What Bankers Need (It Isn’t a Model)
A $1.5 billion valuation for an agent in every banker’s Excel. The bankers are buying the fastest junior analyst in history. What they’ll need, the first time a number gets read back to them, is the file.
The Raise — The News, Straight
Model ML announced a $250 million Series B at a $1.5 billion valuation today, co-led by General Atlantic and QED Investors, with Evolution Equity, Airtree, Forestay, Y Combinator, FT Partners and others participating — $350 million raised in total, after a $75 million Series A in November 2025 and a $12 million seed in February 2025.
The company was founded by brothers Chaz and Arnie Englander, who previously built and sold Fat Llama. The product: a personal AI agent for every banker, consultant and investor, running inside the firm’s own systems and living where the work lives — Excel and PowerPoint.
Revenue splits roughly evenly across banks, consulting firms and asset managers. The company reports third-quarter ARR up 12× year over year and daily active users up 26×. It is not alone: Rogo raised $160 million at a $2 billion valuation earlier this year. The market is pricing the banker agent in the billions, plural, before lunch.
Priced in the billions. Plural.
Two banker-agent companies, two yardsticks. Tap a view to isolate it — the scale never moves.
Combined, the two valuations come to $3.5 billion. Combined capital actually raised: $510 million. The gap between those two numbers is the bet.
Model ML — The rounds, in order
Where the money comes from: revenue splits roughly evenly across the three — and the company reports Q3 ARR up 12× year over year, daily active users up 26×. The company’s own figures, reported as such.
What a Bank Actually Produces
A bank does not produce models.
It produces a file.
Every number a banker puts in a deck is a number the bank may one day have to defend — in a fairness opinion, in litigation, in a regulatory exam, in front of a Congressional staffer with a printout. The work product of banking is not the spreadsheet; it is the defensible account of how the spreadsheet came to be: who ran it, on what inputs, under whose instruction, changed by whom, approved by whom.
- Who ran it?A name
- On what inputs?The data, as pulled
- Under whose instruction?The ask, verbatim
- Changed by whom?Every hand
- Approved by whom?The signature
A human junior analyst leaves that trail by existing — emails, versions, a name a deposition can call. An agent leaves whatever its vendor’s logs leave, for as long as the vendor keeps them, in whatever form the vendor chooses.
Read that sentence again, slowly. The trail still exists. What changed is who holds it, how long they keep it, and in what shape they choose to hand it back. That is the whole story of this dispatch — the rest is arithmetic.
The Certification Gap
Model ML’s trust story, as the company tells it, is real and should be recorded fairly: ISO/IEC 27001:2022 certified, SOC 2 Type II audited, agents running inside the firm’s own systems rather than a vendor’s cloud.
But notice what a certification certifies. It certifies the vendor’s processes, examined periodically, by an auditor the vendor engages. It says the kitchen passed inspection last quarter. It says nothing about what was in a specific dish, on a specific night, for a specific client — and the specific dish is the only thing a deposition ever asks about.
What the certificate says
“The kitchen passed inspection last quarter. Records can be kept here. The process is sound.”
What the deposition asks
“What was in this dish, on this night, for this client — and who put it there?”
A SOC 2 report is not a record of what an agent did. It is a record that somebody checked that records could be kept.
Which raises the only practical question in this piece. Take one ordinary agent task — five steps, the kind run a thousand times a week — and ask, under three different record-keeping regimes, what is still there to be read back two years later. Run the test yourself:
One task. Five steps. Three record-keepers.
Switch the regime. The task doesn’t change — only what survives of it. Tap any step for the question a deposition would actually ask.
Fairness note, in the ledger: the test above is an illustration of record-keeping regimes — what each kind of record is for — not a description of any vendor’s internal systems. The regimes are the point; no vendor was measured.
No Counterparty, In a Suit
This is the same shape Verity documented this morning in “The Vultures and the Rodent”: where no independent record exists, the account of what happened gets rebuilt afterwards, by the actors, at reconstruction prices.
Banking is the industry that invented the word counterparty — the party on the other side who keeps their own books precisely so yours can be checked. An agent whose only record is its vendor’s log has no counterparty. The vendor grades its own homework; the bank cannot grade it either, because in the dispute the bank is the defendant.
What bankers need is not a smarter model. It is the sealed record — proposal, decision, rule — kept at the time, by someone who was not in the room.
Now do the last sum.
Graphic A priced the agents. This is the line item nobody priced.
— of the $3.5 billion in combined valuation above, the amount assigned to an independent record of what the agents do.
Press it. The arithmetic is the argument.
How This Ends
The first agent scandal in banking will not be a wrong answer. Models give wrong answers now; banks survive them, because a person signed the page and the file shows the work. The scandal will be an unrecorded answer — a number nobody can reconstruct, in a deal nobody can re-run, defended by a certification that certifies something else.
Bankers, of all people, know how this ends: they’ll have to bail it.
The gate is open to financial services.
Sealed, independent records for agent work — proposal, decision, rule — kept at the time, by a party who is neither the vendor nor the bank. If your firm is putting an agent in the model, put a record under it first.
pilot@verityverdict.com →