- Financial models are named by buyer programs, and models still fail at workbook work: the best passed 38.3% of attempts in a November 2026 benchmark.
- Models the firm built for itself, or rebuilt from public data, are the easiest to license. Client models need the client’s consent and a check of the engagement letter.
- Anything holding material nonpublic information about a listed company or a live deal stays out until it is public or stale.
- No buyer has published a price for a model library on its own. Published ranges cover a whole company’s operational data.
Can a firm license its financial models and workbooks to AI companies?
Yes, with two conditions. The firm has to own the model, and the model cannot carry information the firm holds in confidence for someone else. A three-statement template your team built, an FP&A forecast of your own business, or a valuation model rebuilt from public filings usually passes both tests. A model built for a client, with the client’s numbers in it, usually passes neither until the client agrees.
The table shows who asks for models today, what each source says, and the rules that decide what can go into a license. Every row was read on 19 November 2026.
| Buyer or source | Stated range or rule | Records asked for | Date checked |
|---|---|---|---|
| Handshake (data partnerships page) | No price published for models | For finance and insurance: "Financial models, process documentation, underwriting workflows" | 19 Nov 2026 |
| Labelbox Prism (categories page) | Financial reasoning traces flagged URGENT; no price per model | Underwriting, valuation, forecasting, investing and M&A workflows "with key assumptions and rationale" | 19 Nov 2026 |
| micro1 (data partnerships page) | $100K to over $3M, for a company’s operational data as a whole | For finance and accounting: processes, SOPs, reconciliations, approvals, reporting workflows | 19 Nov 2026 |
| AfterQuery (page for individual sellers) | Paused: "not taking new" financial models from individuals | Workbooks an individual built; the company route is separate | 19 Nov 2026 |
| GDP.xlsx benchmark (Surge AI, arXiv) | Best model passes 38.3% of attempts; median 19.3% | 70 professional requests over complete workbooks, 67 held back | Published Nov 2026; read 19 Nov 2026 |
| AICPA Code 1.700.001 | No disclosure of confidential client information without the client’s specific consent | Any model or workbook holding a client’s data | 19 Nov 2026 |
| SEC insider trading rules (Investor.gov) | No trading or tipping on material nonpublic information | Models of listed companies or live deals built on inside information | 19 Nov 2026 |
Sources: Handshake, Prism, micro1 and AfterQuery are quoted from their own pages as read on 19 November 2026; those are programme intake pages, so they are not linked. The benchmark is GDP.xlsx on arXiv. The rules are the AICPA Confidential Client Information Rule and the SEC’s insider trading summary.
This page is general information, not legal advice.
Do AI companies buy financial models and Excel workbooks from finance firms?
They ask for them by name. Handshake’s data partnerships page lists "Financial models, process documentation, underwriting workflows" as what it wants from finance and insurance companies. Labelbox Prism, which publishes a demand board for its buyers, marks "Financial reasoning traces" as URGENT and describes them as underwriting, valuation, forecasting, investing and M&A workflows with the assumptions and rationale behind them. micro1 names financial processes, reconciliations, approvals and reporting workflows for finance and accounting teams.
What the labs do with a workbook is easiest to see in Handshake’s own demo. In a November 2026 interview on Latent Space, the finance segment showed an agent that "has to find its assignment in a bank’s email, work out which files govern the deal, and deliver a workbook", and then a banker grading that spreadsheet. A model library teaches exactly that last step: how a professional lays out drivers, links tabs and checks the result.
One door has closed. On 19 November 2026 AfterQuery’s page for individual sellers said it was "not taking new codebases, CAD designs or financial models right now." That pause covers analysts selling models they built on their own. It says nothing about a firm licensing a library it owns, which goes through a company agreement.
Why do AI labs want real workbooks now?
Because models still fail at them. GDP.xlsx, a benchmark Surge AI published in November 2026, gave 20 model setups "70 realistic workplace requests written by professionals in 13 domains", each with the full workbook or CSV file. The top model "passes only 38.3% of attempts, and the median score is 19.3%". The tasks were not hard sums. They were the reading problems an analyst solves every day: finding the tab that governs an answer, noticing a filter that hides rows, reading a comment that changes a number.
Finance-specific tests show the same gap. TechTimes reported on 1 October 2026 that "the best artificial intelligence model in the world scores 51% on a realistic private-equity due-diligence simulation", built from 88 tasks taken from anonymized real transactions. The company behind that benchmark says four of the five top US labs are its customers, as reported by TechTimes.
Two things follow for a firm that holds models. First, labs need real workbooks with the mess left in: hard-coded overrides, stale tabs, version notes, the analyst’s check sums. A clean template teaches less than a model that was used on a live decision. Second, labs need private test sets. GDP.xlsx holds "the other 67" tasks back "to protect the leaderboard from contamination". A set of real models and the questions an expert asked of them can be licensed for evaluation only, which leaves the firm’s files out of training data.
