The weekend that changed the AI governance conversation

TL;DR: The recent withdrawal of Anthropic's Fable 5 and Mythos 5 models highlights an emerging challenge for enterprise AI: model availability is becoming a governance concern. As organisations increasingly embed AI into business-critical workflows, dependence on a single model provider creates a new form of operational risk. The firms best positioned for long-term success will not be those that simply select the "best" model, but those that invest in knowledge architecture, governance, and composable AI platforms that allow models to evolve without disrupting users, workflows, or access to institutional knowledge.

(post first published on LinkedIn)

Over the past two years, enterprise AI conversations have often been dominated by a recurring question: which model should we choose?

The pace of innovation has certainly justified the attention. Every few months a new generation of models arrives with improved reasoning capabilities, larger context windows, better multimodal support, and benchmark scores that quickly become the focus of vendor marketing and technology discussions. Organisations evaluating AI initiatives have naturally spent considerable time comparing providers and trying to identify which model is most likely to deliver a competitive advantage.

However, the recent events involving Anthropic's Fable 5 and Mythos 5 models suggest that this focus may be too narrow.

Only days after Anthropic publicly released Fable 5, the company was instructed by the US government to disable access to both Fable 5 and Mythos 5. Anthropic complied within hours while publicly expressing disagreement with the rationale behind the decision. The details surrounding the decision will no doubt continue to be debated, but for enterprise leaders the more interesting question lies elsewhere.

The incident highlights a reality (a risk really) that many organization's yet to fully confront within their AI strategies: the availability of a foundation model is ultimately determined by parties outside the organisation's control.

For most technology investments, governance frameworks assume a degree of stability in the underlying platform. Organisations evaluate products, establish operating models, train users, define policies, and build processes around them with the expectation that the core capability will remain available long enough to justify the investment. Foundation models introduce a different dynamic. They are controlled by a relatively small number of providers operating within rapidly evolving commercial, regulatory, and geopolitical environments.

As AI becomes more deeply embedded in business operations, this distinction starts to matter.

Model Availability as a Governance Concern

When organisations discuss AI risk, the conversation typically focuses on now familiar themes such as hallucinations, bias, security, privacy, and regulatory compliance. These concerns are important and deserve continued attention. However, the Anthropic situation highlights another dimension of risk that has received far less discussion.

What happens when a model that underpins business processes is no longer available in the form originally expected?

The answer is clearly not always straightforward.

In many organisations, AI is moving beyond experimentation and becoming embedded within operational workflows. Knowledge assistants are being integrated into daily work. Copilots are becoming part of document creation, research, analysis, and decision support activities. Legal teams are incorporating AI into drafting and review processes. Professional services firms are exploring how AI can improve knowledge access and service delivery.

As adoption increases, the cost of change increases alongside it.

A change in the underlying model may require new validation exercises, adjustments to governance controls, updates to operating procedures, retraining of users, and in some cases a reassessment of risk assumptions. Even when alternative models are available, transitioning between them is rarely frictionless if the surrounding architecture has been tightly coupled to a particular provider.

The challenge is not that model providers are unreliable. Rather, it is that organisations consuming those models have limited influence over the decisions that shape their future availability.

From a governance perspective, that reality deserves greater attention than it currently receives.

This matters even more within Professional Services

The implications are especially significant for law firms and professional services organisations, where knowledge sits at the centre of value creation.

Unlike consumer applications, professional services firms are not adopting AI simply to automate generic tasks. They are integrating AI into environments where trust, expertise, provenance, and professional judgement matter. The quality of outcomes depends not only on the capabilities of the model but also on the quality of the knowledge and context available to it.

This is one reason why AI outcomes vary so widely between organisations.

As I discussed in a previous newsletter, AI systems consistently perform better where content is current, structured, governed, and clearly owned. They perform far less effectively when content is fragmented, duplicated, outdated, or poorly maintained. AI often reveals the state of an organisation's knowledge environment more than it transforms it.

The same principle applies here.

An organisation whose AI strategy is heavily dependent on a particular model may find itself exposed to changes that it cannot easily accommodate. An organisation that has invested in its knowledge architecture, however, has considerably more flexibility. It can evaluate new models, introduce specialist capabilities, and adapt to changing market conditions without having to rebuild the foundations underneath.

The distinction may seem subtle, but it has important strategic implications.

The real asset was never the model

Much of the current AI market still behaves as though the model itself is the primary source of value.

That assumption becomes harder to defend as models continue to converge in capability and as access to frontier models becomes increasingly commoditised.

A law firm's competitive advantage does not originate from access to a particular model. The same model is generally available to competitors. Nor does advantage come from adopting a new capability a few weeks before the rest of the market.

The more durable source of value lies in the firm's accumulated expertise, its precedents, client experience, methodologies, playbooks, and institutional knowledge. It lies in the quality of its information architecture, the governance of its knowledge assets, and its ability to connect people with the information they need to perform effectively.

The model provides a powerful interface to those assets, but it does not replace them.

This perspective has become increasingly important as organisations move from AI experimentation to operational deployment. The firms generating the most sustainable value are often those that treat knowledge as a strategic asset and AI as a means of unlocking it more effectively. They invest in governance, ownership, content quality, metadata, retrieval strategies, and information architecture because they recognise that AI outcomes depend heavily on those foundations. So, while the model matters, it is not where the lasting advantage resides.

The emergence of multi-model architectures

A second trend reinforces this point.

Increasingly, enterprise architectures are moving away from the assumption that a single model will serve every use case. Different models exhibit different strengths. Some may perform better for reasoning-intensive tasks, others for research, summarisation, multilingual work, or cost-sensitive workloads. Regulatory requirements may also influence model selection in ways that vary across jurisdictions and business functions.

The logical outcome is an architectural approach that treats models as interchangeable components within a broader ecosystem.

This is one reason why concepts such as MCP, orchestration layers, model routing, and multi-source retrieval architectures are attracting growing attention. As I discussed in my February newsletter, the strategic challenge is increasingly about connecting models to governed knowledge, systems, and workflows rather than selecting a single model and building everything around it.

Microsoft's direction with Azure AI Foundry reflects this broader industry shift. The emphasis is increasingly on orchestration, evaluation, governance, and retrieval across multiple sources and models rather than on any individual model itself.

For organisations designing their long-term AI strategy, this is an important development. The objective is no longer simply to choose the best model available today. It is to build an architecture capable of adapting as models, providers, and requirements inevitably evolve.

Looking beyond the headlines

The lesson from the Anthropic episode is not that organisations should avoid a particular provider, nor that regulatory intervention is inherently problematic.

Rather, it serves as a useful reminder that foundation models are becoming a critical dependency within enterprise architectures. As that dependency grows, resilience becomes just as important as capability.

For CIOs, Chief Knowledge Officers, innovation leaders, and technology teams, this raises a useful question: if a primary model became unavailable tomorrow, how easily could the organisation adapt without disrupting users, workflows, governance controls, or knowledge access?

The answer says a great deal about the maturity of the underlying architecture.

Models will continue to evolve. Some will improve dramatically. Others may disappear. New entrants will emerge, regulations will change, and commercial dynamics will continue to shift. Organisations cannot control those developments, but they can control how dependent they become upon them.

The firms best positioned for the next phase of enterprise AI are unlikely to be those that simply selected the right model at the right moment. They will be the firms that invested in knowledge architecture, governance, and composability, creating an environment where models can change without forcing the organisation to start over.

In the long run, that may prove to be the more important design decision.