AI for Due Diligence: 6 Checks AI Agents Run Before the Data Room Opens

Due diligence starts before the data room opens. Before a fund signs an NDA, a lender commits or a partner signs a supply agreement, someone has to read what is already public: the 10-K and its notes, the litigation history, the Form 4s, the federal contract records, the patents and the trial registry. That reading is where AI for due diligence earns its keep. AI agents for due diligence can do it in minutes, provided they read those primary records in full and return every finding with its source.

This guide follows six outside-in checks, each for a different deal team. We ran each one as a Valyu DeepResearch task in fast mode on November 9, 2026, then checked the findings at the source: SEC EDGAR, USAspending.gov, Google Patents, ClinicalTrials.gov and PubMed.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for due diligence do

An AI agent for due diligence is a language model that can search and read primary records, such as SEC filings, insider filings, federal awards, patents and trial registries, and return cited findings. It doesn't replace the data room or the deal team's judgement. It builds the outside-in picture and the list of questions to ask once the data room opens.

Your own agents can call Valyu's search when they need one fact. For diligence, DeepResearch does the job properly: it plans the research, reads the full filings and records, follows cross-references and returns a cited brief in the format you ask for.

Figure 1. Six outside-in checks we ran as DeepResearch tasks in fast mode on November 9, 2026, with the sources used, what came back and the question each one raises for the data room.

1. Customer and supplier concentration: who the business depends on

The problem: a credit analyst looking at Super Micro Computer's convertible notes needs to know how much revenue depends on a few customers, and how much supply depends on one vendor.

What we ran: a DeepResearch task on the fiscal 2026 10-K.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
task = valyu.deepresearch.create(
    query="Super Micro Computer: how concentrated are its customers and suppliers, according to its fiscal 2026 10-K?",
    mode="fast",
    research_strategy="Use the 10-K for the fiscal year ended June 30, 2026. Report the share of net sales from its largest customers for fiscal 2024, 2025 and 2026, the supplier share of purchases, and accounts receivable concentration. Keep sales concentration and receivables concentration separate. Quote the filing language for each figure.",
    report_format="A short brief: a table of each concentration figure with fiscal year, basis and quote, then three questions for management.",
    search={"included_sources": ["valyu/valyu-sec-filings"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited brief in markdown

What came back (in 68 seconds): a table of every concentration figure, each with its denominator and the filing's exact words. One customer was 28.1% of fiscal 2026 net sales, up from 20.0% for the largest customer in fiscal 2024; in fiscal 2025, four customers each passed 10%, the largest at 20.9%. One supplier accounted for 63.1% of total purchases, down from 65.4% two years earlier. Three customers held 23.0%, 17.1% and 12.5% of receivables at year end. The brief also found related-party purchases from Ablecom and Compuware, 2.1% of cost of sales, and explained why the four measures can't be added or compared.

What we checked at the source: the 10-K on EDGAR, filed August 31, 2026. Every figure and quote matches.

The catch: the 10-K doesn't name the 28.1% customer or the 63.1% supplier, so no amount of research will. Those names belong on the data room request list.

The decision: model the loss of the largest customer, and ask management who the 28.1% customer is, on what contract terms, and what supply agreement covers the 63.1% supplier.

2. Legal and regulatory exposure: read the notes, not just Item 3

The problem: a compliance team running counterparty review before a supply agreement with Super Micro needs every open legal and regulatory matter.

What we ran: a DeepResearch task that told it where the detail sits.

task = valyu.deepresearch.create(
    query="Super Micro Computer: every open legal proceeding, government investigation and regulatory inquiry disclosed in its fiscal 2026 10-K.",
    mode="fast",
    research_strategy="Item 3 of the 10-K incorporates Note 15 by reference. Read Note 15, Commitments and Contingencies, and the Item 1A risk factors as well as Item 3. List each matter with the parties, court or agency, date, status and whether the company is a named party. Quote the filing for each.",
    report_format="A table of legal and regulatory matters with matter, parties, date, status and quote, then a short summary of exposure for a counterparty review.",
    search={"included_sources": ["valyu/valyu-sec-filings"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 4.3 minutes): a matter-by-matter table drawn from Note 15, sorted into three groups:

