Deal AI

Deal AI runs analytical work on the documents inside a deal. Combine three years of P&Ls into one comparable view. Extract every deposit from a stack of bank statements. Compare a tax return against the financials. Every output cites the source files it pulled from, so you can verify the numbers.

Overview

Deal AI shows up in two places inside a deal. They share the same understanding of the deal and hand off to each other:

• The sidebar — A persistent panel on the right side of every deal. You type a prompt; Deal AI does the work; the output streams in with citations.
• File summary cards — A sparkle on every classified file. Click for a structured popover with the most useful fields already extracted. No prompt required.

Both surfaces are grounded in the same knowledge base: the files, requests, Q&A, and chat inside this specific deal. Deal AI does not pull from the public internet, and it does not see other deals.

Why It Works on Vetting Vault

AI quality is bounded by the structure of the data it works on. Most data rooms hand AI a folder of files with names like Financials_FINAL_v3.pdf and Q3_bank_stmt_revised.pdf. That tells an AI almost nothing.

Vetting Vault is request-first: every file is uploaded against a specific request (“most recent three years of bank statements,” “current AR aging report,” “Q4 P&L”). Because the system knows what each file is supposed to be, Deal AI can:

• Find the right documents for a question without guessing at filenames.
• Apply the right extraction template (P&L vs balance sheet vs tax return).
• Cite back to specific files when it makes a claim.
• Cross-check figures across documents that share a known relationship.

This is the part folder-based VDRs with bolted-on AI structurally can’t replicate. Request-first intake isn’t a UX choice — it’s what makes the analysis layer function.

The Sidebar

The Deal AI sidebar is a persistent panel that lives on the right side of every deal. It stays available no matter where you are in the deal — viewing a request, reviewing a file, reading chat — so you can ask Deal AI to do work without losing your place.

Opening the Sidebar

There are two ways to open the sidebar:

• From the deal toolbar — Click the Deal AI / Sparkle icon at the top of the deal page.
• From a file summary card — Click Ask Deal AI about this file in any file’s summary popover. The sidebar opens with that file already loaded as context.

Asking Deal AI to Do Work

Deal AI is framed around doing analytical work on the documents, not answering trivia about the deal. Concrete examples of what works well:

• “Combine the last three years of P&Ls into one normalized summary table.”
• “Pull deposits, withdrawals, and average balance from every bank statement in the financials section.”
• “Reconcile the seller’s working capital schedule against the trial balance.”
• “Compile every candidate resume into an Excel ranked by years of relevant experience.”
• “Extract every change-of-control clause from the contracts folder and tell me which contracts have one.”
• “Compare revenue across the tax return, the P&L, and total bank deposits and flag any mismatch greater than 5%.”

Status-style questions also work as a secondary use case — “what’s overdue?”, “what’s still missing?”, “summarize this deal” — but the sidebar is most valuable when you ask it to do the kind of work you’d otherwise have to do file by file.

Citations and Confidence

Every output from the sidebar includes two trust signals:

• Source citation chips — Below each answer, Deal AI lists the specific files and requests it pulled from. Click any citation to open the source file at the relevant page.
• Confidence pill — A small badge on each answer indicating how confident Deal AI is in the output. Most answers land at “high” or “medium.” Low-confidence answers show an amber “AI — review” flag.

File-Context Handoff

When you open the sidebar from a file summary card (via Ask Deal AI about this file), the sidebar pre-loads that file as context. You’ll see a Context: <filename> chip and a set of seeded follow-up prompts tailored to the document type.

Example: opening the sidebar from a P&L summary card seeds prompts like “compare this to the prior year,” “reconcile against the bank deposits,” and “flag any line items that grew more than 25%.”

File Summary Cards

Every classified file in the deal carries a small sparkle badge on its icon. Click the sparkle and a popover opens with a structured summary of the file — the most useful fields for that document type, already extracted. No prompt required.

