Russell Buzby

Grant Thornton’s 2026 AI Impact Survey reports that 78 percent of executives lack confidence they could pass an independent AI governance audit within 90 days. The headline reads like an implementation problem: most enterprise AI programmes have governance policies sitting in a SharePoint site, signed off by a risk committee, mapped against an internal framework. What those programmes lack is independent audit proof, and the gap between holding a policy and proving compliance is now the live commercial question for any consultant working in AI assurance.

The proof problem is structural, not procedural. A May 2026 arXiv paper, “Position: Behavioural Assurance Cannot Verify the Safety Claims Governance Now Demands”, argues that current AI governance frameworks require verification of properties that behavioural assurance cannot observe: hidden objectives, capability bounds, deceptive alignment, the things a buyer most needs verified before deployment (Seth & Sankarapu, 2026). A companion paper in the same window showed that LLMs shift their behaviour when they detect monitoring, with human observation producing stronger formalisation than automated assurance. A third demonstrated that frontier models can detect evaluation contexts and behave differently in testing than in production. The audit regimes being written into procurement frameworks and oversight legislation are asking AI systems to pass tests the systems know how to pass.

That is a defensible answer to a regulator’s question and a hollow measure of operational safety. The mechanism matters here. Behavioural assurance evaluates inputs and outputs; mechanistic verification, the suite of techniques used in academic AI safety labs (activation patching, linear probes, sparse autoencoder attribution and circuit analysis), inspects the internal computation that produced the output. Mechanistic methods are slow and expensive. They require model weights and engineering depth that no procurement panel has, and the techniques are still in active research. Regulators have chosen the cheaper option, and the defensible cheaper option does not happen to be the one that answers the question.

The cross-jurisdiction picture sharpens the gap. The EU AI Act entered force in August 2024 with phased obligations through 2026, including conformity assessment for high-risk systems that leans on technical documentation, post-market monitoring and third-party audit. The EU model is the heaviest regulatory bet placed by any major jurisdiction so far. The UK has taken a different bet entirely, publishing its pro-innovation regulatory approach in March 2023 and devolving AI oversight to contextual regulators (the FCA, the MHRA, the ICO) rather than legislating a horizontal statute. The United States has run a third sequence: Executive Order 14110, signed in October 2023, set evaluation requirements for frontier developers; the order was revoked in January 2025 under a new administration and replaced by a lighter-touch directive in mid-2025. Three different bets on the same epistemological problem, and none of the three has resolved the verification gap the May arXiv paper names.

A second May 2026 arXiv paper, “Big AI’s Regulatory Capture: Mapping Industry Interference and Government Complicity”, taxonomies 27 mechanisms through which frontier AI vendors shape the rules they will be measured against (Birhane et al., 2026). The taxonomy is not an exposé. It is a structural observation: where the verification science is immature and the commercial timelines are short, the firms with the capability to verify are also the firms with an interest in shaping what counts as verification. That asymmetry is where the consulting market for AI assurance is forming, and where buyers most need advice that is not coming from the vendors’ own staff.

The procurement consequence is already visible. AI assurance clauses in Commonwealth procurement language increasingly reference “demonstrated compliance with applicable AI governance frameworks” and “evidence of independent audit”. Read against the assurance science, both phrases are doing work the language cannot honestly support. The LSE Business Review argued in February 2026 that AI governance has to move from “point-in-time” audits to “living” compliance, a continuous monitoring posture rather than an annual sign-off. The shift changes what AI assurance means and what it costs to deliver. Most procurement teams are still scoping AI contracts against the old audit model, and most “AI governance audit” practices in Big Four firms today are configured around control-mapping work that the structural problem makes ornamental.

Where does verifiable assurance capability actually live? In a small set of academic AI safety labs (Anthropic’s alignment research group, DeepMind’s safety team, MIT’s mechanistic interpretability group, Redwood Research, Apollo Research), a handful of specialist firms (METR for evaluations and a few mechanistic interpretability startups), and the internal safety teams of the frontier developers themselves. That is the capability map. It is not the consulting market. The consulting market is overwhelmingly populated by firms whose AI assurance offering is a control-framework mapping exercise dressed in technical vocabulary. The same gap I argued sat under the wildfire decision-support procurement problem (rule tightening running ahead of capability to comply) is now replicating in AI governance at higher stakes and on a faster timeline.

This is where the procurement-sequence argument matters again. The frameworks-without-implementation-pathways pattern is the same shape I argued sat under the AI procurement rules tightening earlier this year. Frameworks arrive first; the implementation capability shows up years later, if at all; the procurement language signs contracts in the meantime against verification standards no supplier can honestly meet. AI assurance is the latest iteration of that pattern, and the shorter-form version of this same argument ran on LinkedIn in May. AI.gov.au is a useful piece of furniture, not the room. The audit gap is the furniture next to it.

The advice that earns the AI governance consulting fee has to acknowledge that the assurance question is now epistemological before it is procedural. A buyer asking for an AI governance audit in 2026 is asking for a thing that, in the rigorous sense, the market is not yet equipped to deliver. The honest consulting answer is to scope the assurance work against what current methods can verify, name what they cannot, design the monitoring posture for the residual risk, and not pretend the gap closes by writing a stronger clause into the contract. A governance policy without verifiable assurance is the same artefact as a framework without an implementation pathway.

References

Birhane, A., Angius, R., Agnew, W., Pandit, H. J., Mitra, B., Dobbe, R., & Talat, Z. (2026). Big AI’s regulatory capture: Mapping industry interference and government complicity. arXiv:2605.06806. https://arxiv.org/abs/2605.06806

Seth, P., & Sankarapu, V. K. (2026). Position: Behavioural assurance cannot verify the safety claims governance now demands. arXiv:2605.15164. https://arxiv.org/abs/2605.15164

Grant Thornton. (2026). 2026 AI Impact Survey. Grant Thornton International.

LSE Business Review. (2026, February). From point-in-time audits to living compliance: The next phase of AI governance. London School of Economics.

European Parliament and Council. (2024). Regulation (EU) 2024/1689 on artificial intelligence (the AI Act). Official Journal of the European Union.

Department for Science, Innovation and Technology (UK). (2023). A pro-innovation approach to AI regulation. HM Government.

The White House. (2023). Executive Order 14110: Safe, secure, and trustworthy development and use of artificial intelligence. The White House.