Why Your Elite M&A Law Firm Is Lost on AI

It is 9:14 p.m. on a Thursday. The fourth-year associate at one of the most prestigious corporate law firms in the world has the seller’s data room open in one tab and the firm’s merger agreement template, last meaningfully revised in 2019, in the other. The IP rep she is drafting uses the seven-category definition of “Intellectual Property” the firm has been using since 2014. It assigns the seller’s “Software” to source-code repositories and SaaS deployment manifests. It assumes the seller’s value lives where her template says it lives.

She has never written a system prompt, never assembled a library, never seen a fine-tune, never wired up an agent. Her partner has not either. Her firm bought Legora last year for $800K; the Slack channel celebrating use cases has 20 members, the firm has 1,400 lawyers. [S: it’s the same story at every large firm.]

A material portion of the target’s value is locked in a new AI asset class. None of the lawyers on either side, with thousands of deal closings between them, has ever encountered it. The deal does not account for it. The merger agreement does not represent it. The escrow does not protect against its loss. Both parties will live with material ramifications they did not negotiate.

This is the story of elite big law firms closing M&A tech deals in 2026.

Forty years of asset evolution, encoded in your purchase agreement

Walk forward through the asset classes M&A actually priced over the last forty years. Tangible assets in the 1980s, factories and inventory and real estate, tracked by fixed-asset schedules and title reps. Registered IP in the 1990s, patents and trademarks and copyrights, tracked by the IP rep that grew teeth and the watch services that became standard. Source code and customer contracts in the late 2000s, when SaaS multiples replaced revenue multiples, open-source compliance became a rep, and the top-50 customer schedule became its own exhibit. Data and platforms in the 2010s and early 2020s, when GDPR and CCPA forced data-protection reps and breach reps into the standard form, and SOC 2 attestations became diligence items.¹

Every era added a schedule. Every era added a rep. The standard 2024 purchase agreement is the layered accumulation of forty years of asset evolution, each layer added when the asset emerged, none removed. The schedules read like a deposition timeline.

Schedule 3.1: fixed assets. Schedule 3.2: real property. Schedule 3.4: inventory. The 1980s.

Schedule 3.7: registered IP. Schedule 3.7(a): patents. Schedule 3.7(b): trademarks. Schedule 3.7(c): copyrights. Schedule 3.7(d): domain names. The 1990s and early 2000s.

Schedule 3.8: software. Schedule 3.8(a): proprietary software. Schedule 3.8(b): open-source components. Schedule 3.9: material customer contracts. The 2010s.

Schedule 3.12: data privacy. Schedule 3.13: security incidents. The early 2020s.

There is no Schedule 3.X: AI application layer. No Schedule 3.X(a): system prompts and custom instructions. No Schedule 3.X(b): library and provenance. No Schedule 3.X(c): fine-tuned weights and embeddings. No Schedule 3.X(d): agent credentials and tool integrations.

The merger agreement does not have a slot for the asset class that is increasingly driving a tech company’s valuation.

What the AI application layer absorbed

This is the part the merger agreement misses, and the part that matters most.

For forty years, M&A documents tracked the asset by tracking its proxies. The asset itself, the company’s accumulated capacity to do what it does, was never directly inventoried. It was inferred from the things that surrounded it.

The IP schedule was a proxy for the judgment that had been formalized into protectable form. The patents stood in for the inventions. The trademarks stood in for the brand. The registered copyrights stood in for the authored works.

The key-person schedule and the employment agreements were a proxy for the people who carried the unformalized judgment in their heads. The senior engineer. The founding salesperson. The principal architect. The non-compete, the retention bonus, the earn-out vest were the buyer’s hedge against the proxy walking out the door.

The source code repository was a proxy for the executable encoding of every business decision the team had ever made. The customer contract schedule was a proxy for the relationships, the negotiated terms, the trust accumulated over years of operation. The trade secret rep was a proxy for the ad hoc institutional know-how that lived in internal systems.

The AI application layer collapses all of it.

This is not an additive change. The AI application layer is not a new asset class added to the existing list. It is the destination toward which every prior asset class has been quietly migrating for two years. The IP that mattered is in there. The people who mattered are in there. The knowledge that mattered is in there. The customer history that mattered is in there. The workflow that mattered is in there. Strip the layer out and the company is a brand, a customer list, and a cash flow that will struggle to survive 18 months of competitive pressure.

