AI Accountability

Yes, and the record is clear. Courts, regulators, and clients have consistently rejected ’the AI did it’ as a defence: an airline held liable for its chatbot’s advice, lawyers sanctioned for fabricated citations, a major consultancy partly refunding a government client. Agency for the work can move to a machine; accountability for the outcome stays with the humans who deployed it. I set out the evidence, and the strategic choice it forces, in The Accountability Gap.

Since 5 February 2026, the UK GDPR as amended by the Data (Use and Access) Act 2025 has required four safeguards for any significant decision taken solely by automated processing: information about the decision, the ability to make representations, human intervention, and the right to contest. On the page these are procedural rights. In operation they are capability tests, and most probabilistic systems cannot pass them unless the capability was engineered in at design time. The Reasoning Gap works through what that means for the systems a Board has already approved.

The evidence points the other way: the more an organisation delegates to AI, the more expertise it needs to verify the outputs. Verification capability comes from years of doing the work, which is why replacing junior roles to fund AI efficiency quietly destroys the pipeline that produces the senior experts who check the machine. The choice between augmentation and replacement, and its ten-year consequences, is the second half of The Accountability Gap.

In domains that can be formally specified, yes. Automated reasoning proves properties across every possible state of a system rather than the scenarios someone thought to test, which is why civil aviation and the TLS 1.3 protocol already use it, and why financial institutions are beginning to apply it to capital adequacy calculations. For a Board this arrives as a fourth indicator type: reasoned indicators, which prove what must hold true rather than estimate what is likely. From Probable to Provable explains the discipline, and Maximum Fidelity shows all four types applied to real decisions.

Unless the Board has done deliberate work, the provider’s. Every foundation model arrives with a value system built upstream in pre-training and alignment, and system prompts, retrieval, and guardrails constrain that standard without re-authoring it. The real decision, taken deployment by deployment, is to accept the provider’s standard, reject the deployments where it bears directly on people, or build alignment the organisation owns. Ethical AI sets out how to make that choice deliberately rather than inherit it by default.

In most business cases I see, nobody is named. The case counts the hours AI saves and leaves out the hours it adds: the reviewing, correcting, and deciding-whether-to-trust its outputs demand. That labour lands on existing people on top of existing jobs, and BCG Henderson Institute research published in 2026 found the strain attaching to oversight load, not to AI use itself. A right to human intervention on demand needs standing capacity to answer it. The credible answer names the roles, states the capacity removed from their workload, and prices it fully loaded. The Balancing Item sets out the three questions that test whether it has been priced.

It should be able to, and the Institute of Directors’ 2025 paper on AI governance in the Boardroom expects the Board to retain exactly that authority. The authority is only as real as the mechanism beneath it. A foundation model’s value standard moves with every new version, without a fresh approval, so the Board needs to know which deployments it has accepted, which it has rejected, and what would trigger intervention on each. Agentic AI names the kill-switch as a governance minimum for any system running its own loop, and Ethical AI sets out the accept, reject, or build choice that gives the authority something to act on.

Few of the Boards I meet can say, and the inventory is the first priority The Reasoning Gap sets. A rule-based system carries its reasoning on the surface: the rule applied, the facts it operated on, the output that followed. A probabilistic system, of the kind approved for credit decisioning and fraud detection, does not, and the four safeguards UK law has required since 5 February 2026 have to be engineered in at design time. Most production systems are hybrid, so the question belongs at the point where the consequential output is produced, with the engineering function in the room. The Reasoning Gap sets out the inventory.