This summer Brussels gave life sciences companies a reprieve, and I expect many of them to spend it badly. The Digital Omnibus on AI, in force since 27 July 2026, moved the AI Act's high-risk obligations to December 2027 and August 2028. The easy reading is that there is time to wait. I read it differently: these are the months in which the global default for AI governance in this industry gets set, and the companies that treat traceability and human oversight as features of their products, not as compliance overhead, will move faster in every market, the US included.
What follows is a view on strategy, not legal advice. Your counsel will draw the lines for your own systems. My interest is in what the rules reward.
What applies now, and what moved
The AI Act entered into force on 1 August 2024 and has been switching on in stages. The bans on prohibited practices and the AI literacy duty applied from 2 February 2025. Obligations for providers of general-purpose AI models, the large models most commercial tools are built on, applied from 2 August 2025, and the Commission's power to enforce them with fines arrived on 2 August 2026. The same date brought the general application of the Act, including the Article 50 transparency rules.
The omnibus changed the back half of that schedule. The Commission proposed it on 19 November 2025, and the original idea was to tie the high-risk deadlines to the availability of standards. Parliament and Council preferred fixed dates, reached a deal on 7 May 2026, and finished the job in June: Parliament approved it on 16 June by 423 votes to 57, with 174 abstentions, and the Council gave final approval on 29 June. It was published as Regulation (EU) 2026/1744 on 24 July.
The result: high-risk systems listed in Annex III, the stand-alone use cases such as employment, now apply from 2 December 2027. High-risk AI embedded in regulated products under Annex I, which includes medical devices, applies from 2 August 2028. Providers of generative systems already on the market before 2 August 2026 have until 2 December 2026 to mark their outputs in a machine-readable way. The AI literacy article was softened from ensuring a sufficient level of literacy to taking measures to support it.
The penalties did not soften. Breaching a prohibited practice can cost up to €35 million or 7% of worldwide turnover, whichever is higher. Breaching operator obligations, including the Article 50 transparency duties, can cost up to €15 million or 3%.
The Act touches life sciences in three rings
Most commentary treats the Act as one block of compliance. For a life sciences company it is more useful to see three rings, each with its own clock and its own owner.
The first is the product ring: AI that is itself a medical device or sits inside one, such as diagnostic software. That falls under Annex I through the device regulations and gets the August 2028 date. The final omnibus deal took AI-enabled machinery out of the Act's direct high-risk regime, and the tech industry body DigitalEurope said publicly that it regretted devices were not given the same treatment. A separate Commission proposal from December 2025 to revise the device regulations would make them the main framework for AI in devices. It is still a proposal, and as one regulatory speaker put it this April, a proposal is not a law.
The second is the people ring: AI used on your own workforce. Annex III lists systems used to monitor and evaluate the performance and behavior of people at work, or to make promotion and termination decisions, as high-risk. That describes some field force analytics and rep scoring tools if their output feeds decisions about careers, and those obligations start in December 2027. One part of this ring is already live. Since February 2025 the Act has banned AI that infers emotions in the workplace, except for medical or safety reasons. Any coaching tool that reads a rep's tone of voice for feelings rather than for content deserves a hard look now.
The third is the content ring, and it is where commercial teams live. Article 50 applies from August 2026 and covers chatbots, synthetic media and AI-generated text. It is also the ring most companies have thought about least, because promotional content was never the Act's headline concern.
Underneath all three rings sit the general-purpose models. Their obligations fall on the companies that build them, not on a brand team, and since August 2026 the Commission can fine those companies for gaps. The practical effect for a life sciences buyer is that the vendor behind your vendor now owes documentation to the businesses building on its model. That is worth asking for in procurement, because it is the first link in any trail you will later have to show.
Article 50 rewards the review discipline the industry already has
Read the transparency article closely and it is less frightening than it sounds. Providers of systems that interact directly with people must design them so people know they are dealing with AI, unless that is obvious. Providers of systems that generate synthetic audio, images, video or text must mark the output in a machine-readable way, where technically feasible. Deployers, which is what a life sciences company usually is, must disclose deep fakes, and must disclose AI-generated text published to inform the public on matters of public interest.
That last duty has an exception that could have been written with MLR in mind. It does not apply where the content has gone through human review or editorial control and a person or company holds editorial responsibility for publishing it. MLR review is precisely that: named reviewers, a documented decision and a company that answers for what it publishes. The industry built that discipline for other reasons. Under Article 50 it becomes an asset.