The AI application layer is increasingly the company.

The merger agreement and your law firm don’t know that yet. [M: the merger agreement is a museum.]

Why the lawyers cannot see it

Three reasons, in increasing order of awkwardness.

Templates are conservative. M&A practice iterates by reference to last quarter’s deal. The merger agreement gets marginal updates. New schedules get added when an asset class becomes undeniable: when the CLE programs cover it, when the rep insurance carriers underwrite it, when the firm’s training program names it. The AI application layer has not yet crossed those thresholds. The merger agreement is two years behind, and two years is the entire life of the asset class.²

The asset is invisible to inspection. Tangible assets sat on the balance sheet. Registered IP sat in a government database. Source code sat in a repository. Customer contracts sat in DocuSign. The AI application layer sits on someone else’s platform, under credentials that may or may not transfer, in a configuration that may or may not be exportable. There is nothing to inspect physically. There is nothing to search. There is no inventory unless someone built one. The diligence checklist was built around inspection. The asset evades inspection.

The lawyers have not lived inside it. This is the part the profession is not yet talking about. None of the lawyers on either side of the deal, associate, partner, seller’s counsel, has assembled a library, written a system prompt, wired up an agent, or seen a fine-tune. They are valuing an asset class they have not, in their own practice, encountered. They cannot describe to the client what the seller built because they have never built one.

We believe that if the best M&A lawyers in the world have not used the AI application layer, the work they produce cannot represent it.

It is not just the IP rep

Walk the merger agreement top to bottom and the gap repeats.

The asset-class representation. Schedule 3.X does not list the AI application layer. The seller represents ownership of patents, trademarks, copyrights, source code, and trade secrets. The seller does not represent ownership of the asset that drives the premium.

The retention mechanism. Key-person retention is tied to the seat, not the AI application layer. Earn-out milestones are calibrated to revenue, not to layer transfer. Non-compete carve-outs were drafted to protect customer relationships, not configurations on third-party platforms.

The post-closing claims framework. Survival periods, baskets, caps, and indemnity scope are tuned to the assets the merger agreement represents. The escrow releases on a calendar that does not contemplate the layer’s drift, the embeddings’ degradation, or the platform’s terms changing under the buyer’s feet.

The material adverse change definition. The buyer’s walk right between signing and closing does not reach a change in the layer. The seller can lose the admin, swap the library, or migrate the fine-tune in week eight, and the buyer has no out.

The closing conditions. The list of things that have to be true at the bring-down (accuracy of reps, no MAC, third-party consents) does not include the AI application layer surviving in working condition into the buyer’s hands.

Five different parts of the merger agreement. Same gap. It does not see the asset. [S: blind in five places.]

And that is in a stock or merger deal, where everything the company owns transfers by operation of law. In an asset purchase agreement, the gap widens to a chasm. The buyer in an APA acquires only the assets specifically named on the transfer schedule. Assets not named stay with the seller. The standard asset-transfer schedule does not name system prompts, custom instructions, fine-tuned weights, embeddings, retrieval indexes, or platform tenancies as Acquired Assets. Under the literal text of the form, the seller keeps the layer at closing. The buyer pays for a company that just lost its asset class.

The source-code asymmetry

The strongest counter to this piece is that the merger agreement is more elastic than we have portrayed it. The IP rep sweeps “all Intellectual Property necessary to conduct the business as currently conducted.” The Software definition reaches “scripts, configurations, data, databases.” The trade-secret rep covers “confidential information.” In a stock deal everything transfers by operation of law. The further-assurances covenant patches misses. The catch-all indemnity backstops everything else. The buyer is fine.

The merger agreement is elastic. Elasticity is not enforceability.

Compare what the merger agreement does with source code, an asset class the elite M&A firms have agreed for forty years is too valuable to leave to a sweep clause.

Source code gets a definition in Article 1. A representation of exclusive ownership. A schedule listing every material proprietary application. A separate open-source compliance representation, drafted to head off copyleft and disclosure obligations clause by clause. A separate trade-secret representation that names source code as the protected asset. An employee-assignment schedule confirming every developer’s contribution was assigned to the company. A repository-transfer covenant with admin handover at closing. A third-party-licensee schedule. A mature diligence playbook with enterprise vendors running open-source license scans, dependency audits, and vulnerability reviews on the seller’s repositories. In an asset deal, source code is named on the asset-transfer schedule and assigned at closing alongside the build pipelines and credentials.

