The Second Payroll: The Founder’s Bet Nobody’s Pricing In Yet
AI tokens are becoming the 2nd payroll, and here I explore the ramifications from a founders POV and how we think through what comes next
Welcome back to The B2B Stack. It’s been a while since I told a founder story rather than an adtech one and this week’s edition is that, seeded off a genuinely brilliant piece from Kyle Poyar’s Growth Unhinged. Worth reading in full if you haven’t. This is my extension of it: a thought experiment on what happens when AI adoption stops being subsidised, why that’s a founder’s decision and not an employee’s, and why getting it right might be the thing that lets the next generation of founders build bigger companies with fewer people than we ever could.
Over the 6 years since the pandemic, it’s often been said that this era is the exponential era. This is driven by the rapid changes in technology, and nothing personifies that better than the AI revolution. You’d have to have been living in a cave to not have come across the impact this technology has on everything from work to day to day life, but as the VC funded ‘grow at all costs’ era of AI starts to draw to a close, we have a new paradigm to consider alongside the usual questions around AI adoption, how to use it, where not to, how people fit alongside this tech and all the other big questions.
The one question that hasn’t really been part of this argument is cost. To date the technology has been priced ridiculously cheaply to drive adoption and reliance. As the VC era closes and the AI companies need to drive closer to sustainable growth, the costs will rise. There’s been some viral stories around this out there, and I’m picking it up to run a wider founder focussed through experiment on AI
Every piece I’ve read on the AI cost crisis treats it as an operations problem. Set smarter model defaults, like for example, maybe you don’t need the highest octane model to figure out where Microsoft’s head quarters are or what the bath temperature should be for a 6 month old. Cap the tokens. Route to something cheaper and more aligned to the task. Get finance to build a dashboard.
All of that is possibly true, and all of it is downstream of the actual decision, which doesn’t sit with an ops team or a procurement function. It sits with the founder or business ownership - but in this piece I’ll refer to it as the founder, as that’s my lens.
Whether a business ends up with a second payroll and how well that second payroll is managed is a founder-level capital allocation call, made with the same weight as a hiring plan or a fundraise, not a line item that gets discovered after the fact when the invoice lands. Or at least it should be, so my message to any reader of this is to help bring this awareness to your organisation if it’s not there yet.
Employees experience AI adoption as a tool getting better or worse, faster or slower, more or less annoying to use. More or less job threatening. More or less helpful in their day to day role and responsibilities
Founders experience it as a structural decision about what the company is actually made of. Those are not the same lens, and most of what’s been written about this crisis has been written through the wrong one.
What you’ll learn in this article:
Why a Series B founder’s viral Anthropic bill is a warning sign, not an edge case
The two cost lines that defined my first business, and why founders, not employees, are the ones who actually feel them
Why AI spend behaves like a UK income tax cliff edge, and why a founder’s job is to plan for the cliff before it arrives
Why the founder’s real skill here is balance: using AI as a genuine catalyst for speed and quality without losing the human in the loop
Why this cost crisis balances the books with headcount - meaning AI is much less job threatening and the time for employees to lean in is now - to ensure they are part of the AI native thinking and org structure of tomorrow
Why getting this balance right might let a new generation of founders build serious companies without the headcount their predecessors needed
Where I think this settles, and why it could end up healthier for everyone — including the graduates and early-career people the current system has been squeezing out
The bill that went viral
Marty Kausas, co-founder and CEO of Pylon, posted that his company’s Anthropic bill was about to jump from $400,000 to $1.4 million a year. At this point we can see that mature AI adoption is rapidly becoming a real cost line to the business and therefore needs addressing differently to the AI 1.0 ‘adoption era’
The bit that grabbed my attention was that this bill leap was not because usage exploded.
It was because Pylon was about to cross 150 seats, the threshold at which Anthropic moves customers onto enterprise pricing and switches off the subsidised token rates that came with the smaller plan. Every token, overnight, priced at full API rate.
The post went viral for a reason that has nothing to do with schadenfreude. It resonated because almost every founder scaling headcount right now is sitting somewhere on the same curve and doesn’t know exactly where the trap door is.
Kyle Poyar’s follow-up piece on Growth Unhinged is the best breakdown I’ve read of what happens next: engineering burning roughly £3,100 a head a month equating to roughly the cost of a paired junior developer working alongside each team member (hence the second payroll aspect to this thought experiment), support running an AI-driven investigation on every single ticket at real inference cost (which may have been cheaper with junior employees), marketing spending hundreds a head on copy that, in Marty’s own words, still isn’t very good (AI slop costing real money to produce when a skilled copywriter could have nailed it better). He talked to CFOs at Vanta, ClickUp and SentinelOne for that piece, and the pattern across all of them was the same — spend had climbed faster than anyone’s ability to say with confidence whether it was justified. Marty’s own line stuck with me: most people, including him, weren’t conscious of what they were spending. That has to change, he said. I’d go further. It has to change at founder level first, because everyone below that level is simply responding to the incentives the founder set.