How much would AI labs pay for a firm’s library of professional Excel financial models?
No buyer has published a price for a library of financial models on its own, and we found no reported deal for one as of 19 November 2026. Any number you see for a model library alone is an estimate or a negotiating position.
The published figures cover a company’s operational data as a whole. micro1 says its payments "typically" range "from $100K to over $3M". Those figures price a whole program of records, usually several systems and years of history, not a folder of spreadsheets. A model library is one part of that picture.
What moves the value of the model part: how many distinct models there are, whether each has its version history, whether the inputs and the memo or board pack the model fed are kept with it, and how unusual the work is. A private credit underwriting model with the credit memo and the loan’s later performance is worth more than ten copies of a standard LBO template. For deal-side records such as data rooms, memos and credit files, see the guide to licensing deal and credit data from investment firms. For how deals are structured (one-time fee, recurring delivery, evaluation-only), see data licensing pricing models.
If you want a figure for your own records, check what your data may be worth. The estimate starts from a description, not from your files.
Which models can we license, and which stay out?
Sort the library before anyone talks price. The question for each model is who owns it and whose information is inside it. The table gives the usual answer for each kind of workbook a finance firm holds. Your engagement letters and client contracts can change any row.
| Kind of model | Who usually owns it | Whose information is inside | Licensable? |
|---|---|---|---|
| Firm templates and training models | The firm | None, or sample data | Usually yes |
| The firm’s own budget, forecast and board models | The firm | The firm’s own figures | Yes, if the owners agree and figures are scaled or masked |
| Models rebuilt from public filings | The firm | Public information | Usually yes |
| Client models built under an engagement | Depends on the engagement letter | The client’s confidential figures | Only with client consent, or with client data removed and the structure kept |
| Deal models for live or recent transactions | Firm or client, by contract | Possibly material nonpublic information | No, until the deal is public and the information stale |
| Models built from tax returns | The preparer’s file | Taxpayer return information | Not without the taxpayer’s consent under Section 7216 |
The common middle path for client models is to keep the structure and replace the data: the tab layout, the formulas, the driver logic and the analyst’s notes stay, and every client figure and name is replaced or scaled. Buyers often value the structure and the reasoning more than the numbers. Agree in the license who does that work and who checks it.
Who owns a financial model built for a client?
Start with the engagement letter. Many letters say that deliverables belong to the client, or that the firm keeps its templates and tools while the client owns the finished output. If the letter is silent, US copyright law decides who owns the spreadsheet itself. The Copyright Office’s Circular 30 explains that for a work made for hire "the party that hired the individual is considered both the author and the copyright owner of the work", and that "there are two situations in which a work made for hire is produced": an employee’s work in the regular course of the job, and certain types of work made under an express written agreement. In plain terms: your employees’ models usually belong to the firm, while a model you built as an outside adviser may belong to you unless the contract says otherwise.
Owning the file is not the end of it. The client’s numbers inside the model are the client’s confidential information whoever owns the spreadsheet. For CPA firms the rule is direct. The AICPA Code says: "A member in public practice shall not disclose any confidential client information without the specific consent of the client." Its plain meaning: a CPA firm needs each client’s consent before that client’s information goes to anyone, including an AI company. The guide to licensing audit workpapers covers how the AICPA applies this to data sales.
Where a model was built from a client’s tax return, a second rule applies. 26 U.S.C. 7216 covers "any person who is engaged in the business of preparing, or providing services in connection with the preparation of" tax returns, and makes it a crime to disclose return information knowingly or recklessly without consent. Keep return-based models out unless each taxpayer has consented in the form the regulations require.
Is it legal to license models that hold material nonpublic information?
Treat it as a no until the information is public or stale. The SEC’s Investor.gov summary says: "Illegal insider trading refers generally to buying or selling a security, in breach of a fiduciary duty or other relationship of trust and confidence, on the basis of material, nonpublic information about the security." It adds that violations "may also include ‘tipping’ such information". In plain terms: if a model shows a listed company’s unannounced results, a pending deal or a financing that has not been disclosed, handing it to a third party can create insider trading risk for whoever receives it, and you may owe a duty of confidence to the company it describes.
Most libraries handle this with a time rule and a scope rule. The time rule: only models for deals that closed and were announced, or periods already reported, and often only those older than a set number of years. The scope rule: no live mandates, no models from wall-crossed work, and no models built on a data room still under an NDA that limits use. Write both rules into the license so the buyer knows what it is not getting.
This page is general information, not legal advice.
How can an FP&A consultancy license its spreadsheets and financial models for AI training?
We help you license your data. Origin is an independent data partner. Your firm decides what is shared and signs any license itself, and a buyer may pay Origin when a deal closes. The work on the firm’s side usually runs in five steps.