Eight distinct touchpoints across the merger agreement, plus a parallel diligence shakedown. One asset class. The elite firms have decided source code is too important to leave to operation of law and a wish.

For the AI application layer, the asset that absorbs IP, people, knowledge, customer relationships, and workflow into one queryable artifact, the merger agreement has zero analogous touchpoints. No definition. No representation. No schedule. No open-source-style compliance subsection. No trade-secret call-out. No employee assignment confirming the system prompt is the company’s. No repository-transfer covenant for the project file. No platform-tenancy schedule. No admin-handover requirement at closing. No vendor scanning the seller’s Claude or OpenAI tenants for orphaned projects. No tool auditing the library for license provenance or training-data origin. No diligence checklist line item asking the seller to list every fine-tuned model trained on the company’s data, every embeddings index queried by production traffic, every system prompt deployed against a customer-facing surface. The diligence playbook for the AI application layer is empty. Nobody is scanning for it because nobody on the deal team is asking.³

The asymmetry is not that the AI application layer is too new to draft for. The asymmetry is that source code, an asset class with established economic value, gets every safeguard the elite firms have invented since 1990, while the AI application layer, where most of the company’s accumulated value now lives, gets a sweep clause and a wish. [S: lawyers run the last war.]

A weaker version of the counter says the AI application layer is just a configuration on a third-party platform, like a Salesforce instance or a Looker dashboard, handled informally. That was true two years ago. Not now. The AI application layer in 2026 is the senior solutions engineer’s eleven years on a thumb drive, the litigator’s six hundred and forty briefs in a project file, the VP of Sales’s three thousand calls in a configuration she built on her personal account. Not a vendor tool. The asset class M&A has spent forty years building proxies for, finally captured directly in one AI application layer that travels at the speed of an API call.

One more defense, the one every senior partner reading this is thinking. That’s not market. Show me where this is in the SRS Acquiom Deal Terms Study. The reflexive reach for a survey of what every other elite firm is already drafting, deployed against the suggestion that the elite firms might be missing something. The SRS Acquiom study covers what the market is already drafting. It does not cover what the market is failing to draft. Not in the study is not a defense. It is an admission that the entire elite tier is ignoring the same material asset class.

And here is the sharper irony. AI is collapsing the value of source code itself. Code generation has compressed development cycles to a fraction of what they were two years ago. Reproducing a competitor’s proprietary codebase used to take a team of engineers 18 months. It now takes a project, a corpus, and a weekend. The “proprietary code” defense was the centerpiece of B2B SaaS valuation for two decades. The dramatic compression in SaaS multiples since 2023 is the market repricing source code in real time. Code is becoming the commodity. The AI application layer is becoming the moat.⁴

M&A attorneys are investing more drafting effort in protecting an asset class whose value is dissolving, and ignoring the one that has absorbed it. The merger agreement is a lagging indicator. We strongly believe the market has already moved. Law firms need to catch up.

What this means in deal pricing

Two failure modes, both real.

The buyer pays an AI premium for a capability the IP rep does not represent. The AI application layer at the seller is on personal Claude accounts. The founder leaves at month three of the earn-out. The AI application layer leaves with her. The buyer’s only contractual recourse is through reps that do not name the asset, indemnities that do not cover it, and an escrow that releases on a calendar the buyer’s lawyer set against assets the rep was about. The price the buyer paid was for an asset the merger agreement will struggle to enforce against the seller losing.

The seller does not get credit for the layer being dialed in. Version-controlled prompts, a tagged library, an enterprise-tenant fine-tune with transferability, service-account credentials owned by the company. None of this raises the multiple, because the merger agreement does not have a place to give credit. The seller does not get paid more for being more buyable. The same merger agreement that lets the unprepared seller skate also flattens the prepared seller. [S: that preparation should get you much more.]

Both failure modes are produced by the same gap. The merger agreement encodes what M&A used to value. The deal is paying for what M&A values now. The mismatch is not closing on its own, and the lawyers tasked with closing it have not used the tool that would let them see it.

