THE AI EDGE | Issue No. 1 | Weekly AI intelligence for executives on what's actually working in enterprise

Also published as The AI Edge on LinkedIn. Subscribe here → The AI Edge

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Covering: US Government Orders Anthropic to Suspend Fable 5 · The AI ROI Reckoning · ChatGPT Below 50% for First Time · EU AI Act 41 Days · EMEA: MFSA Deadline 8 Days

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## THIS WEEK AT A GLANCE

- US government orders Anthropic to suspend Fable 5 — the world's most capable AI model pulled offline over security concerns raised by Amazon's CEO to the White House. Access ends June 23. Enterprise AI supply chain risk is no longer theoretical

- The ROI reckoning lands: MIT finds 95% of GenAI projects have "little to no measurable" P&L impact. S&P Global: 91% of boards approved AI programmes; only 26% had the capability to execute them

- ChatGPT falls below 50% market share for the first time — 46.4%, down from dominance. Gemini at 27.7%, Claude at 10.3%. The AI market is fragmenting

- EU AI Act: 41 days to full applicability on August 2. MFSA Tokenisation Consultation closes in 8 days. Both clocks are live

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1: THE BIG STORY

When Governments Pull the Plug: The Fable 5 Shutdown and What It Means for Enterprise AI

This week, the US government issued a directive ordering Anthropic to suspend Fable 5 — currently the world's most capable AI model — citing national security concerns. The intervention was triggered by Amazon CEO Andy Jassy raising concerns directly with the White House. Anthropic's free trial window for Fable 5 closes today. After June 23, access requires paid usage credits. Negotiations between the administration and Anthropic are ongoing; President Trump described them as "going fine.

This is the first time a government has ordered a frontier AI lab to suspend access to its most capable model. Whatever the outcome of the negotiations, that precedent now exists. Governments can — and will — intervene in AI model availability when they judge it necessary. For enterprise executives who have built production systems on any single frontier model, the question this week is not about Anthropic specifically. It is about architecture.

Single-model dependency is an enterprise risk. It is the AI equivalent of single-vendor dependency in any other critical infrastructure — tolerated until it isn't, catastrophic when it materialises. The organisations that have built multi-model architectures — routing workloads across OpenAI, Anthropic, Google, and open-source alternatives — are happy this week.

The Fable 5 situation also accelerates a question that boards should already be asking: what is our AI continuity plan? If the model or vendor our operations depend on becomes unavailable how long does it take us to recover, and what does recovery cost?

Most enterprises do not have a good answer. This week is a reasonable moment to find one.

2: REGULATION & GOVERNANCE

EU AI Act: 41 Days. What Is Actually Enforceable on August 2.

August 2, 2026 is 41 days away. Clarity on what actually takes effect on that date matters now — because the Digital Omnibus amendments have created some confusion about which deadlines moved and which did not.

Here is what is enforceable and when:

Already in force: Prohibited practices (since February 2, 2025) — social scoring, manipulative AI, real-time biometric surveillance in public spaces — are live and have been for over a year. GPAI model rules have been in force since August 2, 2025.

August 2, 2026 — 41 days: Full applicability. Transparency obligations under Article 50 take effect. Any system generating AI content that interacts with humans must disclose it is AI. Any AI-generated audio, video, or image must be labelled as synthetic. This applies to customer-facing AI in financial services, iGaming, real estate, and any regulated sector. Fines: up to €15 million or 3% of global turnover.

December 2, 2027: Annex III high-risk AI systems (recruitment, credit scoring, law enforcement AI, critical infrastructure). The Omnibus extended this from August 2026 — which is the relief that was widely reported. But August 2 still lands in 41 days, and Article 50 with it.

The Digital Omnibus was politically agreed on May 7 but is not yet finally adopted. Monitor for formal adoption, which will trigger the revised timeline officially.

Action this week: Map every customer-facing AI touchpoint against Article 50. If your organisation uses AI in client communications, chat, or content generation, the disclosure architecture needs to be in place before August 2.

