# AI for private equity research, stage by stage through a small deal

AI for private equity is mostly sold to large funds for reading data rooms and drafting memos. This page is for the other end of the market: search funds, independent sponsors and small funds with one or two people on a deal and no analyst bench. It follows a deal from thesis to close and, for each stage, lists the questions that come up, the public datasets that answer them, and where an AI tool helps or quietly gets things wrong. Everything here works by hand. The tool we make appears near the end.

Stanford's 2026 study of US and Canadian search funds [reports](https://www.gsb.stanford.edu/insights/search-funds-keep-offering-proven-path-ownership) that acquiring a company typically takes around 20 months, that 58% of search funds since 1996 have made an acquisition, and that the median purchase price in 2024 and 2025 was $16 million. Much of those 20 months is desk research.

## Building a thesis on an industry

At this stage the questions are about the field, not a company: how many businesses operate in the industry and region, how many are large enough to buy, what they pay workers, whether labor is getting scarcer, and whether the count is growing.

- **Census County Business Patterns (CBP)** gives establishments, employment and payroll by industry code for the nation, states, counties, metro areas and ZIP codes. It [covers](https://www.census.gov/programs-surveys/cbp/about.html) establishments with paid employees only. The [latest year is 2023](https://www.census.gov/programs-surveys/cbp/data.html), and Census says it has no planned release date for 2024. The [Census API guide](/guides/census-api) shows how to pull it county by county.
- **Nonemployer Statistics** [counts](https://www.census.gov/programs-surveys/nonemployer-statistics.html) businesses with no paid employees. Owner-operators without staff are rarely targets, so keep them out of the count.
- **Statistics of U.S. Businesses** [reports](https://www.census.gov/programs-surveys/susb.html) firms and establishments by enterprise size. This is the dataset for a fragmentation argument: how many firms have 20 to 99 employees, against how many have 500 or more.
- **BLS QCEW** [publishes](https://www.bls.gov/cew/) quarterly employment and wages by county and industry, covering more than 95% of US jobs. First-quarter 2026 data came out in August 2026, so it is fresher than CBP.
- **BLS OEWS** [estimates](https://www.bls.gov/oes/) wages for about 830 occupations down to metro areas, most recently for May 2025. For an HVAC thesis, the BLS outlook page [lists](https://www.bls.gov/ooh/installation-maintenance-and-repair/heating-air-conditioning-and-refrigeration-mechanics-and-installers.htm) 440,900 mechanic and installer jobs in 2025, a median wage of $61,010, and 11% projected growth from 2025 to 2035.

AI is good at the mechanics here, such as finding the industry code and pulling a dozen counties into one table. It misleads in a specific way. CBP counts establishments, and Census [defines](https://www.census.gov/programs-surveys/susb/about/glossary.html) an establishment as a single physical location, classified by its major activity. A contractor with four branches is four establishments. A model asked "how many HVAC companies are there" will often hand back the establishment count as a company count, which overstates the number of targets. Since 2007 CBP has also added deliberate [noise](https://www.census.gov/programs-surveys/cbp/about.html) to protect individual businesses, so small-county cells are approximate.

Here is a real answer to the first thesis question, from [Answers](/answers), our one-question service, asked on 2026-11-19 and shown unedited.

**Question:** How many plumbing, heating and air-conditioning contractor establishments (NAICS 238220) were there in the United States in the latest Census County Business Patterns, and what was their annual payroll and employment?

**Answer:** "In the latest Census County Business Patterns (2023), U.S. NAICS 238220 had 111,207 establishments, 1,214,761 employees, and $89,519,407,000 in annual payroll ($89,519,407 thousand)."

Check result: supported. Sources consulted: 23.

| Source | Quote the answer rests on | Check |
|---|---|---|
| [Census CBP 2023 table, NAICS 238220, United States](https://data.census.gov/api/access/data/table?codeset=naics~238220&g=010XX00US&id=CBP2023.CB2300CBP) | "Plumbing, heating, and air-conditioning contractors ... 1214761 ... All establishments, 111207 ... 89519407" (excerpt of the table row) | verified |
| [County Business Patterns Data](https://www.census.gov/programs-surveys/cbp/data.html) | "The most current CBP data are for reference year 2023." | verified |

The answer says "establishments", which is the right word. To turn counts like these into a market size, see [how to calculate market size](/guides/how-to-calculate-market-size).

## Screening a target before the LOI

Once a target is in view, the questions narrow: who else is buying in this industry and what they pay, what listed peers report for margins and growth, and whether lenders finance deals like this one and how those loans have performed.

