/failure becomes impossible below a learning threshold (the ceiling from Everyone’s Wrong About Creative Volume). Past that point, more variants has a negative expected value. ε for creative is high right up until it crosses that threshold, then falls off a cliff and into the ocean (i.e. negative).
Content has been a curious but cautionary tale. Since ChatGPT arrived on the scene a century ago in 2022, the raw volume of new content has exploded. The elasticity of volume is enormous. Yet the hit rate (i.e. the rate at which pieces of content actually drive qualified traffic to the sites that publish them) has collapsed. That theory was the entire premise of Nothing The Model Doesn’t Already Know. The k that matters there isn’t pieces per person; it is hits per headcount. A copywriter who uses AI to write 4x more articles at 30% the old hit rate has a “real” productivity gain of 1.2x, which is a lot smaller than the talking heads + X doomsayers would have you believe.
Then there’s taste and judgement, neither of which moves at all (in fact, it might well be negative). AI allowing teams to draft a plan 10x faster doesn’t make the person approving it (or funding it) 10x faster at figuring out whether it’s worth doing at all. End result? k for judgment/taste/style is likely <1, so the multiplier is ~1 regardless of ε. Judgment, taste & discernment – if anything – get MORE valuable as the cost of production declines thanks to AI. That’s probably good news for the CMOs + Marketing Directors out there.
The k In The Demo Isn’t Your k
Now, everything I’ve just outlined above assumes k in the real world is equal to the k they think is true, based on a demo or an anecdote or a sample of n=1. The quintessential example of this is something I heard from our team this week – they had a demo with a tool that reduced the time required for a creative analysis task from ~1 hour (as documented in Harvest) to ~15 minutes, so they assumed k = 4.
The reality is that a demo tends to take place in as close to ideal conditions as you’ll ever encounter, which makes that number a bit of a fiction. That’s not entirely true, but it leaves out most of what it costs to make the 15-minute version work every time.
A few examples:
Failure: A tool that gave you a correct output once has been tested on a sample of n=1. That tells you nothing about how often it fails across the inputs you’ll actually give it (because we all know that no two situations are exactly the same, the data we get is never quite identical, etc.). We also know that the more varied the inputs, and the more steps you chain together, the higher the error rate climbs. Maintaining a failure rate of ~10% to within 5 points requires ~100 varied inputs. Mathematically, this is a variant of The Tampering Tax, just applied to a workflow instead of an ad. Until you’ve run them, you don’t know whether you built a process or got lucky once.
Review: once the AI thingy has done its miracle work, a human still has to review the output and identify where (and if) the system made a mistake.
Correction: when (not if) you find some issues, a person (maybe augmented with AI) has to re-do the part that’s wrong.
Calibration: virtually every AI system is built from a set of instructions; creating those is often a far more complex process than it seems, because so much of what ends up in any particular output above is latent knowledge – a senior attorney just knows that a given judge is wholly unreceptive to a particular argument; an accountant knows that a particular deduction can’t be taken this year because it’s being saved for next; a media buyer knows that a given result is just *off* based on years of accumulated knowledge about the client’s business and industry; a marketer understands that the client will NEVER go for that influencer because the VP of Marketing hates his guts or that topic is completely off limits because ::reasons::.
Maintenance: every AI system is nondeterminant in a way that mathematical operations are not; this is why model drift occurs. The same prompt, the same inputs given to the same model 2x will produce slightly different outputs. You need someone to – periodically – reassess the output of a given AI system to ensure it remains within the acceptable bounds.
Tokens: AI systems aren’t free labor; they just convert a fixed cost (i.e. a salary) into a double-variable cost (i.e. tokens) – you don’t know precisely how many tokens a given task will require to reach a satisfactory output, and the cost of those tokens themselves is variable.
Now, if you express each one in hours per hour of original work, you’ll see the the actual productivity gain is:
Net k = 1 ÷ (1/(k + review + failure rate + build + tokens))
Let’s start with the “claimed” 4x faster example from the demo.