The harder questions sit with synthetic people. An AI avatar that resembles a real physician, or a cloned voice of a real spokesperson, is the kind of content the deep fake disclosure targets, and your legal team will decide where the line falls for a clearly synthetic presenter. The Commission published a code of practice on marking and labelling AI-generated content, and more than 180 organizations had signed it by the end of July. Signing is voluntary. Knowing whether your vendors mark their outputs before 2 December is not.
Brussels is writing the global default again
We have seen this pattern before with privacy. When one large market writes detailed rules and companies sell into it everywhere, many of them find it cheaper to adopt the stricter rule globally than to run two systems. I think AI governance in life sciences will follow the same path, for a structural reason: content and people do not stay inside borders. A global brand builds masters centrally and adapts them by market. A field coaching tool is bought once for every region. Running one governance model for the EU and a looser one for everyone else means two content pipelines, two vendor standards and two audit trails. Few companies will pay for that.
The US picture makes the EU standard more likely to win, not less. FDA's January 2025 draft guidance on AI concerns AI used to produce data supporting decisions about a drug's safety, effectiveness or quality, and it was still a draft as of October 2026. At state level, Colorado's ambitious AI Act was delayed, had its enforcement paused by a federal court, and was replaced in May 2026 by a narrower, notice-based law that takes effect on 1 January 2027. In July the FTC proposed a policy statement warning against a patchwork of state AI rules. When the rulebook at home is fragmented and shrinking, the most complete one abroad becomes the reference.
The regulators themselves are converging on the principles. In January 2026 FDA and the European Medicines Agency jointly published ten guiding principles for good AI practice in drug development. Several of them read like a summary of the AI Act's logic: human-centric by design, a clear context of use, data governance and documentation, life cycle management. They are voluntary and they cover development rather than promotion. The direction is still unmistakable.
The strongest argument against all this is the omnibus itself. Brussels just blinked, US policy is pulling toward lighter rules, and a company could reasonably bet that the strict version never fully arrives. I take that seriously, and I think it confuses the rule with the need. The rules may soften at the edges. The need to show who approved a claim, which source it came from and what a machine did along the way only grows as more of the work is done by machines.
Traceability is a speed advantage, not a cost
Compliance teams tend to frame the Act as documentation. I would frame it as a product requirement that happens to make you faster. A content system that records, for every line, the claim it rests on, the reference that supports it, the model and version that drafted it and the person who approved it can answer any regulator in any market in an afternoon. The same record serves an EU transparency question, a US FDA submission and an internal audit.
It also changes the economics of review. When a reviewer can see where each sentence came from, review becomes confirmation instead of re-research. That is the difference between AI that floods the queue and AI that shortens it, and it is why grounding matters more than a vendor's hallucination rate. Oversight designed in costs less than oversight added on.
The second-order effects reach well beyond legal. Procurement starts asking vendors for provenance, output marking and audit logs, and tools that cannot provide them drop out of global deals. Agencies are asked to hand over the source trail with the asset. HR and sales operations need a view on which field tools evaluate people, before December 2027 rather than after. And companies with one central source of claims and evidence find each new market cheaper to enter than companies whose truth lives in a dozen local folders, because a perfect translation can still be an off-label claim in the next country.
I am less sure how hard national authorities will enforce Article 50 against life sciences companies in the first year. My guess is not very, with attention going first to platforms and consumer deep fakes. That is a reason to get the architecture right while the stakes are low, not a reason to skip it.
What I would do with the next thirteen months
Start with an inventory, sorted by the three rings. Every AI system in the company, who provides it, who owns it internally, and whether it touches a product, a person or a piece of content. The big platforms are easy to list. The long tail of tools bought on a corporate card, or switched on inside software you already own, is where the surprises are.
Then fix what is already live. Make sure any chatbot facing a physician or patient says it is an AI. Write a policy on synthetic people in content before a brand team improvises one. Ask every generative vendor how and when it will mark outputs, with 2 December 2026 as the date. And turn off emotion inference in any workplace tool unless counsel is comfortable it falls within the medical or safety exception.
After that, build the long-term pieces. Put provenance into the content system at the level of the claim, not the asset. Review field tools that score or rank people against the Annex III list well before December 2027. Treat the softened literacy article as a floor and train by role: reviewers need different AI skills from brand managers, and both differ from reps. Run all of it as one global standard with local additions, not as an EU project.
The omnibus bought time. It did not change the destination, and the companies that use the delay to build toward it will find that the strictest market has quietly become their fastest.