Amortising a hallucination

We are currently adjusting our EBITDA to account for a sophisticated parrot that read the internet.

Not just my company. We — collectively, as an industry. Across every enterprise on earth, the fastest-growing line item in the budget is a thing that didn’t exist three years ago, doesn’t have a desk and absolutely refuses to explain its own pricing.

Your company is currently trying to amortise a hallucination. Somewhere down the hall, your finance department is locked in a windowless room, sweating over spreadsheets, trying to figure out how to capitalise a vibe.

Gartner says there is $2.52 trillion in global AI spend this year.¹ Up 44% year-over-year. Yet only 14% of CFOs report clear, measurable ROI.² Ninety-five percent of generative AI pilots fail to deliver tangible P&L impact.³

We are spending at unprecedented scale on something we cannot meaningfully track. And it’s about to get six times worse.

For every dollar on tools, six on the work

Sequoia published a piece last week called “Services: The New Software.” It’s a brilliant treasure map showing exactly where the AI gold rush is happening. What it leaves out is the fact that enterprise CFOs are the ones being forced to blindly fund the shovels.

Their thesis is simple: for every dollar a company spends on software tools, it spends six dollars on the services those tools support. A $10K QuickBooks licence is nothing next to the $120K accountant who actually closes the books. So the real AI opportunity isn’t building a better QuickBooks. It’s building the thing that closes the books.

To be clear, you’re not just replacing Jim from Accounting. You’re replacing Jim with an algorithm that processes a million invoices a second, drinks virtual gin, has no concept of what money actually is and aggressively refuses to show its working. Jim’s $120K salary is gone. You’re now paying $400K in API calls to an entity that cannot be put on a performance improvement plan.

Sequoia calls these “autopilots.” Not copilots that help a professional do the work — autopilots that deliver the outcome directly. Draft the NDA. Process the claim. Triage the IT ticket. The buyer isn’t the law firm or the broker anymore. The buyer is the company that needs the job done.

They’ve mapped the TAM, and it’s massive. Insurance brokerage: $140–200B. IT managed services: $100B+. Accounting and audit: $50–80B. Recruitment: $200B+. Management consulting: $300–400B.

Go read it — the framework is phenomenal. But I want to talk about the part it leaves out: how exactly do you get an itemised receipt from a bird?

3,600 photocopiers

Last night, I went slightly unhinged on LinkedIn about photocopiers. The metaphor: companies are ordering 600 photocopiers without knowing what they want to print. AI spend is exploding, and nobody’s connecting it to revenue.

I was thinking too small. Sequoia’s 6:1 ratio changes the maths.

Say you’ve got 600 engineers. You give each one a $20/month AI coding assistant. That’s $144K/year. It’s a lot, but it’s trackable. You know who has a seat. You know what they shipped. Your CFO can draw a straight line from the spend to the headcount plan.

Now apply the 6:1.

That $144K in tool spend implies $864K in AI services spend — the outsourced autopilot accounting, the agentic IT support, the automated claims handling. That money doesn’t sit against a headcount line. It arrives as vendor invoices, API bills and consulting agreements with zero connection to a project, a product or a financial outcome.

You haven’t bought 600 photocopiers. You’ve bought 3,600. And 3,000 of them are in a field somewhere, currently jammed, aggressively printing blank pages into the wind, billing you $140 an hour for “strategic throughput” and sending invoices to your accounts payable team with the subject line “AI-powered transformation — DO NOT REPLY.”

79% of total AI spend. 0% attributed.

The visibility gap

There are two trend lines in enterprise AI that should legally not be allowed in the same chart together.

One line goes vertical. The other one barely twitches. The space between them — roughly $2 trillion — is money that went somewhere, and nobody in finance can explain where. It’s the parrot’s expenses. It’s the blender’s consulting fees. It’s $2 trillion of “look, the outputs are great, can we just talk about this next quarter?”

Sequoia’s map has a missing axis

Sequoia’s opportunity map is built on two dimensions: how much of the work is “intelligence” (rules-based) versus “judgement” (instinct), and how much is already outsourced. Autopilots win where intelligence is high and the work is outsourced. Frictionless vendor swap.

Great framework. But it’s missing a dimension.

The third axis is cost attribution difficulty. And it correlates with exactly the wrong thing.

X-axis: Ease of AI automation. Y-axis: Cost attribution difficulty. The easier it is to automate, the harder it is to account for.

The verticals Sequoia identified as the biggest autopilot goldmines are the exact same verticals where cost attribution is a complete nightmare. Insurance brokerage? The per-policy AI compute cost is invisible. IT managed services? Token consumption per ticket is unknowable without tooling that doesn’t exist. And accounting — the irony should be physically painful — has the worst attribution of all, because multiple AI tools touch each engagement and nobody tracks which parrot actually reconciled the ledger.

