Anthropic restrictions expose a new risk: AI dependency

For much of the past two years, the marketing industry's conversation around artificial intelligence has focused on capability. Which model writes better copy? Which generates stronger creative assets? Which can improve productivity, automate workflows, reduce costs, or unlock new ways to reach consumers?

The recent curbing of access to Anthropic Claude's advanced AI models, Mythos and Fable, have shifted that conversation in a different direction. The two models are considered among the most advanced in the world, and they saw a large-scale rollout last week before being forced to pull back abruptly, without any warning to the millions of people and companies around the world who were reveling in its use cases.

The move, triggered by new US security restrictions on access to some of Anthropic's frontier AI systems, has sparked debate around AI sovereignty, export controls and the geopolitical implications of artificial intelligence. 

But for advertisers, agencies, publishers and technology firms, the episode raises a more immediate question: what happens when business-critical workflows increasingly depend on AI systems whose availability may ultimately be shaped by decisions beyond their control?

That question comes at a moment when AI is rapidly moving from experimentation to infrastructure. Across the marketing ecosystem, AI is increasingly embedded in content production, campaign planning, analytics, coding, audience segmentation and customer engagement. Agencies are building AI-assisted production workflows. Publishers are experimenting with AI-powered editorial tools. Technology firms are weaving AI into products and internal operations.

Read more on agencies going AI-native 

Across the industry, AI is rapidly becoming embedded into content creation, campaign planning, analytics, audience segmentation, coding, customer service and operational workflows. Agencies are building AI-native production layers. Publishers are experimenting with AI-assisted editorial processes. Technology companies are increasingly incorporating AI into their products and services.

The question, therefore, is no longer whether businesses should use AI. It is whether they are becoming too dependent on a handful of providers.

“Most companies haven't really thought about this,” says Prashant Puri, Co-Founder and CEO of AdLift. “They're using one or two AI tools, their teams are built around them, and nobody's asked what happens if access just... stops.”

According to Puri, that is precisely what makes the Anthropic episode significant. “A government order came in on a Thursday evening, and two products went dark globally, overnight, without any warning or transition period.”

While relatively few marketing teams were directly dependent on Mythos or Fable, he believes the incident exposes a broader vulnerability.

“The marketing industry is particularly exposed. AI is now sitting inside content workflows, campaign planning, analytics, and most of it is dependent on a handful of providers. That's a real business risk, not just a tech problem,” adds Puri.

The concern reflects a broader shift underway across the industry.

Over the past year, marketing conversations have increasingly moved beyond search engines and social platforms toward AI-driven discovery, recommendation engines and agentic systems. Agencies are investing heavily in AI-enabled production. Publishers are exploring how AI could reshape content creation and distribution. Marketers are beginning to optimise not only for visibility in search results, but also for inclusion in AI-generated recommendations and responses.

In many ways, the industry's current relationship with AI resembles its earlier relationship with cloud infrastructure. Most organisations do not build their own cloud platforms. They rent computing capacity from a handful of providers. The arrangement is efficient, scalable and generally reliable.

AI is increasingly evolving along similar lines. Rather than building foundational models themselves, most businesses access intelligence through a small group of companies including OpenAI, Anthropic, Google and Meta.

Ad models behind AI wars. Read more here 

Dr Siddhant Sethi, Senior Manager, Founder's Office at White Rivers Media, argues that the reality is more complicated. “Most companies are model-polygamous, not model-portable,” he says, noting, “They use several AI tools, but that does not mean they can switch between them smoothly.”

According to recent research from ADP, India is among the world's most active adopters of workplace AI, with 80% of employees using AI tools multiple times a week and 41% using them daily. Meanwhile, BCG's latest Global CMO Survey found Indian marketers among the most optimistic globally about AI-driven revenue growth.

So, while changing a model may seem technically straightforward, maintaining the same quality and consistency after the switch is often much harder.

“Prompts, workflows, tone, checks and internal logic are shaped around how one specific model behaves. Move to another, and things get weaker, less consistent or simply different in ways that are invisible to end users but critical to the people designing the system,” Sethi explains.

In other words, organisations may be using multiple models, but many workflows remain optimised around the behaviour of specific systems.

That distinction, Sethi argues, is where the real operational risk lies, pointing out, “What is new is the precedent. A company approaching a nearly hundred-billion-dollar public listing had its flagship products disabled overnight by a single government order. That tells you something about the power dynamics at play.” 

The challenge is likely to become more significant as AI becomes embedded deeper into business processes. Earlier generations of marketing technology largely enhanced existing workflows.

AI increasingly performs parts of the workflow itself. It writes, analyses, summarises, codes, recommends and increasingly acts. That distinction matters because replacing a tool is easier than replacing a capability around which teams have already reorganised themselves.

Gopa Menon, Co-Founder and COO of Theblurr, believes businesses have largely focused on the productivity benefits of AI while overlooking basic questions around resilience and risk management.

According to Menon, the implications extend beyond operational disruption. “As we get more and more dependent on these AI models and platforms, risks are real. The answer is now clear: access to frontier AI or newer models is a privilege that can be revoked without notice, without consultation, and without regard for the commercial relationships it disrupts.”

For Indian businesses, he argues, the issue is also strategic. “The concern isn't just operational, it's strategic. American AI models are bound to American geopolitics.”

That reality may force organisations to think about AI dependency in much the same way they already think about data privacy, media fraud or brand safety. 

“Over-reliance on one platform or model makes everything risky. That's a fundamental operating risk, and it deserves to sit in the same conversation as data privacy, brand safety, and media fraud,” says Menon.

Not everyone believes the industry faces an immediate crisis.

Prabhvir Sahmey, Founder of Stratpulse Labs, notes that large language models are not yet deeply embedded enough in most organisations to cause widespread disruption if a single provider becomes unavailable. “LLMs are yet to become business critical. So generally, everyone is safe,” he says.

However, he believes the episode offers a valuable lesson for organisations building AI-powered workflows. “My personal experience tells me that if we are going to use AI on core business functions, then it has to be an on-premise deployment, which prevents a situation where a complete blockade happens.”

Sahmey also points to a growing consensus among practitioners that redundancy is becoming essential. “It is now evident and more practitioners are recommending that especially if you are coding, then you need to have two LLMs. Relying solely on one can pose challenges.”

Taken together, the prevailing sentiments suggest the Anthropic restrictions may be less significant because of the models involved and more significant because of the questions they have raised. The marketing industry has spent the past two years debating which AI systems are the most capable. The next phase of the conversation may focus on something different: resilience.

As Raahul Seshadri, Director of AI and Tech at WebEngage, says,  “The key takeaway from Anthropic's episode is that long-term success in AI will come from the quality of your customer data, the business context in which that data is used, and how it is orchestrated, rather than the specific model you have access to. With AI becoming an integral part of marketing, customer engagement, analytics, and decision-making workflows, there will be an increasing number of foundation models that are interchangeable as part of a larger stack of technology within organizations.” 

As AI becomes embedded across content creation, campaign planning, audience analytics and customer engagement, organisations may increasingly need to ask not only what their preferred model can do, but what happens if it becomes unavailable.

Or, as Puri puts it, “If our primary AI provider went away tomorrow, what breaks?”