Two cost lines, and the one we’ve been lucky to avoid
My first business ran on two major cost lines. Cloud computing, specifically the internet’s RTB ad calls representing programmatic advertising’s glutinous volumes of ad requests repressing ad opportunities, which at scale which now gets to 1 trillion + a day, is one of the most punishing bid-request economics in software. Everything else cost wise was a rounding error by comparison. Those two lines set the ceiling on how fast you could grow and how much margin you kept while doing it. It was a cost line that squeezed margin and starved people based investments which would have driven growth, and left the business reliant on VC money to scale.
In this AI 1.0 era of subsidisation, there’s still a window to build AI enabled business at artificially reduced costs, which has to be front of mind for any founder right now. The AI companies VC’s are sponsoring by proxy this wave of startups, and FunnelFuel is one of them
it won’t last though and then we’ll be back to what I faced previously, scaling up to cloud computing once the subsidy era was over
Here’s the thing about running a business against two structural cost lines like payroll + cloud (or AI): it isn’t something an employee experiences. An engineer sees the compute bill as a technical constraint to optimise around. A salesperson never sees it at all. The founder is the only person in the building who feels both lines simultaneously, who has to decide in real time which one absorbs the next unit of growth, and who carries the consequence if that call is wrong. That’s not a criticism of employees — it’s simply a different job. Founders hold the balance sheet in their head in a way nobody else in the org is structurally required to.
FunnelFuel has, deliberately and somewhat luckily, avoided that dynamic so far. We’re not running compute-heavy infrastructure at the core of the business model, and our biggest cost line by a distance is still people. AI is still subsidised. That’s helped us grow the way we have, from £1.2M in year one to £12.1M in three years and somewhere around £20m in year four, a business that made it to 33rd on the Sunday Times Fast Growth 100, without a second major cost line dragging on the model as we scaled. I’ve been quietly grateful for that every year we’ve grown. But I don’t think we get to keep it forever, and increasingly I think the founders who assume they will are the ones who’ll be caught out.
AI adoption is starting to look like a second payroll
Here’s the thought experiment. If AI tools genuinely drive the level of adoption and dependency that vendors and investors want them to — if Claude, or whichever frontier model your org standardises on, becomes as load-bearing to daily output as your people are — then AI spend stops behaving like a software subscription and starts behaving like a second payroll. It’s a starker choice to make
Payroll has a headcount lever, a comp band, a hiring plan, a budget owner, and a finance function that reviews it quarterly on principle, because founders and boards have spent decades building the discipline around it. It’s easier to manage
AI spend, for most fast-growing companies right now, has none of that scaffolding. It scales invisibly with adoption, sits several budget owners removed from anyone with headcount-level authority, and as per Marty’s own admission about a $4,000 weekend query he ran himself, even the CEO often doesn’t know what he’s spending until the invoice lands. As a founder, I would put myself squarely in that category too. I’m running around using Ai to prototype, fix problems, run thought experiments, iterate and to generally make myself more productive. I’m not, yet at least, thinking about the cost.
That’s the first version of the second payroll: shadow, unbudgeted, growing with usage rather than with a deliberate decision anyone actually made. The second version is worse, and it’s the one I think matters more for founders specifically. It’s what happens when the subsidy currently propping up frontier model economics goes away, and a cost line that felt like a convenience starts behaving like a second headcount plan nobody signed off on. A plan with no interview process, no comp band, and no natural ceiling except the vendor’s next pricing announcement.
The UK tax cliff edge, but with more ways round it
The UK income tax system has a genuinely brutal cliff edge. Earn over £100,000 and you start losing your personal allowance at 50p for every additional pound, creating an effective marginal rate north of 60% in a specific band, before it flattens out again above roughly £125,000. It’s not a smooth curve. It’s a trap door that a huge number of people don’t see coming until they’re standing on it, and the ones who do see it coming plan around it; pension contributions, salary sacrifice, timing of bonuses etc all need factoring months in advance, because that’s what a good accountant does for a client who’s about to cross a line.
Pylon’s 150-seat threshold is the AI-cost version of that trap door. Below it, subsidised usage. Above it, full enterprise API pricing, applied instantly, with no smoothing and, per Marty’s account, very little warning. Cross a headcount line that has nothing to do with how much value you’re actually getting from the tool, and the bill triples.
