A data room containing 20,000 documents takes a diligence team of five professionals three to four weeks to review manually. Each professional reads approximately 1,000 documents, extracting terms, identifying risks and cross-referencing across their assigned section.
The same 20,000 documents, processed through an AI-powered document review platform, can be classified, key-term-extracted and risk-flagged in 48 to 72 hours. The AI does not replace the three to four weeks. It restructures them: instead of starting with document one and reading sequentially, the team starts with the AI’s exception list, the flagged contracts, the identified risks, the cross-document inconsistencies.
The weeks are still required. They are spent on the documents that matter rather than on the documents that do not. The difference is not time saved. It is findings gained: the AI identifies connections and inconsistencies across the full population that sequential human review, constrained by working memory and section assignments, would almost certainly miss.
What AI Should Screen For
Change-of-control provisions
Across every customer contract, supplier agreement, lease, licence and financing document, identify clauses triggered by a change in ownership or control. These provisions can allow counterparties to terminate, renegotiate or accelerate obligations upon acquisition.
A 500-contract portfolio screened by AI identifies the 35 contracts containing change-of-control provisions within hours. Those 35 contracts represent Rs 180 crore of annual revenue. The buyer now knows, before detailed review begins, that Rs 180 crore of revenue is contractually at risk upon acquisition.
Without AI screening, finding all 35 requires reading all 500 in full. Under time pressure, the team might review 350 and miss 15 of the 35, leaving Rs 50 crore of at-risk revenue unidentified.
Termination and cancellation rights
Every contract with a termination-without-cause provision on less than 90 days’ notice. Every contract where the customer can cancel without penalty. Every agreement where the supplier can discontinue service on short notice.
The AI produces a matrix: contract, counterparty, annual value, termination provision, notice period. The matrix gives the buyer a single-page view of revenue and supply-chain fragility.
Guarantees, indemnities and contingent obligations
Every guarantee the target has provided to subsidiaries, lenders, customers and landlords. Every indemnity given. Every obligation contingent on a future event. These may not appear as liabilities on the balance sheet. They can crystallise post-close.
The AI extracts the guarantee amount, the beneficiary and the triggering condition. A target with Rs 40 crore of outstanding guarantees that is not separately disclosed in the financial statements has a contingent exposure the buyer must evaluate.
Related-party references
Every document that references a related party, an affiliate, a group entity or a director’s interest. Cross-reference with the disclosed related-party list. The gap between referenced and disclosed identifies connections that are documented but not formally reported.
IP ownership and licensing
Ownership, licensing and assignment provisions for every piece of intellectual property. Is the IP owned by the entity being acquired, by a group entity, by an individual or by a third party? A target whose core technology is owned by the promoter personally, not by the company, has an IP risk that the AI can flag before the human team spends days reviewing product documentation.
Litigation and regulatory references
Every reference to pending, threatened or potential litigation, regulatory proceedings, investigations and compliance obligations across all documents, not just the legal disclosures. A settlement discussion in a board minute that is not reflected in the litigation provision is a finding AI can surface.
Cross-document inconsistencies
Revenue in the financial statements that differs from a revenue figure in a customer contract. Headcount in the HR records that differs from the management presentation. A facility description in a lease that differs from the fixed-asset register. Pricing in a vendor contract that differs from the procurement data.
Each inconsistency is a question. Some have innocent answers (definitions differ, dates differ, rounding). Some do not. The AI identifies all of them. The human team determines which matter.
What AI Cannot Do
Assess materiality. AI flags a change-of-control clause. It cannot determine whether the customer will exercise it, what the probability is, what the revenue impact would be, or whether the risk justifies a price adjustment. That assessment requires commercial judgment, relationship intelligence and negotiation experience.
Evaluate management credibility. AI identifies inconsistencies between management representations and underlying documents. It cannot determine whether the inconsistency is an honest error, a different interpretation or a deliberate misrepresentation.
Handle false positives without human review. AI generates false positives: standard boilerplate flagged as a change-of-control provision, entities flagged as related parties because they share a common name with no actual connection. Every flag requires human evaluation.
Guarantee confidentiality without controls. Processing transaction documents through cloud-based AI platforms creates data-security risks. Confidentiality agreements, data-processing protocols and, where necessary, on-premise deployment are prerequisites.
In Northrop Management Private Limited’s due diligence practice, AI is deployed as a first-pass intelligence and exception-identification layer. The AI processes the full document population. The forensic team investigates each exception using professional judgment, primary documentation and management inquiry.
Ashish Chaudhary, frames the technology positioning directly: “AI does not conduct due diligence. It identifies what due diligence should focus on. A 20,000-document data room screened by AI before human review begins produces a team that starts with the 200 documents most likely to contain findings, rather than one that starts with document one and hopes to reach document 20,000 before the deadline. The AI does not replace the judgment. It ensures the judgment is applied where it matters.”
Questions for the Boardroom
- Has the data room been screened by AI for change-of-control, termination, guarantee, indemnity and related-party provisions before detailed human review began?
- How many contracts contain change-of-control provisions, and what is the total annual revenue contractually at risk upon acquisition?
- What is the total value of guarantees, indemnities and contingent obligations identified across all documents, and how does this compare to what is disclosed in the financial statements?
- Were any cross-document inconsistencies identified (headcount, revenue, pricing, contract terms stated differently in different documents), and have all been investigated?
- For every AI-generated flag, has a qualified professional assessed the finding and determined its commercial significance and transaction impact?
Practical Checklist
Before data room access: Configure AI with extraction rules for change-of-control, termination, guarantee, indemnity, IP, litigation and related-party clauses.
Within 72 hours: Complete AI screening. Generate exception list ranked by risk category and estimated value.
Human review priority: Every flagged change-of-control provision. Every guarantee above Rs 1 crore. Every cross-document inconsistency. Every related-party reference not in the disclosed list.
Throughout engagement: Re-run screening as new documents are uploaded. Track management responses to AI-identified questions.
Before signing: Confirm every material flag has been resolved: investigated, explained and quantified.
Closing Implication
AI in due diligence is the mechanism that ensures professional judgment is applied to the right documents. The board that insists on AI-screened diligence gets a team that starts with the highest-probability findings. The board that accepts manual-only review accepts the risk that the team ran out of time before reaching the document that contained the most consequential finding.