- Count the library: number of models by type (three-statement, budget, cash flow, valuation, LBO, credit, unit economics), years covered, and whether version history and the related memo or deck survive.
- Sort each model into the ownership table above. Pull out anything with live deal information or tax return data first.
- Read the engagement letters for the client models you want to include. List which clients need a consent request and which letters already give the firm rights to its working papers and tools.
- Decide the use: training, evaluation only, or both. Evaluation-only licenses keep files out of training data and suit firms worried about their own know-how. See how to structure a data licensing agreement.
- Prepare a de-identified sample of five to ten models, with client names and figures replaced, so a buyer can judge quality before anyone sees the full set. See how to de-identify company data before licensing.
Fractional CFOs and FP&A firms often hold more than models: month-end close files, board packs and the approvals behind them. Those are covered in licensing ERP and accounting records. Your CFO or finance lead can also read the CFO guide to data licensing for how license revenue is booked. When you are ready, start with a description of what you hold.
Which companies pay for real financial models and spreadsheets to train AI?
On 19 November 2026 these sources named financial models or finance workbooks in public: Handshake (financial models for finance and insurance companies), Labelbox Prism (financial reasoning traces, flagged urgent), micro1 (finance and accounting processes) and AfterQuery (which paused individual model submissions but runs a separate company route). Benchmark builders such as Surge AI, which published GDP.xlsx, and finance-specific environment builders reported by the press also use real workbooks to build tests. A name on this list is a market fact, not a recommendation.
Compare at least two routes before you sign, and check every offer against what to review in a data licensing offer or the offer checker.
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
Who buys investment memos, deal files and diligence records from an investment firm for AI training?
Buyer programs that name finance records include Handshake, which lists financial models and underwriting workflows, and Labelbox Prism, which flags financial reasoning traces such as valuation and M&A workflows as urgent. The most useful records are the full chain for a closed deal: data room, model and memo together. Records from live deals or holding inside information stay out. Origin’s guide to deal and credit data covers investment firms in detail.
How much would AI companies pay a hedge fund or private equity firm for its research and deal data?
No buyer has published a price for a fund’s research or deal files on their own. Published ranges cover a company’s operational data as a whole, such as micro1’s $100K to over $3M. The value of a fund’s records turns on how many closed deals have the full chain (model, memo, outcome), how far back they go, and whether limited partner and portfolio company agreements allow the use.
Which companies help investment firms and lenders license their financial models and data to AI labs?
Some buyer programs deal with firms directly: Handshake, micro1 and Labelbox Prism publish pages for companies. Origin is an independent data partner that helps a firm describe its records, sort what it owns from what it holds for clients, and compare offers. A buyer may pay Origin when a deal closes. The firm signs any license itself and decides what is shared.
An AI data broker offered to buy our firm’s old financial models. Is that legit, and what should we check before we sign?
Ask who the end buyer is, what use the license allows (training, evaluation or resale), and whether payment is fixed in writing before delivery. One commenter on a public post about such offers said they start high and "end up offering 10-100x less". Check that client models, tax-based models and anything with inside information are excluded, and that resale to other labs needs your approval.
I am the CPA for an FP&A consultancy. Can the firm license client models if the names are removed?
Removing names may not be enough. The AICPA rule bars disclosing confidential client information without the client’s specific consent, and figures, structure and notes can identify a client even without its name. The safer routes are client consent, or a version that keeps only layout and formulas with every client figure replaced. Return-based models also need Section 7216 consent. This is general information, not legal advice.
Can we license our models for AI evaluation only, without them being used for training?
Yes, and buyers already work this way. Benchmarks keep test tasks private: GDP.xlsx released 3 of its 70 tasks and held 67 back so models cannot train on them. An evaluation-only license lets a lab test models against your workbooks and expert questions, with no training use, no copies kept after the term, and deletion confirmed in writing. Prices for evaluation sets are not published either.
Do buyers want polished templates or the models our analysts actually used?
The models people actually used. The benchmarks that show models failing test real files with hidden rows, stale tabs, comments that change a number and drivers spread across sheets. A template sold on a marketplace teaches little. A model with its version history, the memo it fed and the later outcome shows how a professional decided, which is what Prism’s urgent category describes.
Sources and further reading
- arXiv: GDP.xlsx, Benchmarking Frontier Models on Professional Workbooks (Surge AI, November 2026)
- Latent Space: What Frontier Labs Buy, a data deep dive with Handshake AI (9 November 2026)
- TechTimes: Finance AI Tops Out at 51% on Due Diligence (1 October 2026)
- LinkedIn: post on an offer of up to $2M for lease-administration procedures (20 October 2026), with comments
- AICPA Code of Professional Conduct, 1.700.001 Confidential Client Information Rule
- SEC Investor.gov: Insider Trading
- US Copyright Office: Circular 30, Works Made for Hire
- 26 U.S.C. 7216: Disclosure or use of information by preparers of returns (LII)