The one meaningful difference, and it’s worth dwelling on, because it’s the whole reason this is a founder problem rather than just a finance problem — is that unlike the UK tax system, there are genuinely more ways round the AI cliff edge. Creative accounting has not been eradicated from the AI game plan yet.
How may we do that?
You can split functions across providers so no single vendor sees your full seat count. You can route by task to cheaper models rather than defaulting everyone to the most expensive one. You can hold a team deliberately below a threshold, the way a small business sometimes holds revenue just under a VAT registration line, buying time to plan the transition properly rather than being forced into it. None of that is available to a UK taxpayer staring down the £100k trap. It is available to a founder who treats model spend with the same forward planning as a good accountant treats a tax year.
But “more ways round it” only helps the founders who are actually looking for the trap door before they hit it. Most aren’t, because right now this still feels like a subscription decision made once in procurement, not a structural bet that needs revisiting every time headcount moves.
The founder’s real job: finding the balance, not picking a side
This is where I think the conversation needs to move on from “control the cost” and toward something more interesting, because I don’t think the founder’s job here is defensive at all.
Used well, AI is a genuine catalyst. A catalyst for speed, for the sheer volume of output a small team can produce, and in the right hands, for quality that a leaner team couldn’t have matched a few years ago. I feel that directly. There are things happening at FunnelFuel right now, in how quickly we can turn a partnership brief into a working demo or a data model into a client-ready narrative, that simply weren’t possible with the headcount we had even eighteen months ago. That’s not a cost line to be managed down to zero. That’s leverage, and any founder who treats it purely as a risk to be contained is going to get out-built by one who treats it as a genuine unlock.
But leverage without a human anchor degrades fast, and this is the part that gets lost when the conversation stays purely financial. Pylon’s own experience is the clearest evidence of this: AI-written marketing copy that, by their CEO’s own admission, still isn’t very good, still repeats the same mistakes, month after month, regardless of spend. Sales tooling that got built, adopted enthusiastically, and then quietly abandoned once the novelty wore off, because it wasn’t good enough to survive contact with a real quota. Spend went up. Judgement didn’t improve to match it. That’s the failure mode a purely cost-driven response to this crisis will never fix, because it isn’t a pricing problem - it’s a problem of where the human stays in the loop and where they don’t.
So the founder’s actual job, as I see it, isn’t “adopt AI aggressively” or “control AI spend tightly.” It’s holding both at once: pushing hard enough that the business genuinely captures the speed and leverage on offer, while staying disciplined enough about where a person still needs to be doing the thinking that the second payroll doesn’t quietly become a worse version of the first one — expensive, sprawling, and short on judgement. That’s a much harder balance to strike than either extreme, and I think it’s going to separate the founders who compound advantage from the ones who either overspend blindly or under-adopt out of fear.
In my world - specialised B2B advertising, there is clearly a value in the person who has been ingratiated with the nuances of the industry which will not get exposed from a pure LLM model. However can that person work better with an LLM tag team buddy? yes I absolutely think they can. The model should blend lifting the best of the persons knowledge with the agility, critical thinking, document ceating, challenging aspects of an LLM.
Why this might be good news for the next generation of founders
Here’s the part of this I find genuinely exciting rather than just cautionary, and it’s the reason I don’t think this piece should read as a warning against AI adoption.
Every generation of founder before mine built against a headcount ceiling that was, in practice, non-negotiable. You needed a certain number of people to run support, to write the copy, to build the first version of the product, to chase the pipeline — and the businesses that won were often simply the ones who could afford to hire that headcount fastest. That’s the model I grew up in. Its the model that favoured the markets, like the USA, with better access to venture capital. It’s the model that gave us the two-cost-line problem I described earlier, where payroll and infrastructure both scaled roughly in step with ambition, and where the founders with the deepest pockets or the earliest funding round had a structural head start that had very little to do with the quality of their thinking.
A founder who gets the balance right; real leverage from AI, real judgement retained where it matters, isn’t just managing a second payroll well. They’re building a company that can credibly compete on ideas and execution quality rather than on how many people they could afford to hire in year one. That’s a genuinely different starting position for a founder than the one my generation had, and if enough people figure out how to do it properly, I think it changes who’s able to start a serious company at all. Not because AI replaces people, but because it removes some of the headcount-as-moat dynamic that’s shaped who gets to compete for the last twenty years. The founders worth watching over the next few years, I suspect, won’t be the ones who spent the most on AI or the ones who avoided it hardest — they’ll be the ones who worked out, earlier than everyone else, exactly where the leverage was real and exactly where a person still had to be in the room.
