---
title: "AI Risk"
date: 2026-07-12
description: The AI risks that bite are seldom on the register, and the bill for sovereignty shocks, readiness gaps, verification costs, and model risk arrives later.
author: Mario Thomas
canonical: https://mariothomas.com/briefings/ai-risk/
---

## Start here

Two short reads before you go deeper: what AI risk actually is and where its costs hide, then the order to take this briefing in.

### The Bill That Has Not Yet Been Presented

On 12 June 2026, a single government directive forced a provider to withdraw two frontier models from every customer overnight, including organisations the order was never aimed at. Nothing degraded and nothing failed. A capability simply disappeared, removed by a party those organisations had no standing to appeal to. That is the character of the risk this briefing deals with: not the risks the AI headlines describe, but the ones created quietly by ordinary, defensible choices, each invisible until the day it is tested.

Most of the Boards I meet treat AI risk as a technology category and delegate it accordingly. The result is a register that prices the visible costs (licences, compute, and programme spend) and misses the structural ones: dependence on a single model and jurisdiction, investment mistaken for readiness, code generated faster than anyone can verify it, and a model chosen because it dominated the headlines rather than because it fitted the work. The evidence on readiness alone is stark. **89%** of enterprises have adopted AI tools, yet only **23%** can measure the return (Larridin 2025). That gap is not a measurement problem. It is a mirage of maturity, and it means much of the risk sits exactly where the register shows capability.

The position these articles take is that every convenient AI choice carries two costs, and only one is presented before it is paid. Frontier capability bought cheaply and governed elsewhere is paid for in sovereign control. Tool adoption without balanced capability is paid for in stalled transformation. AI-generated output without the expertise to evaluate it is paid for in technical debt, which is why I call verification a premium: it is the part of the price most organisations discover they still owe. None of these is a problem to be solved. Each is a trade to be taken knowingly or taken by default, and default is the more common.

What a Board should be able to do by the end of this briefing is ask, of any material AI dependency, three questions: which cost are we paying, did we choose it or back into it, and has the hidden half of the bill been priced? A Board that can answer those has not eliminated its AI risk. It has done something more useful: it has seen the exposure, named an owner, and chosen its position rather than discovering it.

### From Dated Event to Deliberate Choice

Start with [the articles](#core-reading), and take the core reading in order. *The AI Sovereignty Trilemma* opens the briefing with a dated, documented event: a frontier capability withdrawn from production overnight by an order nobody affected could appeal. It establishes the discipline the rest of the briefing works from, that model availability is a continuity risk the Board owns, and that the convenient choice is the one whose price has not yet been presented.

*The AI Maturity Mirage* then turns from the external shock to the internal one: the gap between what organisations have invested and what they can actually do, and the three patterns of overestimation that keep that gap invisible. *The Verification Premium* takes the same gap down to where much AI spend now lands, the generation of code, showing why AI output without the expertise to evaluate it accumulates debt rather than value. *Selecting Your Enterprise LLM* closes the sequence with the decision sitting beneath the others: how to choose a model on strategic value, risk, readiness, and sustainability rather than on hype.

The further reading deepens each thread. The original sovereignty analysis sets out the Trilemma at the level of national regimes, trust against speed against control; the June piece restates it at the model layer, where the surrender of control was first felt. The shadow AI and agentic technical debt pieces trace where ungoverned usage and legacy architecture quietly concentrate exposure. *The Balancing Item* prices the oversight labour that verification demands, the hours a business case seldom counts. *The Accountability Gap* holds that accountability stays with the humans who deployed the system, however much of the work it now does, and the ethical AI piece shows that a model arrives carrying its provider's ethical standard, which the organisation inherits by default unless the Board decides otherwise. And *The Reasoning Gap* shows where these risks now meet the law, inside systems many Boards have already approved.

Then go to [Remake](#remake-assets) for the mechanisms that turn the thinking into apparatus a Board can actually use, and finish with [the questions](#faqs), which are the ones I would want answered before accepting any assurance that AI risk is understood.

