Business Context: The Due Diligence Cost Problem
Due diligence fees in mid-market M&A transactions often represent 1-3% of deal value. A significant portion of these fees pays for junior lawyers and analysts to read documents and extract information that could, with the right tools, be extracted by AI in a fraction of the time. The senior adviser's time, which is what actually commands the high fee rates, is spent on interpretation, negotiation strategy, and client advice. AI due diligence tools redirect the time ratio: less on extraction, more on interpretation and advice. The market for this capability is broadly split between corporate law firms that want to deliver faster, higher-margin due diligence services to clients, and buy-side advisory firms that want to increase their transaction throughput without increasing headcount. Both benefit from the same underlying capability: fast, accurate extraction of key terms and risk flags from a large document corpus, presented in a structured format that supports the adviser's analysis. For UK law firms, the Solicitors Regulation Authority has published guidance on AI-assisted legal work that reinforces the principle of lawyer responsibility for all outputs. The AI system is positioned as a research assistant, not a substitute for qualified legal review.
Architecture: Virtual Data Room, RAG, and Structured Reporting
The system is built around three components. The virtual data room provides a secure, access-controlled repository for due diligence documents. Clerk handles authentication and role-based access: deal principals, legal advisers, financial advisers, and management team members have different document access permissions configured per deal. Documents are uploaded, encrypted at rest in AWS S3, and processed through an ingestion pipeline that extracts text, generates embeddings via Claude or OpenAI Embeddings, and indexes them in Pinecone namespaced by deal. The document Q&A layer allows advisers to ask natural language questions against the document corpus and receive Claude-generated answers with precise document and page citations. The checklist reporting layer maps a customisable due diligence checklist (legal, commercial, IP, employment, property, regulatory) to the document corpus, generates a structured assessment for each checklist item, and produces a formatted due diligence report exportable to Word or PDF. All document access, queries, and generated outputs are logged for the deal audit trail.
AI Components: Document Q&A and Checklist Extraction with Claude
Claude is the AI backbone for two tasks. Document Q&A: given a natural language question from an adviser (for example, 'what are the change of control provisions in the target's key customer contracts?'), retrieve the relevant contract sections from Pinecone and generate a synthesised answer with precise citations to document name, section, and page. This task is where Claude's long context window and instruction-following reliability are most valuable: advisers need accurate, cited answers, not approximate summaries. Checklist extraction: given a due diligence checklist item (for example, 'identify all IP assignments and confirm the target has clear title to its core IP'), retrieve relevant documents and generate a structured response covering what was found, any gaps in the documentation, and any risk flags. The extraction is governed by a structured output schema so that checklist responses are consistent across deals and can be compiled automatically into the due diligence report template. The system prompt for legal document analysis is carefully designed to avoid overclaiming certainty. When documentation is absent or incomplete, Claude is instructed to flag the gap explicitly rather than infer an answer.
Challenges: Document Sensitivity, Confidentiality, and Accuracy Standards
Due diligence documents are among the most commercially sensitive materials in business. Robust security is not optional: it is the table stakes for any product used in M&A transactions. SpeedMVPs builds the security architecture to enterprise standards: encryption at rest and in transit, strict role-based access controls, document access logging, and session timeout policies. Claude API calls process document text within the inference request. Anthropic does not train on API data submitted via enterprise agreements, which is a standard disclosure advisers require before approving AI tools for transaction use. Accuracy standards in legal due diligence are exceptionally high. A missed change of control clause or an overlooked IP encumbrance can have material financial consequences for the buyer. The system is designed to be comprehensive in flagging potential issues rather than conservative. Missing a risk flag (false negative) is a more serious failure mode than surfacing a flag that turns out not to be material (false positive). This is the opposite calibration from a consumer product and it shapes every design decision. Advisers are expected to review all flagged items, not rely on the AI's judgment.
Outcomes: Deal Speed and Adviser Capacity
Law firms and advisory teams that deploy AI due diligence tools report three primary outcomes. Deal cycle time for the due diligence phase reduces. The initial review and checklist extraction that previously took a team of juniors 2-3 weeks to produce can be completed by the AI in 24-48 hours, giving the senior team more time for analysis and negotiation preparation. Document Q&A reduces adviser hours on research queries. Senior advisers can ask questions directly against the document corpus in real time during management meetings or negotiation calls rather than having to pause for junior team members to find and read relevant documents. Deal capacity increases. With AI handling the extraction workload, the same team can run more transactions simultaneously. For advisory practices competing on deal volume, this is a direct capacity expansion without additional hiring.
Lessons: Lawyer Review Is Non-Negotiable, Citability Is the Standard
Due diligence AI tools that try to eliminate lawyer review fail in the market because clients will not accept unreviewed AI outputs in a transaction context. The right product design positions the AI as accelerating and improving lawyer review, not replacing it. Every AI-generated checklist item and every Q&A response must link to the specific document and section it is based on. This is not just good practice, it is the minimum standard for adoption in a legal context. If an adviser cannot verify a claim against the source document in 10 seconds, they will not trust the output. The second lesson is to work with the firm's precedent checklist. Every law firm has its own due diligence checklist, developed over years of transaction experience. An AI tool that imposes a generic checklist will not be adopted. SpeedMVPs builds the checklist extraction system to accept and work from the firm's own checklist format, configured during the build.