Review adds 0.15 hours per original hour (someone still has to read the report it produces and make sure it’s right before it is sent to a client). A10% failure rate means that the portion of the report that is re-done by one of your SMEs adds another 0.15. Build and maintenance, amortized over a year, add 0.12. Tokens and tooling add 0.05. The denominator goes from 0.25 to 0.72. k falls from 4.0 to 1.38.
Then, take that “real” K and run it through the table at ε = 0.5. You’ll see that “cut half the team” (which was the conclusion at k=4) quickly becomes “reduce maybe ~18%.”
You might read that and think the honest version of the AI narrative is that it’s a giant ponzi scheme – that it doesn’t work at all. I do think that’s wrong as well.
The productivity increase from AI is real. I’ve seen it first-hand in our business, and I’ve seen it across dozens of our clients’ businesses, too. The hard truth is that the magnitude of the gain is just a fraction of what’s claimed, because much of “savings” is lost via failure, correction, calibration, maintenance & tokens. None of that is going to change anytime soon, because those are somewhat inversely correlated – reduce the error rate to 5%, and the time required to find the errors increases as does the cost of those errors (at least, in many fields) – which means insuring against them becomes more important.
If we’re all being both candid + materialistic, this also explains why so few people seem to be getting rich off AI: when a productivity gain is available to every competitor at the same time, competition hands most of it to customers in the form of lower prices. Diffuse gains don’t produce overnight monopolies; they show up as margins that don’t move.
Case in point – computers spent ~15 years showing up everywhere except in productivity statistics. The people reliably making money during that period were the ones selling the hardware. Does that sound a bit familiar, NVIDIA stockholders?
Nobody Sues A Chatbot
If AI could replace expertise outright, the first casualties would be the people who sell it by the hour: consultants, lawyers & agencies. A model – with the right inputs – can build a consultant-quality deck in less than an hour. Yet today, plenty of very large, very sophisticated, very profit-oriented companies still pay McKinsey or Bain or Deloitte. They still have DLA Piper or Hogan Lovells or Kirkland Ellis as their outside counsel. They still rely on PwC or KPMG to assemble their financial statements and file their tax returns and 10Ks.
But if you think those companies were ever – at any point – paying for the deck or the brief or the filing, you’d be quite wrong.
What they’re actually paying for is two things: (1) credibility + (2) accountability.
There’s a reason insurance companies don’t fear AI-generated complaints: there’s no enforcement mechanism. I can generate a terrifying-sounding complaint with Claude, but if the other party doesn’t think I can credibly enforce it, it’s worthless.
Similarly, firms want expertise because it comes with accountability. There’s someone credible who puts their name (or their firm’s name) on a recommendation, then sits in a meeting with the CEO or CFO or Board, takes the hard questions and is held responsible for the outcome when things inevitably go sideways.
Could you imagine walking into a meeting with your boss and justifying a consequential decision with, “Well, the AI said to do it” as your rationale? You’d likely (and rightfully) be fired on the spot.
All of this converges into this: AI can make producing the artifact (the deck or the filing or the tax return) cheaper; but the artifact was never the expensive thing to produce. What has always been expensive is judgment, taste and accountability. When options become – effectively – free, the scarce thing is a credible party choosing among them. The people who can do that well are about to get a lot more leverage.
What I’m Not Dismissing
Now, there is another half to this particular debate.
“It’s just a tool” can become its own kind of denial. It’s tempting to read everything I just wrote, exhale deeply, and think you’re safe no matter what.
That’s not exactly true, either.
If you visualize competence as a pyramid, with the lowest-leverage, most routine, mundane tasks at the bottom and the highest-leverage, most judgment/taste oriented tasks at the top, you’ll quickly realize that those tasks at the bottom are the most vulnerable to displacement by AI. In practice, that looks like initial drafts, research, competitive analysis, reporting, QA – all things junior people have always learned from before they’re promoted up the pyramid to progressively more strategic work. Each one is – essentially – execution against mostly inelastic demand.