## Core reading

The core sequence runs from a capability withdrawn overnight to the discipline of choosing a model deliberately, with the further reading tracing where the exposures concentrate.

1. [The AI Sovereignty Trilemma: When a Frontier Model Vanishes and Reality Bites](https://mariothomas.com/blog/ai-sovereignty-trilemma-reality/) (10 minute read, 14 June 2026): The visible cost of sovereignty deters Boards. The hidden cost of the convenient alternative was never shown, and that is the cost 12 June presented. Podcast edition: 11 minute listen.
2. [The AI Maturity Mirage: Diagnosing the Gap Between Investment and Readiness](https://mariothomas.com/blog/ai-maturity-mirage/) (11 minute read, 7 December 2025): Boards overestimate AI maturity by counting tools and pilots rather than capability. Three patterns create the illusion, and each can be diagnosed before it misleads. Podcast edition: 16 minute listen.
3. [The Verification Premium: What Classical Training Reveals About AI Coding Costs](https://mariothomas.com/blog/vibe-coding-vs-classical-training/) (13 minute read, 25 January 2026): AI coding tools amplify the expertise gap rather than closing it: senior developers capture twice the gains. The verification premium is the cost nobody budgets. Podcast edition: 18 minute listen.
4. [Selecting your enterprise LLM: Moving beyond the hype to make the right choice](https://mariothomas.com/blog/choosing-the-right-llm/) (10 minute read, 3 January 2025): With well over a hundred language models available, choosing one is a question of fit, not headlines: match the model to the task.

## Further reading

- [AI Sovereignty: A Board's Guide to Navigating Conflicting National Agendas](https://mariothomas.com/blog/ai-sovereignty-board-trilemma/) (15 minute read, 7 September 2025): AI governance is fragmenting into incompatible systems: Europe's transparency, America's scale, China's control. Boards can no longer serve all three; they have to choose.
- [Shadow AI and the Case for an AI Amnesty](https://mariothomas.com/blog/shadow-ai-amnesty-governance/) (15 minute read, 21 September 2025): Shadow AI is surging and most employees would use AI tools without permission. An AI amnesty turns that hidden risk into governed, employee-validated innovation.
- [How Agentic AI Turns Your Biggest Tech Problem into Competitive Advantage](https://mariothomas.com/blog/agentic-ai-technical-debt/) (11 minute read, 3 August 2025): The legacy estate that constrains agentic AI is also its largest opportunity. Retiring technical debt is what clears the path for autonomous systems.
- [The Reasoning Gap: The Capability the Law Now Demands of Boards](https://mariothomas.com/blog/the-reasoning-gap/) (11 minute read, 3 May 2026): UK law now requires four safeguards for solely automated decisions. Most Boards have approved probabilistic systems that cannot deliver them in operation. Podcast edition: 12 minute listen.
- [The Balancing Item: The AI Oversight Cost Your Business Case Never Priced](https://mariothomas.com/blog/unpriced-cost-ai-oversight/) (10 minute read, 26 July 2026): Every AI business case counts the hours saved. Almost none counts the oversight hours added, and people are silently absorbing the difference. Podcast edition: 13 minute listen.
- [The Accountability Gap: When AI Delegation Meets Human Responsibility](https://mariothomas.com/blog/ai-agency-accountability/) (15 minute read, 16 November 2025): Organisations are transferring decision-making agency to AI while accountability stays with people, and approving deployments without the verification capability that accountability needs.
- [Ethical AI: When the Model Imposes Values Your Organisation Did Not Choose](https://mariothomas.com/blog/ethical-ai-inherited-values/) (14 minute read, 17 May 2026): A foundation model arrives with a value system its provider built and the Board did not choose. The decision: accept it, reject it, or build. Podcast edition: 15 minute listen.

## Remake

The mechanisms beneath the thinking: the model, diagnostic, methodology, and principle from the Remake Library that turn this briefing into apparatus a Board can use.

