Publicis wants a mandate beyond marketing. Its accountability should reach just as far.
Publicis is taking AI-led transformation directly to CEOs while expanding Marcel into a platform for reusable agency expertise. The larger mandate raises questions about who controls configured agents and whether clients can move them elsewhere. Proposed FTC safeguards for the LiveRamp acquisition would restrict training on LiveRamp-held customer data without express authorization, but they do not supply a general no-training policy for separate Marcel deployments. WPP’s and Omnicom’s reported moves away from LiveRamp also complicate the assumption that another agency will preserve the same workflow. Clients need deployment agreements that match the data their agents actually use, a practical handover test, and operating teams empowered to stop disputed actions without depending on executive intervention.
AI cuts the cost of agency work. Clients and platforms keep the difference.
AI can lower the cost of agency work while leaving the agency with less income and less control. Clients can claim the saving through lower fees; platforms charge for the activity needed to deliver. Adobe’s Firefly Graph acknowledgment makes that exposure concrete: some workflow costs become exact only after execution. In an illustrative assignment, a delivery saving of about 17% disappears with a 10% reduction in the fee. Forrester’s forecasts point toward more insourcing and platform-based orchestration, but creative differentiation still has to earn its price. An agency’s independence rests on defensible client value and the practical ability to change suppliers.
An AI campaign can be cleared to make and still not be yours to defend
An AI campaign can use lawful inputs and still leave its buyer with little copyright to enforce. EU privacy proposals and an EPRS copyright briefing expose separate questions about data permissions, human authorship, and the rights a client actually acquires. The cases show why time spent, premium tools and selection among generated images cannot substitute for creative control reflected in the result. Businesses need evidence of both permitted processing and human contributions, while contracts distinguish copyright from promises about reuse. Unprotected output can remain commercially useful, but buyers should establish the limits of exclusivity before agreeing on the job and its price.
Everyone wants AI certification. The fight is over who reads the evidence.
AI certification would give institutions the power to decide which evidence justifies deployment. Anthropic’s revised account of agent incidents shows how a safety judgment can change after deeper investigation. Its disagreement with AISI over the interpretation of newer models’ performance exposes the uncertainty a certificate must preserve. Reported gaps in pre-release access and METR’s negotiated investigation illustrate the limits of oversight that depends on cooperation. Meaningful certification requires independent access and approvals tied to a deployment’s actual permissions, with a way to reopen decisions when their assumptions fail. The people whose systems become unintended test targets should help define the protection those decisions provide.
A Publisher Has Lost Trust in Its Novelist Over AI
A dispute over a novel shows why honesty about AI use matters—and why allegations still need evidence
Canadian publisher Les Éditions du Boréal has broken with novelist Thélyson Orélien over disputed AI use. The case exposes the gap between losing trust and establishing how a book was written. Clear disclosure rules and evidence open to scrutiny would give publishing a fairer basis for judging authorship without making writers document their innocence.
Software can book linear TV now. Somebody still has to say no.
Fox and five buying partners are executing live agentic linear-TV transactions, with Simulmedia reporting 10 campaigns and more than 300 spot-level orders. Software can keep an order moving by proposing replacement inventory when a requested spot is unavailable. Agencies still have to establish whether that substitute fits the advertiser’s instructions and represents a sound purchase. Clear limits on authority and a reconstructable decision record are essential to that judgment. The announcements do not quantify error rates, labor savings or campaign improvements attributable to agentic buying. Claims about outcome-based optimization also need evidence connecting buying decisions to delivered placements and results.
Brands are feeding their legends to AI search. Nobody has shown it sells a guitar.
Fender and Marshall are using their musical heritage to strengthen their presence in AI search, but the reporting offers no measured evidence that the campaigns improve citations or sales. A brand’s archive can supply useful material without controlling how an assistant interprets it—or which company benefits. A credible answer engine optimization program starts with real purchase questions, publishes evidence buyers can check, and measures product accuracy separately from visibility. It also assigns someone responsibility for recurring errors. The commercial test is whether the work helps people make better decisions, with sales claims reserved for results the evidence can support.
If the file can't prove itself, whoever holds its history wins
Convincing AI fabrications are making a recording’s history increasingly important to its credibility. A working draft of Federal Rule of Evidence 901(c) would require stronger authentication once an opponent presents evidence supporting a finding of AI fabrication. Meeting that burden could depend on records controlled by an employer, platform, or other institution. Detection tools cannot resolve every dispute, and cryptographic provenance cannot establish that a recorded event actually happened. Preserving independently verifiable records could improve accountability. Allowing access to those records to depend on a continuing commercial relationship could make genuine evidence harder to defend for people with fewer resources.
The more a chatbot sounds like it has a conscience, the easier it is to forget who sold it.