• Securities and derivative litigation. Federal securities class actions in the Northern District of California, with the 2024 cases consolidated under a lead plaintiff and a motion to dismiss filed in November 2025, and a second group of 2026 cases consolidated in July 2026. Derivative suits in the Northern District of California, the Santa Clara County Superior Court and the Delaware Court of Chancery, with the company as nominal defendant.
• Government investigations. SEC and Southern District of New York subpoenas and inquiries from the Commerce Department's Office of Export Enforcement, linked to the March 19, 2026 indictment of three former associates for alleged export control violations. The company is not named as a defendant or co-conspirator.
• Patent complaints. US International Trade Commission complaints by Netlist naming the company as a downstream customer of Samsung memory. A Samsung settlement on August 5, 2026 is expected to resolve two of them; a third, filed August 11, 2026 against Micron and its customers, had not been instituted at the filing date.

The brief also noted which matters were closed or expected to resolve, and quoted the 10-K's statement that the company is "unable to estimate the possible loss or range of loss" for the litigation.

What we checked at the source: Note 15 and Item 1A on EDGAR. The case names, dates and quotes match.

The catch: Item 3 is one paragraph that points to Note 15. A quick search returns that paragraph; the research strategy told DeepResearch to read the note and the risk factors as well, which is where the matters are. And the 10-K gives no loss estimate, so exposure can't be sized from public filings.

The decision: escalate to counsel for export-control review before signing, and ask for the status of each investigation and for any accruals the filing doesn't show.

3. Government contract revenue: obligated money, not ceilings

The problem: a growth equity team looking at BigBear.ai needs to know which federal contracts drive revenue and which have ended.

What we ran: a DeepResearch task across USAspending and the 10-K.

task = valyu.deepresearch.create(
    query="BigBear.ai: which federal contracts drive its revenue, and which are still active?",
    mode="fast",
    research_strategy="Use USAspending for BigBear.ai Federal LLC's largest awards, with award ID, agency, amount obligated, description and period of performance. Use the latest 10-K for what the company says about federal revenue and changes in Army program volume. Mark each award as active or ended as of November 2026.",
    report_format="A table of the largest awards with award ID, agency, amount, description, start and end date and status, then a paragraph linking the awards to the 10-K's revenue explanation.",
    search={"included_sources": ["valyu/valyu-usaspending", "valyu/valyu-sec-filings"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 6.5 minutes): eight awards to BigBear.ai Federal, LLC with dates, descriptions and status, each linked to its USAspending page:

It linked the ended Army contract to the 10-K's explanation that 2025 revenue fell $30.6 million, mainly on lower volume on Army programs, and noted that the 10-K anonymises its largest customers, so awards can't be mapped to its concentration table.

What we checked at the source: each award on USAspending.gov. The obligations, ceiling, dates and descriptions match.

The catch: a contract ceiling is not money received. The production contract's $329.5 million ceiling sits next to $7.8 million actually obligated. And three of the five awards "in progress" were $500 to $1,000 minimum-guarantee orders on multi-award contracts, which only matter if task orders follow.

The decision: build the revenue bridge from obligated amounts on live awards, and ask management which task orders are expected under the minimum-guarantee vehicles.

4. Insider selling: one sale, two filings

The problem: an investment committee member wants to know whether Super Micro's insiders have been selling before the fund adds to its position.

What we ran: a DeepResearch task on 2026 insider transactions.

task = valyu.deepresearch.create(
    query="Super Micro Computer: insider buying and selling in 2026.",
    mode="fast",
    research_strategy="Use Form 4 insider transaction data for 2026. Separate open-market sales from awards and gifts. Joint or indirect holdings can put one trade on more than one Form 4: deduplicate by date, shares and price before totalling, and note any Rule 10b5-1 trading plans.",
    report_format="A table of 2026 open-market sales after deduplication, with insider, role, date, shares, value and plan, then the deduplicated total.",
    search={"included_sources": ["valyu/valyu-insider-transactions-US", "valyu/valyu-sec-filings"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 5.7 minutes): one open-market sale in 2026: 200,000 shares for $7,718,584, recorded under both CEO Charles Liang (indirect) and director Sara Liu (direct). The brief counted it once, "not 400,000 shares for roughly $15.4 million", and found no open-market purchases in 2026. It deduplicated a 340,000-share gift and a 20,980-share derivative exercise the same way. It marked the trading-plan status "not established" because it couldn't read the Form 4 documents, rather than guessing.

