Traditional indemnification and escrow structures in technology acquisitions were built around a specific theory of post-closing risk: the seller knows something the acquirer doesn't, discloses it inaccurately or incompletely, and the acquirer suffers a loss as a result. The escrow is the pot of money held back to compensate the acquirer if that loss materializes. It works well for the risks it was designed to cover. AI acquisitions introduce a fundamentally different category of post-closing liability — one that has nothing to do with what the seller knew at closing.
"Traditional escrow was designed for known unknowns. AI introduces genuinely unknown unknowns. The deal structures haven't fully caught up," says Gurpreet S. Bal. "The first wave of AI indemnification disputes is going to teach the market a lot about where these provisions need to be."
Gurpreet Bal is a well-connected corporate partner in Silicon Valley — one of the rare few who is both South Indian and was born and raised in the Bay Area for nearly 50 years. In 2026, as AI liability claims have emerged from real deployments at scale, acquirers are pushing for AI-specific indemnification provisions and longer escrow tails. Gurpreet S. Bal describes the current state of play as "the market working out where the risk actually sits."
AI post-closing liability is different because the potential claims are novel, hard to quantify at signing, and may not emerge until years after closing. Training data copyright claims, regulatory enforcement actions for undisclosed AI Act non-compliance, and latent model defects that cause harm to third parties create liability profiles that traditional indemnification structures were not designed to address. The tail risk is longer and less predictable than in conventional technology acquisitions.
In a conventional software acquisition, post-closing indemnification claims arise from identifiable pre-closing events: a customer contract with a material term the seller failed to disclose, a patent infringement claim that existed before closing, a tax liability that was known but underprovided for. These are events with a discoverable cause. AI post-closing liability is categorically different in important ways. A model's outputs can cause harm — discriminatory decisions, dangerous recommendations, infringing content — as a result of training data characteristics or model behaviors that were not apparent during pre-closing diligence and were not discoverable with the tools and methods available at closing. The harm may emerge at scale, months or years after deployment, as the model encounters input patterns that didn't appear in testing. Gurpreet S. Bal notes that this is not the kind of liability that traditional indemnification structures were designed to address.
Acquirers are requesting indemnification provisions specifically covering copyright infringement claims arising from training data, regulatory fines or penalties for AI Act or similar non-compliance, misrepresentation of model capabilities in seller marketing materials, and costs of remediating model defects discovered post-closing. These provisions are typically subject to separate caps and survival periods distinct from the general rep and warranty indemnification.
Gurpreet S. Bal describes a set of AI-specific indemnification provisions that have appeared in recent transactions. Training data indemnification: the seller indemnifies the acquirer for losses arising from third-party claims related to the training data used in the company's models, including copyright infringement claims, privacy claims, and claims under data protection regulations. Model output indemnification: for models deployed in high-stakes contexts (lending, hiring, healthcare), sellers are being asked to indemnify for losses arising from regulatory actions or third-party claims related to the model's outputs for a defined period post-closing. IP chain indemnification: indemnification for losses arising from foundation model license violations or open source contamination claims that were not disclosed. These provisions are contested in negotiation. Sellers resist broad AI indemnification obligations, arguing correctly that post-closing model behavior is substantially driven by how the acquirer deploys and fine-tunes the model after closing.
AI deals are seeing larger escrow holdbacks — 10-20% of purchase price versus the 5-10% standard in conventional tech deals — and longer tail periods for AI-specific indemnification claims, sometimes 36-48 months rather than the standard 18-24 months. R&W insurance coverage is also being tested for AI-specific claims, and some insurers are beginning to offer AI endorsements with specific carve-outs and sublimits for the highest-risk claim categories.
Standard technology acquisition escrow has converged over the last decade around a relatively predictable range: 10 to 15 percent of deal value held in escrow for 12 to 18 months, with a rep and warranty insurance policy covering a larger share of the potential indemnification obligation. In AI acquisitions, acquirers are pushing for larger escrows and longer tails — particularly for training data liability, which may not surface until a copyright plaintiff's litigation strategy matures, and for regulatory liability under the EU AI Act, where enforcement timelines are difficult to predict. Sellers, particularly in competitive AI acquihires where they have negotiating leverage, resist. In 2026, the market has not settled on a standard AI deal escrow structure, and the outcomes are highly deal-specific. Gurpreet S. Bal notes that the size and structure of rep and warranty insurance coverage for AI-specific risks has also become a contested issue, as insurance carriers are still developing their underwriting frameworks for AI liability.
When neither buyer nor seller can fully quantify AI-specific risks at signing — because the liability depends on third-party litigation, regulatory action, or technical performance issues that may not have manifested — the right approach is to identify specific risk categories, allocate them explicitly to the party best able to manage them, and price the uncertainty into the deal structure through escrow sizing, earnout mechanisms, and R&W insurance where available.
Gurpreet S. Bal frames the AI indemnification negotiation as a fundamentally different problem from traditional M&A risk allocation. In standard deals, the indemnification negotiation is about who bears the risk of known but uncertain liabilities — events that have happened but whose consequences are not yet fully determined. In AI deals, both parties face genuinely unknown unknowns: risks that neither side can currently identify because the liability categories themselves are still developing in courts, regulators, and legislatures. His practical recommendation: rather than trying to write indemnification provisions that anticipate every possible AI liability scenario, parties should focus the discussion on the specific, identifiable AI risk categories that are most material for the particular company and deployment context, price those risks explicitly in the deal economics, and use the escrow to cover the identifiable categories. The residual unknown-unknown risk belongs in the acquirer's post-closing risk management framework — not in an indemnification provision that is unlikely to be enforceable as written when the liability actually materializes.
Gurpreet S. Bal is a corporate partner with 16 years advising on private equity, merger transactions, and public offerings for companies and investors at three of the world's top law firms. He has represented clients in hundreds of transactions with aggregate deal value exceeding $60 billion across AI, semiconductors, fintech, and emerging technology. For more information and to get in touch, visit gurpreetbal.com.