Where I think this settles
Here’s my speculation, and I want to flag clearly that this is speculation, not something I’d bet the business plan on.
I think the current AI pricing environment is a temporary subsidy, not a stable equilibrium - venture-funded model providers burning capital to win adoption before the enterprise pricing cliffs kick in, exactly as Pylon experienced. As that subsidy unwinds across the industry, I think a meaningful number of AI-native businesses end up looking a lot more like my first company: two major cost lines, cloud-and-compute (now AI credits rather than RTB calls) and payroll, both requiring genuine operating discipline, both capable of capping how fast you can grow profitably. The founders who’ve spent this window building the muscle to manage that rather than assuming the subsidy is the new normal are the ones who won’t be caught flat-footed when the pricing resets.
If that’s right, the leadership skill of the next few years isn’t “adopt AI aggressively” or “resist AI adoption.” It’s finding the right ongoing balance between headcount growth and AI spend as two competing claims on the same budget, which is precisely ClickUp’s CFO’s framing of zero-based budgeting, treating AI as an equal competitor for spend rather than a costless add-on sitting above the org chart.
And here’s the piece I think gets underweighted in most of the commentary I’ve read. If AI spend stops being artificially cheap relative to junior human labour, the economics of hiring graduates and early-career employees get more attractive again, not less. This pendulum will swing again. It may take some aggressive government intervention to aid it, such as employment cost subsidies for the young, or it may not, because it will normalise again.
A junior hire who can be taught to do a task reliably, who improves with tenure, and whose cost doesn’t triple overnight because your company crossed an arbitrary seat threshold, starts looking like better risk-adjusted spend than a heavily-subsidised AI workflow that might reprice 3.5x on a vendor’s timeline rather than yours. I’d genuinely like to see a version of the labour market that rebalances toward that — less distorted by a false economics window created by investor subsidy, more accretive to the graduates and early-career people who’ve had a genuinely hard few years finding their first foothold, and more welcoming to the next wave of founders who won’t need the headcount my generation did to prove a business works.
I don’t think we’re there yet. I think we’re still mid-subsidy, mid-FOMO, and mid-figuring-out where the real value is. But I’d rather start planning for the world where the subsidy ends, and where the balance between people and AI is the whole game, than get caught at seat 151 wondering what happened to the bill.
Where do you think this settles for your business - are you already managing AI spend like a second payroll with a founder’s discipline behind it, or still treating it like a subscription someone else will notice eventually?
Sources and further reading
Primary source
Kyle Poyar, “The AI cost crisis is entirely self-inflicted (and Fable 5 just made it worse),” Growth Unhinged, July 8, 2026:
https://www.growthunhinged.com/p/the-ai-cost-crisis-is-entirely-self-inflicted-and-fable-5-just-made-it-worse
The viral post itself
Marty Kausas (co-founder/CEO, Pylon), LinkedIn post on the Anthropic bill jump: https://www.linkedin.com/posts/martykausas_our-anthropic-bill-is-about-to-jump-from-activity-7470505066148352000-zqnt
On Anthropic’s enterprise pricing shift
The Information, “Anthropic Changes Pricing, Bills Firms Based on AI Use Amid Compute Crunch” (April 2026): https://www.theinformation.com/articles/anthropic-changes-pricing-bill-firms-based-ai-use-amid-compute-crunch
On broader frontier model pricing dynamics
WSJ, on OpenAI considering “drastic price cuts”: https://www.wsj.com/tech/ai/openai-considers-drastic-price-cuts-anticipating-war-for-users-with-anthropic-9b8c178e
Developer’s Digest, Fable 5 cost-per-task analysis: https://www.developersdigest.tech/blog/claude-fable-5-pricing-cost-per-task-analysis
CFO commentary quoted/paraphrased in Kyle’s piece (and referenced in yours)
John McCauley, CFO, Vanta
Dan Zhang, CFO, ClickUp
Sonalee Parekh, CFO, SentinelOne
Adjacent context referenced in the original piece (not directly used in yours, but worth knowing they’re there)
Tesla’s $200/week AI spend cap: https://www.theinformation.com/articles/tesla-caps-employee-ai-spend-200-per-week-adoption-push
Uber’s $1,500/employee/month cap: https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/
Ramp’s AI spend intelligence launch: https://ramp.com/blog/trillion-dollar-ai-blindspot
Your own framing (no external source — original to your piece)
The UK income tax cliff-edge analogy
The “two cost lines” founder history and FunnelFuel comparison
The founder-vs-employee mindset argument and the “next generation of founders” speculation