AI data labeling financial models is paradise

You get laid off at the beginning of the year from your post-buyside exit op and somehow stumble into this entire “applied AI lab” ecosystem. Which is Silicon Valley jargon for “we’re trying to gig-ify white collar work.” 

Your first project pays $50/hr. Within a week you get “promoted” to reviewer despite not even submitting a completed workflow. Suddenly you’re making $90/hr reviewing LBO models built by other ex-bankers who are equally confused about what any of this actually is.

Life is good.

The project leads answer questions like opposing counsel on a sell-side. 

You ask a simple clarification question: “Should we create separate rubrics for separating operating cases?”

And the response is:

“Great question. We encourage experts to leverage professional judgment while aligning outputs with the broader objectives of the financial reasoning framework.”

A completely non-committal answer specifically engineered to avoid accountability later.

The project lead notifies that the AI lab is putting the entire project on pause for 2 weeks to reassess. Says our hard work may or may not actually be improving model outputs. 

You laugh nervously. It’s fine. It will be back.

2 weeks go by, no response. 3 weeks go by, “we are still re-evaluating, but hey we’ll pay $30 per CIM you can give us.” Yeah bro, I’m gonna break my former NDAs for $30 bucks a pop. 

A month goes by. Slack channel notification from the project lead “No need to DM me or others for a project update we will post here once we have an update. Thanks for the continued patience”

It’s safe to say the project is dead.

No sweat. You find another project that pays $5K a week. Fixed rate too. Paradise again.

Except now they expect 3 full models a day. Meanwhile your first project wanted 2 models a week. Full blown operating models or reverse morris trust analysis (who even does these anymore, pity on the bulge bracket bros/brosephinas as a former MM analyst) that would take a real analyst 2.5+ hours minimum to build properly. Yeah sure bro. You’re gonna get best-of-breed modeling here.

The entire ecosystem incentivizes garbage in and garbage out. Nobody has time to think. The KPI is throughput and model performance based on non-real world workflows.

And let’s be clear here because the Wall Street Journal is like Goldman analysts are personally “training AGI.”

No.

Actual bankers are on the desk. Access is still limited and even the most forward desks aren’t seeing that much value yet beyond basic grunt work acceleration because of their MNPI compliant nuked internal LLM.

What these “applied AI labs” are actually doing is hiring ex-bankers at $175/hr to basically recreate banking deliverables from memory.

You build a CIM. You build a LBO. You build an operating model. Then you write a rubric explaining what “good” looks like. Then you generate prompts. Then you score the AI’s response to your own prompt as the “expert.”

It’s structured IP theft with venture funding.

Then they take the prompts that make the frontier labs look the best and benchmark performance across clients. Then the labs take the highest performing use cases discovered by exhausted former bankers and productize them into enterprise features.

You start digging through your old IB training folders like an archaeologist recovering lost civilization technology. Where is that simple Tesla operating model and DCF? Where is that Nike LBO?

You get your first round of comments and they say your “model intent alignment was insufficient and prompt was overly wordy”

“Uhh.. What is model intent?”

You ask the reviewer for clarification.

The reviewer responds with 4 paragraphs of GPT-generated corporate prose explaining that your “financial reasoning pathway did not sufficiently align with expected expert cognition.”

You stare at the screen for 15 seconds. You hate look up them up on LinkedIn. The reviewer worked at Wells Fargo in the consumer coverage group for 9 months.

At this point everyone is using AI to write prompts, rubric commentary, and review feedback because everyone else clearly is too.

The reviewers use AI to review AI generated rubric commentary for AI models trained on recycled banking templates originally built by a coach from your summer analyst Wall Street Prep bootcamp.

It’s a completely self-referential ecosystem. And somehow there is enough money sloshing around to pay hundreds of ex-bankers $150/hr to participate in this.

But the VC money keeps flowing. You might be helping automate the same industry that psychologically waterboarded you for 2 years straight.

Every daily standup feels like a parody of banking. You got some jabroni who worked back office operations at a mega fund leading calls about “leveraging financial reasoning frameworks at scale.”

He uploaded a broken Excel file yesterday and the Notion page he built for this project is filled with outdated guidance.

The funniest part is there are multiple VPs and MDs on these projects and you immediately remember why you left banking.

Absolute miserable people. Every call has one VP who unmutes himself solely to be aggressive about formatting conventions on synthetic AI training data. “Can we ensure the precedent transaction rationale reflects institutional quality standards.”

Brother this is a fake comp spreading exercise for a fake company being fed into Claude.

There’s also a woman from a bulge bracket who actually seems competent but every time she speaks 4 different bros interrupt her to restate the exact same point in slightly different corporate phrasing. 

You know with 100% certainty these people are violating every OBA compliance policy on earth. Convenient they all leave their last name out of their Slack display name. There is no way the rainmaker MD at a real bank is allowed to moonlight as an “AI financial reasoning expert”

Everyone on these projects talks in this fake corporate hypothetical language too.

“Imagine if your MD sent you a prompt asking for a detailed LBO sensitivity analysis with clear assumptions and output formatting.”

If an MD emailed me an ask it would be:

“Need this by morning.” With a forwarded client email chain below it.

That’s the prompt.

Nobody in real banking communicates like these AI people think they do.

There’s no “desired output specification.”

You get half a CIM from 2021, a broken management model, three contradictory emails, and an associate who’s also getting yelled at by a VP.

Then the VP changes the ask halfway through because the MD had “a thought.”

That’s the workflow.

And somehow these AI companies think they’re replicating investment banking cognition because a former Piper Sandler associate who got PIP’d wrote a 900 word prompt about taking Apple private. Real-world-use-case!

Human-in-the-loop AI?

Every human on these projects are completely out of the loop.

Nobody knows what the client actually wants.
Nobody knows how the benchmark is scored.
The rubric contradicts itself.
The reviewers disagree with each other.
The project leads answer questions like hostage negotiators.
And the frontier lab engineers are probably just looking at aggregate scores in a dashboard somewhere in Palo Alto.

You’re 4 layers removed from reality labeling synthetic CIM prompts at 1AM with a guy named Garrett who used to work at Houlihan and now has “AI Financial Reasoning SME” in his LinkedIn headline.

This is the exit op you dreamed of. 

Literal paradise.

Life is good.