AI Act and Delaware: What Actually Binds an AI Licence

The Delaware decision of 4 September 2026

On 4 September 2026, the Delaware Court of Chancery filed the post-trial memorandum opinion in Candor Compass LLC v. iDentivisuals SRL and iDentivisuals Inc., C.A. No. 2024-1179-MTZ^[1], signed by Vice Chancellor Zurn.

The case, submitted for decision on 1 May 2026, concerns a binding term sheet for the licence of an AI emotion-detection technology. The startup developing it was looking for a partner with experience and contacts in the legal tech market. The partner, acting through its own company, signed after two months of negotiations.

The jurisdiction is Delaware, the corporate home of most US technology companies, and the rule applied is contract law. While the European Commission applies the AI Act to systems placed on the Union market, here an equity judge decides on the basis of what the parties wrote. The two planes coexist, and the case serves to tell them apart with precision.

The starting point is simple: a technology far from market, a nearly empty cash box, and a contract that was supposed to cover both risks.

What the binding term sheet provided

The licensee negotiated contractual insurance against the startup's failure. The technology was still under development, and the partner wanted something usable should development stop.

The term sheet imposed three obligations in sequence. The startup had to make the technology under construction available, so that the licensee had material to work with. It then had to deliver everything needed to use the technology independently and place the remaining intellectual property in escrow.

The third clause provided that, in the event of default, ownership of the licensed technology would pass to the licensee.

The binding nature of the document was the startup's own choice: it intended to book revenue and tell investors that a licence had been signed. The parties agreed to the request and made the term sheet a contract in every respect. The party that asked for the binding commitment is the same one that later failed to meet its obligations.

The hidden risk: developers who had downed tools

The startup had stopped paying the people writing its code. The developers halted work, restricted access to the product and threatened to withhold and liquidate what they had built. The licensee discovered all of this after signing the term sheet.

For nearly a year, the licensee shuttled between the company and the developers to salvage the product. Developing independently was impossible: the technology had stayed outside the escrow and outside the delivery, while the developers threatened to walk away with the work they had done.

In the end, the licensee declared a default and asked the Court for an order compelling the transfer of ownership provided for in the term sheet, as reconstructed in the opinion of 4 September 2026^[1].

The mechanism is straightforward. An AI asset under development is worth as much as its authors' ability to keep writing it, and that ability depends on payments the licensee knew nothing about.

The governance signal for anyone licensing AI technology

The governance signal: for an AI technology still under development, the licensee's protection lies in the delivery, escrow and transfer clauses, together with verifying who pays the developers.

Regulation (EU) 2024/1689, the AI Act, has been in force since 1 August 2024 and governs the obligations of providers and deployers in the European Union. The transparency obligations of Article 50 apply from 2 August 2026. The Delaware case shows a different layer: the contractual governance that decides who owns the system when development stops.

A compliance posture calibrated to the public framework turns out to be over-calibrated relative to the contractual risk. The audit remains necessary; the perimeter has changed and now includes the developers' payment chain.

An emotion-detection system also touches Article 5 of the AI Act, which since 2 February 2025 has prohibited emotion recognition in the workplace and in educational institutions. The question of ownership is therefore joined by a question about the permitted scope of use in Europe.

Three decisions for the board

The case maps the choices a board must make before signing a binding term sheet on AI technology. I single out three.

1. Who is accountable, by name, for delivery and escrow.
2. Which transfer clause is triggered, and on what evidence.
3. Which disclosure accompanies the revenue booked on the licence.

First decision: which named role, in writing, before signing, is accountable for the delivery and escrow of the technology. An obligation without a named owner produces paperwork, and this case proves it. The General Counsel defines who verifies delivery and how often.

Second decision: which ownership-transfer clause is triggered in the event of default, and what evidence activates it. The Chief Risk Officer updates the risk framework to include dependence on external developers and the risk of restricted access to the code.

Third decision: which disclosure the Board Audit & Risk Committee requires on the booked value of the licence. Revenue recognised on a binding term sheet demands verification that the underlying asset exists and can be delivered.

The due diligence question left open

The licensee asked for guarantees on the technology and obtained robust clauses; it failed to ask who was paying the developers.

Classic due diligence checks patent ownership, open source licences and the cap table. For an AI system under construction, the decisive question is a different one: which role, in writing, guarantees that the authors of the code are paid and that the IP is deliverable at all times. The answer goes into the term sheet as a representation and warranty, with a right of periodic verification.

The first question, whether the term sheet was binding, was answered by the parties themselves. A second one has opened up: what is a binding commitment worth on an asset its authors can withhold.

The CEO finds a strategic limit here. The choice to license AI technology from a startup ties product strategy to a payment chain that the contract must make visible and verifiable.

Regulatory horizon

The memorandum opinion is a post-trial decision filed on 4 September 2026 and binds the parties to the case; its weight as precedent depends on the courts that will cite it. It concerns the contract law of one US state, applied to an AI asset under development.

On the European side, Regulation 2024/1689 is in force and the transparency obligations of Article 50 apply from 2 August 2026. Public enforcement proceeds in parallel: the EDPB reports that the CNIL fined Extia 300,000 euros for failing to respect the rights of data subjects (EDPB press release^[2]). The licensee of an AI system is therefore accountable on two planes, that of the contract and that of the regulator.

Organisations that today build in mandatory delivery, IP escrow, transfer on default and a named owner for the payment chain will face the 2027 deadlines with an advantage measurable in months of audit saved.