Sequoia mapped where the money will flow. This chart maps where it will disappear.

Autopilots make the problem worse, not better

This is the bit I keep coming back to.

With a copilot — say, Harvey selling to a law firm — you have a paper trail. A per-seat licence. A professional who used the tool. A matter number. Three cost lines, one outcome. Annoying to reconcile, but entirely possible.

With an autopilot — an AI selling the completed NDA directly to the company — you’re buying the outcome blind. One invoice. No breakdown. You’re paying McKinsey rates for a blender that swallowed a thesaurus, and you have no idea how much was compute, how much was human review and how much was the blender itself.

By default, that cost lands in operating expenditure.

If your finance team can’t prove it qualifies as R&D or capitalised software development under IAS 38 or ASC 350-40, it’s a straight P&L hit. Your margin drops. You miss R&D tax credits worth 6–10% of qualifying spend. The money falls right through the gap between “we bought an outcome” and “we can’t prove what the outcome cost to produce.”

Same spend. $250K swing in margin impact.

Picture the board meeting. Your CFO is sweating through a shirt that cost more than most people’s rent, trying to explain why EBITDA is down 8% despite record output. The real answer — “a sentient PDF reader decided it deserved a Q3 bonus and we legally have to classify its vibes as operating expenditure” — is not going to land well with institutional investors. So instead, they say “we’re investing in AI-native transformation.” Everybody nods, and nobody asks a follow-up question because nobody knows what follow-up question to ask.

We’ve been here before (sort of)

In 2014, companies adopted cloud without cost governance. AWS bills doubled every year. CFOs got invoices they couldn’t allocate to a team, let alone a product.

But here’s the thing: cloud was just renting someone else’s computer. When a forgotten EC2 instance sat running for six months, it just… sat there. Costing money, but inert. In 2014, the worst-case scenario was leaving a server running over the weekend and wasting $5,000.

AI is not inert. In 2026, the worst-case scenario is that you forgot to fire a virtual McKinsey consultant that’s been hallucinating 40,000 useless Jira tickets for six months and billing you by the token. It doesn’t idle. It makes decisions. You’re employing a ghost that works 24/7, confidently generates nonsense at scale and can’t be fired because nobody remembers who hired it.

By 2018, FinOps was a discipline. By 2020, nobody shipped a cloud workload without a cost allocation tag. The industry built an entire governance layer because untracked cloud spend was killing margins.

AI needs the same thing. Except the curve is steeper.

Cloud reached its governance moment at about 3.4x its initial spend. AI has blown past 5x and there’s no equivalent.

And it’s fundamentally harder than cloud for one reason: AI costs aren’t infrastructure costs. They’re workforce costs wearing a different hat.

The autopilot that closes the books used to be a person. A person on your headcount plan, in your R&D capitalisation model, in your tax credit filing. A person your workforce planning team could forecast. A person who — crucially — would show up to a 1:1, explain what they’d been working on and accept feedback.

Now, that person is your highest-paid contractor: a cloud of maths that never sleeps, refuses to fill out a timesheet, occasionally gaslights the juniors and can’t be put on a PIP. The work is still happening. The cost structure changed from “salary + benefits” to “tokens + API calls.” But nobody updated the workforce model because, honestly, what do you even put in the “role” field? Incorporeal knowledge worker (hallucination-prone)?

How do you do headcount planning when half your “headcount” is an API call? How do you forecast capacity when the thing doing the work bills by the millisecond, scales unpredictably and sometimes just makes things up?

This isn’t a finance problem, and it isn’t an engineering problem. It’s a workforce visibility problem. And it’s the one almost nobody is solving.

The punchline

Sequoia is right. The services market dwarfs the software market. Autopilots will capture work budgets worth six times the tool budgets everyone’s been fighting over.

But someone needs to figure out the books. Not the AI-generated books — the actual books. The ones where you explain to your board what you spent $864K on, which project it connected to, whether it qualifies as CapEx and how much of it you can claim back in R&D credits.

Because right now, we’re collectively wiring millions to a highly opinionated cloud of maths. We can’t see its timesheets. We can’t audit its methodology. And if we ask it to explain its pricing, it will generate a very confident poem about synergy.

The photocopiers are still printing. There are 3,600 of them now. And the parrot just sent another invoice.

Will Hackett is CTO and Co-founder of Flowstate, a workforce planning platform for engineering teams.

Gartner, “Worldwide AI Spending Will Total $2.5 Trillion in 2026,” January 2026. ↩
RGP, “The AI Foundational Divide: From Ambition to Readiness,” December 2025. Survey of 200 US CFOs across technology, healthcare, financial services and retail/CPG ($500M–$10B+ revenue). ↩
MIT 2025 AI Report, as cited in Stanford HAI research. ↩