What we checked at the source: the two Form 4s filed September 8, 2026. They report the same trades on September 3 and 4, and the footnotes confirm a Rule 10b5-1 trading plan adopted by Sara Liu, the CEO's spouse, on May 26, 2026.

The catch: plan status lives in the Form 4 footnotes. DeepResearch said what it couldn't confirm, and one look at the filing settled it.

The decision: record one $7.7 million sale under a pre-arranged plan, and weigh it accordingly.

5. Patent ownership: who actually owns the IP

The problem: a venture capital partner preparing a follow-on investment in Commonwealth Fusion Systems, a private fusion company, needs its granted US patents and their owners.

What we ran: a DeepResearch task on USPTO patents that filtered on the assignee.

task = valyu.deepresearch.create(
    query="Commonwealth Fusion Systems: what granted US patents does it hold, and are any co-owned?",
    mode="fast",
    research_strategy="Use USPTO patents. Include only patents where Commonwealth Fusion Systems is a listed assignee; exclude name matches such as the Commonwealth System of Higher Education. For each patent, give the number, title, grant date and every assignee, and flag patents co-owned with another organisation.",
    report_format="A table of patents with number, title, grant date and assignees, a list of co-owned patents, and questions for a venture investor's diligence.",
    search={"included_sources": ["valyu/valyu-patents"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 3.5 minutes): 10 granted US patents listing Commonwealth Fusion Systems as an assignee, granted between July 2024 and August 2026, covering superconducting magnets and cables, tokamak maintenance, tape quality control and a superconducting wind turbine generator. Five were co-owned with the Massachusetts Institute of Technology, including US 12,700,530 on passive quench protection and US 12,555,711 on distributing forces in high-field magnets. The brief called its count "a documented floor, not a verified total", and flagged two patents whose full assignee lists it hadn't confirmed.

What we checked at the source: every patent on Google Patents. The two DeepResearch flagged, US 12,293,871 (stacked-plate superconducting magnets) and US 12,689,204 (a cable joint for superconducting cables), are also co-assigned to MIT. That makes 7 of the 10 co-owned.

The catch: follow up on what the brief marks as unverified; here it changed the answer from half the portfolio co-owned to 70%. Joint ownership matters: under US law (35 U.S.C. § 262), each joint owner can make, use and sell the invention without the other owners' consent unless an agreement says otherwise, and courts have generally let co-owners license it too.

The decision: ask for the MIT agreements that govern the seven co-owned patents before valuing the IP.

6. Clinical pipeline: does the registry match the story

The problem: a biotech investor reviewing Insilico Medicine, which listed in Hong Kong in December 2025, wants the status of rentosertib, its lead drug for idiopathic pulmonary fibrosis.

What we ran: a DeepResearch task across ClinicalTrials.gov and PubMed.

task = valyu.deepresearch.create(
    query="Insilico Medicine: what is the clinical status of rentosertib for idiopathic pulmonary fibrosis?",
    mode="fast",
    research_strategy="Use ClinicalTrials.gov records for every rentosertib (INS018_055) trial with phase, country, enrolment, status, start date, primary completion date and last update date. Flag any trial whose estimated completion date has passed while it is still listed as recruiting. Use PubMed for published results.",
    report_format="A table of trials with NCT number, phase, country, enrolment, status and dates, a list of flags, and a catalyst timeline.",
    search={"included_sources": ["valyu/valyu-clinical-trials", "valyu/valyu-pubmed"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 4.6 minutes): all four registered trials, the published efficacy signal and two flags:

From the Phase 2a results in Nature Medicine, it reported a mean change in forced vital capacity of +98.4 mL at 60 mg once daily against −20.3 mL for placebo over 12 weeks. The flags: the US trial is still listed as recruiting after its estimated completion date, and the registry plans 40 patients where a June 2026 review describes 60.