- **SEC filings of listed peers.** EDGAR full-text search [covers](https://www.sec.gov/edgar/search/efts-faq.html) every electronic filing since 2001, exhibits included, so a listed consolidator's acquisition disclosures can be found by searching an industry term. The XBRL [frames and company facts APIs](https://www.sec.gov/search-filings/edgar-application-programming-interfaces) return reported financials for one company or one figure across every filer, with no API key.
- **Form D.** A company raising money under an exemption must [file](https://www.sec.gov/resources-small-businesses/exempt-offerings/filing-form-d-notice) a Form D within 15 days of the first sale. A recent notice from a known platform in your space tells you a sponsor-backed buyer is raising money. The SEC's quarterly [Form D data sets](https://www.sec.gov/data-research/sec-markets-data/form-d-data-sets) are published as filed, and the SEC says it cannot guarantee their accuracy.
- **SBA 7(a) loan data.** SBA's loan-level [FOIA file](https://data.sba.gov/dataset/7a-504-foia) is updated quarterly, about a month after each quarter ends. Our tabulation of it puts the [median SBA loan to buy a business](/guides/sba-7a-acquisition-loan-data) at $968,000 in fiscal 2026. Our [study of 433,292 7(a) loans](/guides/sba-7a-loan-default-rates) approved in fiscal 2010 to 2018 found charge-off rates of 11.6% for residential remodelers and 9.9% for limited-service restaurants, against 2.7% for dentists and 2.2% for veterinary practices. Lenders read that risk into the terms they offer.

For price, the [valuation multiples by industry](/guides/small-business-valuation-multiples) guide collects closed-deal figures, and the [letter of intent guide](/guides/letter-of-intent-to-buy-a-business) covers what the LOI should commit to.

AI helps most with the filings, and that is also where it has failed in public. In the 2023 [FinanceBench](https://arxiv.org/abs/2311.11944) test of 10,231 questions about public-company filings, one leading model with a retrieval system "incorrectly answered or refused to answer 81% of questions." Models have changed since, but the lesson holds: check any number a model gives you against the filing it came from. A second trap is the private target itself. It files nothing with the SEC, so any revenue figure a model offers for it did not come from a filing; ask where it did come from. If you need private-company deal comps, that is what a [PitchBook seat](/alternatives/pitchbook-pricing) is for. For a structured map of rivals, see [AI competitive analysis](/ai-competitive-analysis).

## Confirmatory work after the LOI

After signing, most of the work is in the data room: a quality of earnings report from an accountant, contracts and leases from a lawyer, and the lender's checklist. The [commercial due diligence guide](/guides/commercial-due-diligence-small-acquisition) explains what each costs and who orders it, and the [due diligence checklist](/guides/due-diligence-checklist-for-buying-a-business) lists the documents.

The public record cross-checks what the seller tells you:

- **OSHA inspections.** The establishment search [finds](https://www.osha.gov/ords/imis/establishment.html) inspections and citations by company name. OSHA warns that citations on open cases can still change through settlement or contest, and that different companies can share similar names, so match on address.
- **EPA enforcement.** ECHO [covers](https://echo.epa.gov/) facility compliance under the Clean Air Act, Clean Water Act, RCRA and the Safe Drinking Water Act, plus formal federal enforcement cases.
- **Liens.** A UCC search shows who has a claim on the business's assets; the [UCC lien guide](/guides/what-is-a-ucc-lien) explains how to run one by state.

That name warning is the main way AI misleads at this stage: a model can merge unrelated companies with similar names into one record. Ask for the address on every hit. For a full public-record screen on a named company, see [AI due diligence](/ai-due-diligence).

## Where a research run fits

Hyperresearch does this desk work as one written report. You describe the deal question as you would brief an analyst, and a run reads the sources above that it has connectors for (Census, BLS, SBA loan data, SEC EDGAR with Form D, OSHA and wage-and-hour cases, EPA ECHO, among the [120+ on the sources page](/sources)), plus the open web. It checks cited claims against their sources and corrects what does not hold before you see the report. Deep runs also carry a [verification receipt](/verification-receipt) with a verdict for each citation, and every source stays in your vault for the next question on the same deal. On the [DRACO benchmark](/benchmarks/draco), Hyperresearch Deep scored 77.7. The [HVAC acquisition sample](/samples/buying-an-hvac-business) is an unedited Deep report on a roughly $2 million revenue company, from 213 sources.

The limits matter for this reader. A run does not read PitchBook, Crunchbase or your data room, and it reads public pages only, logged out. It does not value the business or tell you whether to buy; it gathers and checks sources, and it is not investment, legal or tax advice. Advisors running deals for clients may find [AI for consultants](/ai-for-consultants) closer to their case, and [AI market research](/ai-market-research) covers sizing outside a deal.

## Starting on a live deal

Your first Light report is free when you sign up, with no card. A good first question is the thesis check for the industry on your desk: the establishment counts for your region, wage trends for the key trade, and who has been buying. Sign up at the [console](https://console.hyperresearch.ai/login); [pricing](/pricing) lists run sizes.