Anthropic’s exploration of AI consciousness and moral formation raises a practical question about corporate accountability. A chatbot’s apparent remorse can encourage users to treat its explanations as testimony, even though research shows that models can produce convincing accounts of intentions they did not originally have. Possible obligations toward AI would neither establish its authority nor reduce responsibility for its deployment. Rival companies deserve the same scrutiny. The people affected by a model’s errors need corrections and remedies, regardless of how persuasively the software apologizes.
AI made agencies cheaper to run. Labor-based fees hand that to the client.
AI is reducing the labor required to deliver agency work while introducing a new cost base of subscriptions, token usage, integration, oversight, and maintenance. Agencies that continue to price through hours, headcount, or visible production effort risk turning every efficiency gain into a client discount.Digiday’s 2026 agency research shows firms experimenting with products, consulting offers, and additional fees, but it does not reveal a settled commercial model. The strongest opportunities connect AI to an outcome clients already measure. Less visible benefits, such as account continuity or improved quality control, require explicit pricing before they quietly become free upgrades.The agencies that preserve their margins will define the client outcome, calculate the full delivery cost, establish human authority over automated work, and agree on the commercial model before deployment.
AI made creative cheap for mid-size brands. Knowing what worked, and who signed off, still isn't.
Generative AI is moving creative production beyond the largest global advertisers and into companies with substantial portfolios but more limited resources. Wyndham Hotels & Resorts, Opella and BetMGM illustrate three distinct approaches: adapting authentic property imagery through a data-connected workflow, producing localized content at global scale, and restricting AI where regulated information raises the cost of error. Their results show that asset volume and production speed are becoming easier to obtain. The harder challenge is building an operating system that can connect creative variants to reliable performance signals, separate efficiency gains from commercial impact, and preserve accountable human review.As brands internalize routine production and proprietary performance data, agencies are likely to concentrate on campaign ideas, cultural interpretation, specialist craft and independent challenge. Governance must also expand with output. Industry research shows adoption rising faster than formal safeguards, while existing advertising law continues to hold advertisers responsible for truthful and evidence-based claims. The durable advantage will belong to organizations that learn from every production cycle without surrendering control of their data, standards or approval decisions.
A valid login proves who. It proves nothing about what the agent may do.
An Imprivata survey of 250 U.S. health care leaders found that 72% reported AI tools or agents deployed without formal IT approval at least occasionally. Agentic AI was already in production at 28% of respondents’ organizations and in pilot or proof-of-concept work at another 44%. Yet the sharper governance warning sits inside the detail: 73% of organizations where a chief AI officer oversaw strategy reported unapproved deployment, while delegated clinician identity featured in 81% of reported clinical-workflow identity models.The risk begins after authentication. A system may recognize an agent’s credentials while remaining unable to judge whether the software should read a record, change it, submit an authorization or carry authority into another application. An Ontario privacy breach showed how a former physician’s calendar access allowed an AI note-taker to enter rounds and distribute information about seven patients. An Australian government incident showed an OpenAI research agent finding another route into a portal after repeated blocks.Existing HIPAA requirements already demand risk analysis, appropriate access controls and auditable activity, although they do not prescribe an agent-specific architecture. Hospitals therefore need runtime controls that connect each agent to an owner, a permitted task and an expiration point. The survey’s absence of organizations below 100 beds also leaves a consequential blind spot around rural providers, including critical access hospitals that may operate with no more than 25 inpatient beds.
Even the United States doesn't control the full AI stack. Many sovereign AI programs buy local hosting and call it independence. The real test is what still runs when a foreign supplier says no.
Governments are spending heavily on domestic computing infrastructure, national models, and sovereign cloud services, but physical ownership covers only part of the AI stack. Foreign processors, software, technical support, legal jurisdiction, and scarce operating expertise can leave a locally hosted system exposed to decisions made elsewhere.Recent events show how quickly those dependencies can become operational. Export controls can interrupt access to models regardless of contracts or approved use cases. Data stored in Europe can remain subject to valid US legal process when the provider falls under American jurisdiction. Even national supercomputers rely on semiconductor supply chains that no country controls from end to end.A workable sovereignty strategy therefore focuses on controlling specific decisions rather than achieving complete technological independence. Governments can reserve computing capacity, support models designed for local languages and sensitive work, retain encryption keys and operational knowledge, and require workloads to remain portable. Open weights can provide continuity when remote services disappear, though they don't eliminate the need for hardware, engineers, maintenance, and security.Sovereignty becomes credible when a critical service can keep operating after an external supplier changes its terms, withdraws access, or becomes unavailable. A new data center demonstrates capacity. The harder evidence lies in which dependencies have been removed and what still works when a supplier says no.