What we checked at the source: each record on ClinicalTrials.gov and the Nature Medicine abstract on PubMed. The dates, enrolment and lung-function figures match, and the review does describe 60 patients.

The catch: the Phase 2a's primary endpoint was safety; the lung-function gain was a secondary outcome in about 18 patients per arm. The Phase 3 is the first test of efficacy at scale.

The decision: ask the company for the US trial's enrolment and timeline, and place the Phase 3 readout no earlier than late 2029 in the catalyst calendar.

What the six checks cost

A fast task costs a fixed 25 credits, about $0.16 at pay-as-you-go rates of 1,600 credits for $10, whatever it searches and reads. The six checks cost about $0.94 together. We started them at the same moment, and all six were finished within 7 minutes. For deeper work on one question, standard mode uses 100 credits and heavy 500.

Go further: a full diligence pack in one task

The six checks each answer one question. A deal team's actual output is a document: the first-look brief, the red-flag register, the IC memo. A single DeepResearch task in standard or heavy mode writes these, with files such as an xlsx red-flag register.

task = valyu.deepresearch.create(
    query="Prepare an outside-in due diligence pack on BigBear.ai before data room access.",
    mode="standard",
    research_strategy="Use the latest 10-K and 10-Qs for customer concentration, legal proceedings and contingencies, reading the notes as well as Item 3. Use USAspending for the largest federal awards with obligated amounts, ceilings and periods of performance, Form 4 data for insider activity, deduplicated, and USPTO for patents with every assignee. Cite every finding.",
    report_format="A red-flag brief: Business and customers, Government contracts, Legal and regulatory, Insiders, IP, then 10 questions for the data room. Every finding cited.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-usaspending", "valyu/valyu-insider-transactions-US", "valyu/valyu-patents", "web"]},
    deliverables=["An xlsx red-flag register with finding, source, date, severity and question for management"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six checks: read the notes behind Item 3, separate obligations from ceilings, deduplicate insider filings and list every patent assignee. To run the same pack across a pipeline of targets, the DeepResearch Batch API runs the tasks in parallel with one webhook when they finish.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows for deal teams each come with set inputs, a recommended mode and file deliverables:

To run one from code, pass its slug as workflow_id instead of a query:

task = valyu.deepresearch.create(
    workflow_id="pe-commercial-dd",
    workflow_params={"target": "BigBear.ai", "thesis_question": "Can commercial revenue offset slower Army program volume?"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

To browse the full library, open Workflows in the Valyu dashboard, or see running a workflow in the docs. For worked private equity examples, including a deal-sourcing queue and an acquisition screening brief, see the private equity workflows.

Where AI agents stop in due diligence

The checks above show what DeepResearch can do from public records. Four limits shape how to use it:

• The data room is still the data room. Contracts, cap tables, quality-of-earnings work and management accounts are private. The agent's job is to arrive at the data room with sharper questions, not to replace it.
• Public filings have gaps. The 10-K doesn't name Super Micro's largest customer or estimate its legal exposure, and no research can add what the company didn't disclose.
• Unverified means check it. DeepResearch marks what it couldn't confirm, such as the trading-plan status and two patent assignees. In our run, those were exactly the points worth a look at the source.
• Judgement stays with people. Lawyers assess legal exposure and the investment committee weighs the risks. Bloomberg Law's commentary on AI in due diligence makes the same point about human oversight.

How to set up a due diligence agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp and ask for the same research in plain English. To build your own agent, follow how to build a financial research agent with Valyu. For deeper work on one company, see our guides to writing a first-look brief on an acquisition target, capital structure and debt covenants and comparing 10-K risk factors year over year.

If you already run your own agents, they can call Valyu's search for single lookups. For a check you'd put in front of a committee, use DeepResearch.