What to do

If you are buying, find a lawyer who actually understands the AI application layer and put them on the deal team. Your elite firm still closes the deal. The specialist makes sure the deal is for the right asset.

If you are selling, find counsel who can help you prep before you walk into the data room. The merger agreement does not credit the prepared seller. The right attorney does.

A closing thought

The associate is still at her desk at 11:42 p.m. The IP rep is done. The disclosure schedule is on its third pass. The merger agreement is the one her firm has been using since 2019. She has not opened Legora tonight. She has not opened it in eight months.

The seller’s AI application layer, invisible to the merger agreement, invisible to her, invisible to the partner who will sign her work, is actually what her client is paying for.

The rep her client is getting is for the assets of 1995, 2005, and 2015. Accreted, layered, refined, and silent on the asset class that has absorbed all of them.

In 2026 the AI application layer is increasingly the company. The merger agreement and your law firm don’t know that yet. Your counterparty’s counsel doesn’t either.

But now you do.

Postscript: what the recent agreements actually say

We pulled four recent material public-company merger agreements off EDGAR. We read them in chronological order. Each was drafted by one of the most elite M&A firms in the country. These are the deals every other firm in the country grades itself against.

HPE / Juniper Networks. January 9, 2024. Fourteen billion dollars. Section 4.16, the IP rep, covers patents, trademarks, copyright registrations, domain names, trade secrets (including source code), open-source compliance, and employee invention assignment. The Disclosure Schedule lists every item of Registered Company IP with title, application number, filing date, jurisdiction, and owner. Source code gets its own subsection. The rest of the merger agreement (the asset rep, the retention mechanism, the indemnity scope, the MAC, the closing conditions) is the standard 2018-vintage language. The full text of the agreement, scanned end to end, contains zero mentions of “artificial intelligence,” “machine learning,” “fine-tuning,” “system prompt,” “training data,” “model weights,” “large language model,” or “generative AI.” A fourteen-billion-dollar acquisition of a networking and security company in 2024 closed under a merger agreement that was unaware that AI exists.

Salesforce / Informatica. May 26, 2025. Eight billion dollars. Section 3.14, the IP rep, lists five categories of Intellectual Property, all tied to a government filing: patents, registered trademarks, domain names, copyright registrations, and “any other Intellectual Property that is the subject of an application, certificate, filing, registration or other document issued by, filed with, or recorded by” a government authority. The AI application layer fits into none of them. The agreement does mention AI in Section 3.9, the privacy and compliance rep, where the Company represents that it complies with laws governing the development, training, fine-tuning, and validation of “AI Systems,” and that it had legally enforceable rights to use the Training Data. It is a compliance rep, not an asset rep. It says the Company did not break any AI laws. It does not represent the existence, scope, ownership, or transferability of any system prompt, custom instruction file, agent configuration, fine-tuned weight, embeddings index, retrieval library, or tool integration. It puts none of them on a schedule. The retention package, the indemnity, the MAC, and the closing conditions all run on the standard merger agreement language, untouched by AI. The asset class is invisible across the document.

Electronic Arts going-private. September 28, 2025. Fifty-five billion dollars. Acquired by a consortium of the Saudi Public Investment Fund, Silver Lake, and Affinity Partners at $210 per share. Section 5.1(q), the IP rep, works through six subsections of the same forty-year-old asset framework: registered IP, infringement, trade-secret protection, employee assignment of inventions, government funding, source code, open source. Subsection (vii) is new. It reads, in full, that the Company and its Subsidiaries “(A) have obtained all licenses and consents, provided all notices and disclosures, and taken all other steps required by applicable Laws in order to collect and use all AI Inputs in the conduct of their business as currently conducted, (B) use all AI Technologies in compliance with all applicable license terms, consents, agreements and applicable Laws, and (C) have not been subject to any Actions regarding the Company’s or its Subsidiaries’ development or deployment of any AI Technologies.” Three subclauses, all compliance subclauses. Defined terms “AI Inputs” and “AI Technologies” carry the entire AI rep, and they carry it without representing ownership, scope, transferability, or location of any of the underlying components of the AI application layer. EA, a fifty-five-billion-dollar take-private of a software company that uses AI throughout product development, has its AI application layer represented by three compliance subclauses tucked into the seventh subsection of the IP rep. Nothing about the AI application layer is scheduled. Nothing in the retention package, the indemnity, the MAC, or the closing conditions reaches it. The asset class is invisible across the document.

