---
title: "An Interview With Prudence Haircut, the Bank Cost Manager Who Has to Price Checking the Machine"
canonical: "https://current.tinyfish.ai/issue/55/vision/article/65094/an-interview-with-prudence-haircut-the-bank-cost"
publication: "TinyFish"
publisher: "TinyFish"
language: "en-US"
guide: "https://current.tinyfish.ai/llms.txt"
issue: 55
section: "vision"
section_name: "Vision"
published: 2026-11-13
word_count: 1833
source_links: 15
---

# An Interview With Prudence Haircut, the Bank Cost Manager Who Has to Price Checking the Machine

*November 13, 2026*

[Read on the website](<https://current.tinyfish.ai/issue/55/vision/article/65094/an-interview-with-prudence-haircut-the-bank-cost>)

On Friday, November 13, three days after Reuters reported that banks are handing their AI bills to new teams some call **AI FinOps**,[^1] we sat down with someone whose job barely has a description yet. (FinOps is the discipline that grew up around taming cloud bills. Its newest assignment is the meter running on AI agents.)

Prudence Haircut runs AI cost management at a mid-sized bank. The bank, like Prudence, exists only for the length of this conversation. She considers this a professional advantage, since imaginary people are rarely invited to steering committees. Before we've even sat down, she points out that a *haircut* is also what bankers call the discount applied to collateral nobody fully trusts. "Same job," she says. "Different collateral."

Reuters' reporting doesn't settle what her role actually is. There's no standard title, no standard reporting line, no agreed set of powers. The FinOps Foundation spreads the work across engineering, finance, procurement, product, and security,[^2] which in practice means everyone has a stake and nobody has a desk.

What *is* settled is that the meter is becoming visible. Some JPMorgan engineers reportedly work under a $2,000 monthly limit on an AI coding tool, with the option to ask for more.[^3] McKinsey has written up a transaction-monitoring pipeline that cost about €2.50 per case.[^4] Prices like these are starting to sit right next to the work, where the people doing it can see them.

Prudence arrives with a paper notebook and a laptop covered in stickers from conferences she did not attend. She has the expression of someone recently asked to explain a cost line called "other."

---

## Reuters says banks are creating teams like yours. What do you actually do all day?

**Prudence:** Mostly I translate. Engineering thinks in tokens per minute, tokens being the little chunks of text a model reads, writes, and bills by. Finance thinks in euros per month. The provider's rate limits are written in the first language and the budget in the second.[^2] I stand in the gap holding a dictionary that's still being written.

Everyone assumes I spend my days haggling over price. I'd love that. Haggling is clean. But the advertised token rate is nearly the least interesting number on the invoice. A request drags along system instructions, conversation history, retrieved documents, and sometimes reasoning you're billed for and never get to see.[^5]

What the user typed is the tip of the... no. I refuse to say iceberg. It's the part of the invoice you were shown.

## There's a figure going around: AI spending reaching 20% of labor costs. Is that your nightmare number?

**Prudence:** It's a "could" number. I keep it around the way you keep a smoke alarm. Loud, useful, and not evidence of a fire.

Reuters attributes it to McKinsey and frames it around tokens powering agents.[^1] The underlying review says tokens, compute, and AI operations together "can reach as much as" 20 percent.[^6] Those are different fences around different fields. And neither one tells you what the average bank spends today.

## Banks first tried routing simple questions to cheap models. Why isn't that enough?

**Prudence:** Because routing assumes a question has a size. You decide it's small, you send it to the small model, everyone goes home. Fine when one person asks one thing.

Once agents start talking to each other, consumption piles up in places nobody routed. Vik Sohoni at McKinsey told Reuters that's where the approach loses its grip.[^1] Picture a dinner where everyone ordered soup and the bill still comes to four hundred euros, because the soup kept ordering for the table.