Summary cards are how Deal AI gives you value passively, without you having to remember to ask. The first time you open a file in a deal, the relevant fields are already there.

Supported Document Types

Each document type has its own template — a curated set of the fields that matter most for that kind of file:

• P&L statements — Period, Revenue, Gross Profit, Operating Income, Net Income
• Balance sheets — Period, Total Assets, Total Liabilities, Total Equity
• Bank statements — Institution, Account, Period, Beginning Balance, Ending Balance, Total Deposits, Total Withdrawals
• Tax returns — Tax Year, Entity, Gross Revenue, Taxable Income, Taxes Paid
• LOIs / IOIs / Term Sheets — Buyer, Target, Cash, Equity / Rollover, Earnout, Financing Condition, Exclusivity Period, Expiration

New document templates ship regularly. If a file doesn’t match a known template, the sparkle simply doesn’t appear — you can still ask the sidebar to extract whatever you need.

When Summaries Generate

Summaries are generated lazily — the first time someone opens the popover for a given file, Deal AI extracts the fields and caches the result. Subsequent opens are instant.

While a summary is generating, the popover shows a “Generating AI summary…” state with a pulsing sparkle. Generation typically takes 5–15 seconds. After generation, the summary is cached at the file level and shared with every user in the deal.

The “AI — review” Badge

When Deal AI’s confidence in an extracted field is below the trust threshold, the popover header shows an amber AI — review badge.

This typically happens when:

• The file is a scan with poor OCR quality.
• The document is non-standard for its type (a unique balance sheet format, for example).
• Several of the expected fields couldn’t be located.

Treat “AI — review” cards as a starting point, not a final answer. Open the source document and confirm before relying on the numbers.

What Deal AI Doesn’t Do

Knowing the boundaries helps you trust the boundaries:

• Doesn’t pull from the public internet. Deal AI is grounded only in your deal’s files, requests, Q&A, and chat. It doesn’t look up market comparables, public filings, or news unless that information is in the deal.
• Doesn’t share across deals. Each deal has its own scoped knowledge base. Deal AI in Deal A cannot see anything from Deal B, even if you’re an admin on both.
• Doesn’t make business decisions. The agent does extraction, combination, comparison, and reconciliation. Judgment calls — pricing, structure, go/no-go — remain with you.
• Doesn’t silently fix uncertain answers. When confidence is low, Deal AI tells you with the “AI — review” badge. It does not paper over what it doesn’t know.

MCP Server (External Access)

If you already use external AI tools — Claude.ai, ChatGPT, your own custom agent — you can give them direct access to a deal’s knowledge base through Vetting Vault’s MCP (Model Context Protocol) server.

The MCP server exposes the same structured knowledge base Deal AI runs on, with OAuth-based consent and per-deal scoping. Your external agent can read requests, files, and Q&A without an export step.

• OAuth consent flow — You explicitly authorize each external client.
• Per-deal scope — Access is granted at the deal level. An external agent only sees the deals you’ve approved.
• API tokens — Manage and revoke tokens from your account settings.

Setup details are in your account’s API Tokens section. The MCP server is most useful to sophisticated buyers who already have their own analytical tooling.

Best Practices

• Frame prompts as work, not Q&A. “Combine the last three P&Ls” gets a better artifact than “tell me about the financials.”
• Verify against citations before acting on numbers. Click through the citation chips when the answer is going to drive a decision.
• Treat “AI — review” as a yellow light. The information may still be useful, but confirm with the source document.
• Use the file-context handoff. If you’re working in one file’s summary card, click Ask Deal AI about this file instead of retyping the filename in the sidebar — the agent gets cleaner context that way.
• Upload to the right request. The more accurate the request taxonomy, the better Deal AI’s outputs. A bank statement uploaded to the “Bank Statements” request gets matched to the bank-statement template automatically.
• Use the MCP server when you already have an AI workflow. If your team runs analysis in Claude or a custom tool, hand it the MCP connection instead of re-uploading documents.