This just so happens to be the corner where cost savings have the greatest long-term negative impact: if you eliminate these tasks, the savings this year look great. The savings next year probably looks pretty stupendous, too. But what about in 3 or 4 or 5 years? You’ll have saved quite a bit in relatively low-cost labor, but done so at the expense of your leadership pipeline. The unfortunate reality is that no one develops good judgement without first doing a lot of grunt work.
On the flip side of the coin, switching from a manual tool to a power tool also scales potential harm. The “Hack October” is that exact principle: 2 critical flaws in PHP (the language behind roughly 70% of websites) sat undetected for more than a decade before Calif.io, Wiz, OpenAI and Quarkslab disclosed them. PHP shipped emergency patches on October 21, but the attackers (thanks, in large part, to AI tools) were already ahead: Amazon saw exploitation attempts starting October 17. Proof-of-concept exploits were in use within hours of disclosure.
Nothing here happened autonomously; people were there for all of it, on both sides. AI tools allowed the attackers to find the weak points AND execute the attacks far faster than most cybersecurity professionals had planned. That left the companies on the receiving end exposed for far longer – simply because patching still runs on change windows, vendor development queues and somebody in IT remembering that the one thing you use relies on PHP. Honestly, that gap will get wider before it gets narrower.
I also could be completely wrong.
Amara’s Law says we tend to overestimate a technology’s effect in the short run and underestimate it in the long run. That pattern played out near-perfectly with the internet – it killed classifieds, gutted newspapers and erased trillions in the dot-com bust, then built the industries most of the people reading this work in. This argument rests on AI producing work rather than making accountable decisions. If that changes – if an AI system ever reaches the point where it can provide accountability + credibility – then the calculus shifts. I don’t think we’re anywhere close to that today, and getting there will take massive, outside-of-technology changes. Insurance companies aren’t just going to start covering AI attorneys or AI doctors. Lenders still do not underwrite AI-generated forecasts. Right now, AI is a power tool with ambition, not a wholesale displacement of the global economy.
So Where Does That Leave You?
1. Measure net k, not the demo. For every process/workflow you’ve automated, log review time, redo rate + time, build & maintenance time and token spend for a month, across at least 100 varied inputs. Then compute net k. If it comes back at 1.5, then discount it a bit more and make that the number you use to plan expected headcount.
2. Estimate ε for every recurring output. Ask what the business would actually fund if each one cost ¼ as much to produce. If the honest answer is “the same amount,” ε is near 0 and that work shrinks no matter how anyone feels about it. If the answer is “4x as much, and we’d find a use for more” – that’s probably where your savings will go.
3. Decide How To Spend The Savings BEFORE Leadership Does. If AI saves your team 100 hours a week, put the plan for those hours in writing – more tests, more training, more clients, better strategy, incremental service offerings, more comprehensive whatever – with a rationale attached. Without that, the default answer (in the short term) is fewer people. But no sane person cuts headcount if doing so actually harms short-term profitability or long-term viability.
4. Protect the bottom of the pyramid. Give junior people judgment work earlier and pair them with the AI tools instead of replacing them. Today’s junior people are tomorrow’s leadership; if you don’t cultivate it now, you’ll have to buy it later (and trust me, buying leadership is VERY expensive).
5. Put a name on every decision. Let AI expedite the first draft of the plan or the ad copy or the analysis. Resisting that does not help you, especially in an increasingly pro-AI environment. Where you can have an impact is ensuring that a person reviews it, fixes it and signs off on the final deliverable AND that person answers for it. It’s awfully easy (as a CEO now) to mistake the artifact for the value when you don’t see all the inputs that go into that thing.
he headlines/media/twitterverse wants you to believe the machines are coming for everyone. I don’t think that’s true, and I think you can prove it for your company/organization with a little math and a bit of time.