Routing still has its uses. **The unit you're routing just stopped being the unit you're paying for.**

## So the price goes on a dashboard. What happens to people when they can see the meter?

**Prudence:** They turn into taxi passengers. Eyes on the meter, and nobody asks for the scenic route.

At Revolut, David Abitbol, who leads AI operations, described strict budgets pushing engineers to save their tokens for the hard problems.[^7] Sounds efficient. Except the scenic route was where the debugging lived, along with the internal documentation and all the slightly boring work where people learn how a system actually behaves.

Revolut later moved to cheaper self-hosted models. Abitbol reports spending around a tenth of the old invoice, and I'd note those are the company's own numbers.[^8] After the move, engineers leaned on assistance more freely for exactly that surrounding work.

The *how* is what I keep coming back to. He didn't send a memo about thrift. He changed the infrastructure so people had less reason to ration. Nobody in history has been memo'd into curiosity.

## Where does a visible price do the most damage?

**Prudence:** On the line labeled *checking*.

Take that McKinsey case. One large European bank was spending about €2.50 per case before optimization, and roughly half of it went to automated quality control.[^4] Now imagine a brand-new cost manager with a quarterly target opening that breakdown. Half the money pays a machine to watch another machine. It looks like hiring a bureaucrat to supervise a bureaucrat. Every instinct says trim.

To be fair, the authors do list sampling those checks as a possible saving, alongside cheaper models, caching, and smaller handoffs between agents. Then they add the sentence I'd like tattooed inside every FinOps lead's eyelids. In a regulated function, checking less often is a *control* decision. Engineering doesn't get to make it alone.

> To the spreadsheet, a cached prompt and a skipped check save exactly the same euro.

That's what visibility does. The cost of checking becomes very easy to see, while the value of checking stays invisible, right up until the morning a check catches something. Insurance always looks like waste on a quiet Tuesday. Just this week, the ECB reminded banks to keep sufficient budgets for maintaining critical services through disruption.[^11] Nobody needs to write that sentence about things that obviously look worth paying for.

And not all of the checking is done by machines. In one banking customer-service workflow McKinsey studied, tokens were often 20 to 25 percent of variable run costs, while human oversight was 70 to 75 percent.[^10] Early estimates, one setting, salt to taste. Still, stare only at tokens and you're optimizing the quarter of the bill that happens to be easiest to read.

So the number I actually want is **cost per successful outcome**. The FinOps Foundation defines it as the total cost of every run, failures included, divided by the runs that meet the quality bar. It also recommends reporting that number as a range.[^12] Failed runs are part of the price of the good ones, the way rehearsals are part of the price of a concert.

> Cut the checks and your cost per run falls. Your cost per *right* answer? You no longer know. You've stopped measuring the thing that would tell you.

## Then who pays for the checking?

**Prudence:** *[long exhale]* My favorite question. I say that the way a dentist says "interesting."

The FinOps Foundation separates two options. Showback means reporting costs to a team, so they can see them. Chargeback means posting those costs to the team's actual budget, so they feel them. The Foundation says neither is inherently more mature.[^13] It also admits there's no generally accepted way to allocate costs across multiple agents, and that even identifying who consumed the output is hard.[^14]

So picture a quality-control agent protecting three departments. Charge it to whoever triggered the run and you've invented a tax on being careful. Hold it centrally and it becomes platform overhead, which is the first line anyone wants cut, usually by someone who has never read it.

I don't have a clean answer. I have a very long meeting. Recurring, no end date.

## In June, a JPMorgan executive told Semafor some employees spent more on tokens than their salaries.[^15] Now some engineers face a $2,000 cap. Someone asks you to raise theirs. Yes or no?

**Prudence:** Tell me what the extra runs bought. Then ask me again.

*[She writes something in the notebook. She does not show us.]*

## Fair. Zoom out for us. Where does this job go once the meter is everywhere?

**Prudence:** Hm. Honestly? I think it drifts away from cost and toward arguments about what "done" means.

McKinsey points to roughly 60 percent task-time reductions across seven production systems, and also warns that saved time doesn't automatically turn into money.[^6] Reuters says banks increasingly want revenue growth to justify the spend.[^1] So the price of a run is known. What the run bought is still being litigated.