- **Model: AI Sovereignty Trilemma**. The proposition that organisations and jurisdictions can optimise their AI posture for trust, speed or control, but not all three simultaneously, forcing deliberate strategic positioning rather than attempting to serve every market at once. [Remake Library](https://mariothomas.com/remake/library/ai-sovereignty-trilemma/)
- **Diagnostic: AI Amnesty Questionnaire**. The structured instrument for running an AI amnesty: ten sections capturing which tools employees actually use, the use cases they serve, the data they touch, the value already created, and the risks encountered, turning unknown unknowns into a governable inventory. [Remake Library](https://mariothomas.com/remake/library/ai-amnesty-questionnaire/)
- **Methodology: AI Amnesty**. The time-boxed programme that brings shadow AI under governance: a declaration window, typically 30 to 45 days, in which employees disclose all AI tool usage without fear of punishment, followed by the post-amnesty roadmap of rapid triage, governance guardrails, pilot launch, and ongoing operations. [Remake Library](https://mariothomas.com/remake/library/#ai-amnesty)
- **Principle: Minimum Lovable Governance**. Governance embedded in how work happens: proportionate to risk, continuous rather than episodic, and used because it works. [Remake Library](https://mariothomas.com/remake/library/minimum-lovable-governance/)

## Questions

The questions I would put to any Board seeking assurance that its AI risk is understood, owned, and priced.

### Could a provider withdraw the models we depend on overnight?

It already happened to organisations that were never the target. On 12 June 2026 a national-security directive forced a provider to switch off two deployed frontier models for every customer, and those cut off lost the capability as collateral of someone else's compliance. Model availability is a third-party continuity risk that belongs on the risk register with a named owner, and the test is one line: if this model disappeared tomorrow morning, what would stop working by lunchtime? I set out the full argument in [the Trilemma made real](/blog/ai-sovereignty-trilemma-reality/).

### We have AI tools deployed in every function. Why does that not count as maturity?

Because visible activity bears little correlation to capability. **89%** of enterprises have adopted AI tools, yet only **23%** can measure the return on that investment (Larridin 2025), and the gap is produced by three reinforcing patterns: counting tools as capability, mistaking pilot wins for systemic readiness, and hype-driven metrics that reward short-term ROI over capability. [The AI Maturity Mirage](/blog/ai-maturity-mirage/) sets out a diagnostic for finding the organisation's true position before further investment widens the gap.

### Can AI coding tools reduce our dependence on expensive engineering talent?

The evidence suggests the opposite. Senior developers capture roughly twice the productivity gains of juniors (McKinsey 2023), and a randomised controlled trial found experienced developers were actually **19% slower** with AI tools once verification and correction costs landed (METR 2025). Expertise is not what AI coding replaces; it is what determines whether AI-generated code creates value or debt. The Board question is not whether AI can write code cheaper but whether the organisation holds the verification capability to know. I explore this in [The Verification Premium](/blog/vibe-coding-vs-classical-training/).

### Is the most capable frontier model the safest choice?

Not necessarily. The safest model is the one that fits the work, judged across strategic value, risk, organisational readiness, and operational sustainability, and different use cases will justify different models. The choice is made per deployment, not once for the organisation, and it now carries a jurisdictional dimension: two providers in the same jurisdiction, subject to the same directive, are one exposure wearing two names. Start with [Selecting Your Enterprise LLM](/blog/choosing-the-right-llm/) and read it alongside [the Trilemma made real](/blog/ai-sovereignty-trilemma-reality/).

### Our people are using AI tools we never approved. Should we shut that down?

Enforcement fails on the numbers: **54%** of employees say they would use AI tools even if unauthorised (BCG 2025), and **57%** conceal their usage (Gigster 2025), so prohibition simply drives the risk underground. A time-limited [AI amnesty](/blog/shadow-ai-amnesty-governance/) converts the blind spot into governance visibility and captures the innovation employees have already validated. And the approved estate deserves the same scrutiny: many Boards have signed off automated decision systems that cannot yet deliver the safeguards UK law now requires, the exposure I call [the Reasoning Gap](/blog/the-reasoning-gap/).