Adobe / Semrush. Announced November 18, 2025. Closed April 28, 2026. $1.9 billion. Five months from signing to close, the most-watched mid-cap take-private of the cycle, and the sharpest test yet of the gap between deal narrative and deal document.

Adobe’s announcement does not bury what Adobe says it is buying. It leads with it. “Generative Engine Optimization (GEO).” “Agentic Search Optimization (ASO).” “How their brands appear across owned channels, large language models (LLMs), traditional search and the wider web.” Anil Chakravarthy, Adobe’s CXO Business president, on signing: “the rules of brand discovery and commerce are being rewritten in real time, and marketers who aren’t optimizing for that world today will find themselves invisible tomorrow.” Five months later, on close, Adobe doubled down. Semrush is the “AI-driven insights” platform and the “leading brand visibility platform” in the agentic era. Adobe’s two press releases, signing and close, use some version of “AI,” “LLM,” “agentic,” “generative,” “GEO,” or “ASO” on nearly every paragraph. Adobe paid $1.9 billion, in Adobe’s own words, for an AI application layer.

The merger agreement, filed as Exhibit 2.1 to Semrush’s November 19 8-K, was the operative document through close. Section 3.14, the Intellectual Property representation, runs from subsection (a) through (m). Subsections (a) through (l) walk the same forty-year-old framework. Registered Intellectual Property listed on the Company Disclosure Letter (patents, trademarks, copyrights, domain names). Ownership and exclusivity. Sufficient rights to conduct the business. Third-party software licensing. Non-infringement. Employee and contractor invention assignment. Trade-secret protection. Source-code release triggers. Open-source compliance. The same architecture HPE / Juniper used in 2024. The same architecture Salesforce / Informatica used six months earlier. Subsection (m), the AI rep, has four subclauses, and they will be familiar. (i) Semrush maintained policies and procedures, in compliance with applicable Law, governing the “development, training, fine-tuning, validation, testing, improvement, use and deployment of AI Systems.” (ii) For each AI System trained or fine-tuned, Semrush had “sufficient rights to collect, use and process” the data used. (iii) No Governmental Entity has asked Semrush about an AI System since January 2023. (iv) Semrush does not develop or deploy any “High-Risk AI Systems” or any AI activity “unacceptable” or “prohibited” under the EU Artificial Intelligence Act. Four subclauses. All compliance. The AI rep represents that Semrush did not break the law training a model. It does not represent that Semrush owns or can transfer the system prompts, the libraries, the fine-tunes, the embeddings indexes, or the agent configurations that make Semrush the “AI-driven” platform on the cover of the press release. Nothing in the IP schedule lists any of them. The merger agreement underneath, unchanged.

Either Adobe and its counsel were buying an AI application layer and failed to capture it in the deal document, or Adobe was not actually buying an AI application layer. Both are failures. Big Law underdrafted, or Big Tech oversold. We read this one closer to the second. Semrush is a strong SEO and GEO platform, but the deal cover oversells the institutional-judgment-encoded asset class we have been describing. The broader point holds either way. When a buyer pays $1.9 billion on an AI thesis, the merger agreement should be able to tell you which AI assets actually transferred. This one cannot.

Four deals, spanning twenty-two months. Each more recent. Each one slightly more aware that AI exists. None of them representing the AI application layer as an asset. None of them protecting the buyer who paid for it. None of them giving the seller credit for having built it.

In every one of these merger agreements, source code, the moat of 2010, got eight schedules, six representations, and a parallel diligence shakedown that likely ran into the millions. The AI application layer, the moat of 2026, got nothing.

We are not saying these were bad firms drafting on a bad day. We are saying elite M&A firms, drafting at the top of their game on the most-watched deals of the last two years, did not see the AI application layer. The merger agreement they produced does not see the AI application layer.

Sources: HPE / Juniper Merger Agreement, Ex. 2.1; Salesforce / Informatica Merger Agreement, Ex. 2.1; Electronic Arts Merger Agreement, Ex. 2.1; Adobe / Semrush Merger Agreement, Ex. 2.1; Adobe announcement (Nov. 19, 2025); Adobe announcement on close (Apr. 28, 2026).