One possibility, and I hold it loosely, is that the useful number moves. Instead of the run price, you'd measure what it costs to reach the right state, verify it, keep the evidence, and recover when things go wrong. That's a framing I've seen floated, and nobody has standardized it. But it's a far more honest map of where a bank's money goes.

What worries me is quieter. Delegated work might outpace an institution's ability to verify and repair it. Or it might not. Cheaper checks, sensible sampling, and fewer handoffs can all shrink that load. At Commerzbank, Hans-Georg Beyer describes humans keeping the judgments while AI gathers information and writes more structured reports.[^16] That's a design where checking gets *easier*.

If my job ends up being about that, I'll be delighted. If it ends up being about rationing, I'll be very good at it and slightly sad.

## Last one. It's Friday the 13th. Superstitious?

**Prudence:** Only about lines labeled "other."

*[She closes the notebook.]* Also, I'm fictional, so this whole run was paid for in your readers' attention. Do let me know whether it met the quality bar. Failures included.

[^1]: Tatiana Bautzer, Reuters, "Banks seek to rein in soaring AI agent costs, McKinsey says," November 10, 2026. https://www.investing.com/news/economy-news/banks-seek-to-rein-in-soaring-ai-agent-costs-mckinsey-says-5015595
[^2]: FinOps Foundation, "Managing AI Token Costs from Direct Model Provider APIs." https://www.finops.org/wg/token-economics-saas/
[^3]: Alice Tecotzky, Business Insider, September 17, 2026. https://www.businessinsider.com/jpmorgan-claude-spending-limit-security-engineers-2026-9
[^4]: McKinsey, "Agentic AI in financial-crime prevention: Lessons from winners," October 2026. https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/agentic-ai-in-financial-crime-prevention-lessons-from-winners
[^5]: FinOps Foundation, "GenAI FinOps: How Token Pricing Really Works." https://www.finops.org/wg/genai-finops-how-token-pricing-really-works/
[^6]: McKinsey, "Global Banking Annual Review: The AI disruption dividend," November 2026. https://www.mckinsey.com/~/media/mckinsey/industries/financial%20services/our%20insights/global%20banking%20annual%20review%202026%20part%202/global-banking-annual-review-2026-the-ai-disruption-dividend.pdf
[^7]: David Abitbol, "Shipping like a startup, regulated like a bank," AI Engineer NYC, October 13, 2026. https://ai.engineer/nyc/2026/talks/revolut
[^8]: The Register, November 3, 2026. https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345
[^10]: McKinsey, "Where AI agents pay off: A practical guide to the economics of agentic workflows," August 24, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/where-AI-agents-pay-off-a-practical-guide-to-the-economics-of-agentic-workflows
[^11]: ECB Banking Supervision, written overview for the exchange of views with the Eurogroup, November 9, 2026. https://www.bankingsupervision.europa.eu/ecb/pub/pdf/ssm.written_overview261109.en.pdf
[^12]: FinOps Foundation, "FinOps for AI Overview." https://www.finops.org/wg/finops-for-ai-overview/
[^13]: FinOps Foundation, "Invoicing & Chargeback." https://www.finops.org/framework/capabilities/invoicing-chargeback/
[^14]: FinOps Foundation, "FinOps for AI: Technology Category." https://www.finops.org/framework/technology-categories/ai/
[^15]: J.D. Capelouto, Semafor, "AI token costs are exceeding some employees' salaries," June 3, 2026. https://www.semafor.com/article/06/03/2026/ai-token-costs-are-exceeding-some-employees-salaries
[^16]: McKinsey, "Agentic AI in transaction monitoring: How banks can scale AI in compliance," October 26, 2026. https://www.mckinsey.com/capabilities/risk-and-resilience/our-insights/agentic-ai-in-transaction-monitoring-how-banks-can-scale-ai-in-compliance

## More

- [Vision — Issue 55](<https://current.tinyfish.ai/issue/55/vision/markdown>)
- [Issue 55](<https://current.tinyfish.ai/issue/55/markdown>)
- [Publication guide](<https://current.tinyfish.ai/llms.txt>)