3: ENTERPRISE & INDUSTRY

The ROI Reckoning. The 26% Who Are Actually Executing

The numbers published this month are the most important data in enterprise AI right now, and they are not flattering.

MIT's NANDA report found that 95% of generative AI projects have delivered "little to no measurable" impact on profit and loss. S&P Global's banking survey found that 91% of boards have approved AI programmes — and only 26% of their organisations have the capability to actually execute them. The gap between board approval and operational capability is a governance and execution problem.

The broader dataset is consistent. 79% of organisations report real challenges in AI adoption. Only 29% see significant ROI from generative AI. Only 23% from AI agents. And 54% of C-suite executives say AI adoption is actively creating internal organisational stress.

None of this means AI does not work. It means most organisations are deploying it without the foundations that make it work: clean data architecture, clear ownership of AI outputs, governance frameworks that define what decisions AI can make and which require human sign-off, and measurement frameworks that connect AI activity to business outcomes rather than activity metrics.

The 26% who are executing share a common pattern. They did not start with the most sophisticated model or the most ambitious use case. They started with a specific, measurable problem, built the data and process infrastructure to support it, and measured outcomes against a defined baseline. That is not a technology insight. It is a management discipline.

Meanwhile, the AI market itself is fragmenting. ChatGPT's share of the global AI assistant market fell below 50% for the first time this week — to 46.4%. Google Gemini rose to 27.7%, Claude reached 10.3%. Critically, Claude's paid conversion rate is 13% — the highest in the field. That is a signal about who is actually deriving value: not the largest user base, but the most committed one.

4: EMEA LENS

Eight Days, Forty-One Days

Three things converging for EMEA operators this week.

MFSA Tokenisation Consultation: 8 days. The Malta Financial Services Authority's consultation on RWA tokenisation closes June 30 — eight days from today. Submissions to fintech@mfsa.mt. For firms advising on or operating Malta-based tokenisation structures, the submission must address the ownership model question and ensure consistency with EU Commission MiCA review positions, particularly where additional MiFID authorisation may be required alongside MiCA classification for some RWA structures. Do not miss this window.

EU AI Act: 41 days. As covered in Section 2, Article 50 transparency obligations land August 2. EMEA national competent authorities — MFSA, CBI, CSSF, BaFin, AFM — are the enforcement point for regulated entities. They are not waiting for Brussels to act first.

The broader picture for EMEA: 71% of financial services firms are now actively adopting generative AI, and 52% are deploying agentic AI specifically — up from around 10% just 18 months ago. The adoption curve is real. The execution gap documented in Section 3 above is equally real. The EMEA firms that close that gap in the next two quarters will be positioned ahead of everyone else.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## WATCH LIST

| 23 June 2026 | Anthropic Fable 5 free access ends — paid credits required | Today |

| 30 June 2026 | MFSA Tokenisation Consultation closes — fintech@mfsa.mt | 8 days |

| 2 August 2026 | EU AI Act full applicability — Article 50 transparency live | 41 days |

| 2 December 2027 | Annex III high-risk AI systems compliance deadline | 528 days |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## MY TAKE 

The Fable 5 shutdown is this week's headline. The MIT finding — 95% of GenAI projects with no measurable P&L impact — is this week's actual story.

We are two years into the enterprise AI era and the majority of organisations cannot demonstrate that their AI spend is producing returns. Boards approved the budgets. Vendors took the contracts. Pilots ran. Agents were deployed. And most of the time, the needle did not move in any way that a CFO could point to on a quarterly call.

That is not an argument against AI. It is an argument for doing it properly — which means starting with a problem worth solving, measuring outcomes against a defined baseline, and building the governance infrastructure that lets you know, with confidence, what your AI systems are actually doing and whether it is working.

The 26% who are executing are not smarter or better-resourced than the 74% who are not. They are more disciplined. That is the gap worth closing.

George Kakouras

---

The AI Edge is published weekly for informational purposes only and does not constitute legal, financial, or investment advice. Each edition covers enterprise AI deployment, strategy, and regulation for executives operating in EMEA

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━