### Which of our critical processes depend on a single model or jurisdiction?

In my experience few Boards can answer this from the chair, because the dependency was assembled by default, through procurement choices each defensible on its own and never examined together as a position. The credible answer is a per-deployment map: which processes rely on one model, which of those are customer-facing or regulated, and whether the fallback would survive the specific event. The map is only honest if it counts jurisdiction as well as vendor, since a second provider under the same directive is no fallback at all. The AI Sovereignty Trilemma model frames the trade, and [the Trilemma made real](/blog/ai-sovereignty-trilemma-reality/) shows the bill once it is presented.

### Who owns model availability on our risk register?

Often nobody, because model availability has been filed under IT as an operational setting rather than carried as a third-party continuity risk with a named owner. Continuity planning was built for outages, degradation, cyber incidents, and supplier failure, events the provider is working to reverse. A compelled withdrawal sits outside all of them: the provider stays up, the infrastructure stays up, and the capability goes anyway. The Board's task is not to choose the vendor or the architecture but to require that the exposure is known, owned, and priced, which is the Minimum Lovable Governance principle applied where it matters. [The Trilemma made real](/blog/ai-sovereignty-trilemma-reality/) sets out the argument.

### Does our AI business case price the oversight labour it assumes?

Almost never, in my experience. Business cases count the hours AI saves and seldom the hours it adds: the reviewing, correcting, and supervising that outputs demand before anyone can rely on them. That labour lands on existing people on top of existing jobs. A survey of 1,488 US workers found **14%** of those using AI reporting mental fatigue, the strain attaching to oversight load rather than to AI use itself (BCG Henderson Institute 2026). A control asserted but not resourced exists on paper only, so every proposal should state who performs the oversight, what share of their capacity it consumes, and what that costs fully loaded. [The Balancing Item](/blog/unpriced-cost-ai-oversight/) prices the line the business case left out.

## References

The external research and primary sources these articles draw on, for directors who want the evidence at first hand.

- **METR** (10 July 2025): [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/). Randomised controlled trial finding experienced developers 19% slower with AI coding tools, the clearest measurement yet of the verification premium.
- **GitClear** (2025): [AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones](https://www.gitclear.com/ai_assistant_code_quality_2025_research). Analysis of 211 million changed lines of code showing an eightfold rise in duplicated blocks in AI-assisted codebases.
- **Google Cloud** (22 October 2024): [Announcing the 2024 DORA report](https://cloud.google.com/blog/products/devops-sre/announcing-the-2024-dora-report). The 2024 DORA report, finding a 25% increase in AI adoption correlates with a 7.2% decrease in delivery stability.
- **MIT Sloan Management Review** (18 August 2025): [The Hidden Costs of Coding With Generative AI](https://sloanreview.mit.edu/article/the-hidden-costs-of-coding-with-generative-ai/). Economic modelling of how technical debt from AI-generated code quickly eclipses the short-term productivity gains.
- **J.P. Morgan** (17 November 2025): [Vibe Coding: A Guide for Startups and Founders](https://www.jpmorgan.com/insights/technology/artificial-intelligence/vibe-coding-a-guide-for-startups-and-founders). Investor due-diligence questions on AI-generated code, from verification process to technical debt profile.
- **Gartner** (30 June 2025): [Gartner Survey Finds 45% of Organizations With High AI Maturity Keep AI Projects Operational for at Least Three Years](https://www.gartner.com/en/newsroom/press-releases/2025-06-30-gartner-survey-finds-forty-five-percent-of-organizations-with-high-artificial-intelligence-maturity-keep-artificial-intelligence-projects-operational-for-at-least-three-years). Survey evidence that high-maturity organisations sustain AI value for three years or more and earn markedly higher stakeholder trust.
- **McKinsey & Company** (5 November 2025): [The State of AI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai). The State of AI research identifying leadership alignment, not technology capability, as the bottleneck in enterprise AI value.
- **BCG** (June 2025): [AI at Work 2025: Momentum Builds, but Gaps Remain](https://web-assets.bcg.com/fd/0d/bcc5dfae4cbaa08c718b95b16cf5/ai-at-work-2025-slideshow-june-2025-edit-02.pdf). AI at Work 2025, finding 54% of employees would use AI tools even if unauthorised, the workforce reality beneath shadow AI.
- **legislation.gov.uk** (2025): [Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents). The Data (Use and Access) Act 2025, the statute behind the four safeguards now required for solely automated decisions.
- **Institute of Directors** (2025): [AI Governance in the Boardroom](https://web.archive.org/web/20251121075957/https://www.iod.com/app/uploads/2025/09/AI-Governance-in-the-Boardroom-1c7612e872fa3fce3f9d6cad78b0b4ba.pdf). The IoD business paper on AI governance in a sector-led UK regulatory environment, with the ICO as a central actor.
- **Harvard Business Review** (1 March 2026): [When Using AI Leads to “Brain Fry”](https://hbr.org/2026/03/when-using-ai-leads-to-brain-fry). BCG Henderson Institute research in HBR: a survey of 1,488 full-time US workers identifying mental fatigue that attaches to heavy AI oversight loads, with replacement-pattern AI use associated with lower burnout.
- **Anthropic** (12 June 2026): [Statement on the US government directive to suspend access to Fable 5 and Mythos 5](https://www.anthropic.com/news/fable-mythos-access). The 12 June 2026 provider statement on the US directive that suspended two frontier models for every customer: the continuity risk that opens this briefing.
- **Larridin** (20 November 2025): [The State of Enterprise AI in 2025: From Experimentation to Accountability](https://www.larridin.com/blog/state-of-enterprise-ai-in-2025). The 2025 enterprise survey behind the 89% adoption against 23% measurable return figure, the gap the maturity mirage describes.
- **McKinsey & Company** (14 June 2023): [The economic potential of generative AI: The next productivity frontier](https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier). McKinsey's 2023 analysis of generative AI's economic potential, the source of the finding that senior developers capture roughly twice the productivity gains of juniors.