Three habits made the checks above reliable:

• Put the known traps in the research strategy. Telling DeepResearch to read Note 15, separate obligations from ceilings and deduplicate Form 4s produced briefs that did exactly that.
• Ask for the basis of every figure. Denominator, period and source quote turn a number into something you can defend in committee.
• Follow up on anything marked unverified. It takes minutes, and in our run it changed the patent answer.

FAQ

What is AI for due diligence?

AI for due diligence is the use of language models and AI agents to gather and read the evidence behind a deal: SEC filings and their notes, litigation disclosures, insider filings, federal contracts, patents, clinical trials and news. Most of the value today is in fast, cited outside-in research before data room access, not in making the investment decision.

What are AI agents for due diligence?

AI agents for due diligence are language models that read primary records such as 10-K filings, Form 4s, USAspending awards, patent databases and trial registries, and return cited findings. Research agents such as Valyu DeepResearch plan the work, follow cross-references and write a brief, building the outside-in picture of a target and the questions for management.

Can AI do due diligence?

AI can do much of the research behind due diligence, but not the judgement. In a November 2026 test, six Valyu DeepResearch tasks in fast mode ran outside-in checks on public and private targets for 150 credits, about $0.94, and finished within 7 minutes, with every key finding matching EDGAR, USAspending, Google Patents or ClinicalTrials.gov. Legal assessment, data room review and the investment decision still need people.

What can AI agents check before data room access?

Before data room access, AI agents can check customer and supplier concentration in the 10-K, legal proceedings and investigations in its notes, federal contract awards and their periods of performance, insider buying and selling, patent ownership and co-ownership, and the status of clinical trials. Each finding comes with the filing, award, patent or trial record behind it.

What is the best AI tool for due diligence?

AI due diligence tools fall into three groups. Data room and document review platforms, such as Datasite and V7, analyse documents the target provides. Research platforms, such as AlphaSense and Hebbia, search licensed content inside their own products. Valyu DeepResearch researches primary records, such as SEC filings, federal awards, patents and clinical trials, and returns cited briefs, IC memos and commercial diligence packs through an API or inside Claude and ChatGPT.

What is the best API for building a due diligence agent?

The best API for a due diligence agent reads the primary records diligence depends on, follows cross-references and returns a source for every finding. Valyu DeepResearch covers SEC filings, insider transactions, USAspending, USPTO and EPO patents, ClinicalTrials.gov, PubMed and the web, and returns cited reports with xlsx, docx and pptx files, including ready-made workflows for commercial due diligence and IC memos.

How much does AI due diligence cost?

With Valyu DeepResearch, a fast-mode check costs a fixed 25 credits, about $0.16, and a standard report 100 credits, about $0.63. In a November 2026 test, six fast-mode due diligence checks cost 150 credits, about $0.94, in total. A heavy-mode report uses 500 credits, about $3.13. Free credits are included every month.

Can AI do due diligence on private companies and startups?

Yes, from the public record a private company leaves: patents, federal grants and contracts, clinical trials, papers and news. In a November 2026 test, Valyu DeepResearch found 10 granted US patents for Commonwealth Fusion Systems, and a check of the two it marked unverified showed that 7 of the 10 are co-owned with MIT. Private financials still come from the data room.

Can AI write an investment committee memo?

Yes, as a cited first draft. Valyu's Investment Committee Memo workflow (pe-investment-memo) writes the thesis, company, market, competition, financials, value creation, risks and exit sections, with a supporting xlsx. The deal team reviews the figures and the judgement before the memo goes to committee.

How accurate is AI due diligence?

AI due diligence is as accurate as the records it reads and the checks you run. In a November 2026 test of six DeepResearch tasks, every key finding matched its primary source. The points that needed care were the ones DeepResearch itself marked as unverified: a trading-plan status it couldn't read and two patent assignees, both of which turned out to be co-owned with MIT.

Start with one check

Pick the check that takes your team the most time on every deal, such as reading the litigation notes or mapping a target's federal contracts, and run it as a DeepResearch task on a company you already know. Compare the brief with your own work.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For the analyst side of the same research, see AI Agents for Financial Research.