## The ideas beneath this briefing

Ideas I've named and matured writing about AI Risk: what each one means, and where it started.

- **Sovereign Control**: The ability to keep an AI capability running on terms the organisation sets, rather than terms a distant provider, or a government, can revise without consultation; more than data residency. [Read more](https://mariothomas.com/blog/ai-sovereignty-trilemma-reality/)
- **AI Maturity Mirage**: Mistaking visible tool deployments and isolated pilot wins for genuine organisational capability, a systematic overestimation that derails transformation strategies. For a Board, correcting it means diagnosing actual capability against AISA and the Five Pillars rather than trusting the appearance of activity. [Read more](https://mariothomas.com/blog/ai-maturity-mirage/)
- **Hype-Driven Assessment Metrics**: Judging AI progress by short-term ROI and perceived importance rather than actual integration, an overestimation pattern that hardens where organisations fail to track AI impact at all. [Read more](https://mariothomas.com/blog/ai-maturity-mirage/)
- **Pilot Success Trap**: Isolated pilot wins that create a seductive appearance of advancement while revealing nothing about systemic readiness across an organisation whose functions sit at different stages. [Read more](https://mariothomas.com/blog/ai-maturity-mirage/)
- **Tool-Centric Illusion**: Counting AI tools deployed as evidence of maturity when, without integrated infrastructure, those deployments create capability silos rather than organisational transformation. [Read more](https://mariothomas.com/blog/ai-maturity-mirage/)

All concepts: https://mariothomas.com/glossary/concepts/

## More Board Briefings

More complete resources on AI and emerging technology for the Boards that need the full picture.

- [AI Governance](https://mariothomas.com/briefings/ai-governance/): Governance people route around fails to govern; the task is governing AI the Board cannot fully see without strangling adoption.
- [AI Regulation](https://mariothomas.com/briefings/ai-regulation/): AI regulation has split into incompatible regimes, and the Board's task is not compliance with each but a deliberate position across all of them.
- [AI Accountability](https://mariothomas.com/briefings/ai-accountability/): Agency can move to the machine; accountability cannot, and answering for what AI decides now takes capability that policy alone does not supply.
- [AI Infrastructure](https://mariothomas.com/briefings/ai-infrastructure/): AI infrastructure now runs from the power station to the protocol, and energy access decides what an organisation can do with AI.