AI Doesn't Fail on Capability. It Fails on Governance.

Grant Thornton surveyed 950 US business leaders in 2026 and asked what's actually behind AI projects that underperform or fail. The top answer wasn't model quality, data problems, or a talent gap. Forty-six percent named governance or compliance failures – well ahead of the second-place factor, insufficient training, at 31%. That fifteen-point gap is the whole argument this piece wants to make: for most organizations right now, "which model should we use" is the wrong question. "Who signed off on this, and against what standard" is the one that actually predicts whether the project delivers.

I want to flag upfront how I found this, because the number is specific enough to deserve scrutiny before you build an argument on top of it. I traced it past a secondhand repost to Grant Thornton's own published piece, then cross-checked it against an independent legal-industry outlet reporting the same 950-person sample and the same 46%/31% figures – including a detail about the training-gap ranking that wasn't in the version I started from, which is a good sign the underlying report is real rather than a repost that mutated somewhere along the way. Two things I found reported alongside it but couldn't independently confirm – an industry-specific breakdown for insurers, and an exact survey field period – I'm leaving out entirely rather than repeating secondhand.

The gap that should actually worry you

Two more figures from the same survey sharpen the point. Seventy-eight percent of the leaders surveyed said they lack confidence their organization could pass an independent AI governance audit within 90 days. And 48% of boards had approved major AI investment without ever setting governance expectations for how that AI would be overseen – despite three-quarters of boards approving that spend in the first place. Read together, those three numbers describe an organization that's comfortable funding AI and deeply uncomfortable being asked to account for it.

Sit with the shape of that gap for a second, because it's a specific and slightly strange one. It's not that boards are unaware AI carries governance risk – 75% approving major spend suggests real conviction that this is worth funding. It's that the approval and the oversight got decoupled somewhere between the boardroom and the deployment. Someone said yes to the investment. Almost nobody, in nearly half of these organizations, said what "responsible" looks like once the money moved. That's not a knowledge gap. It's a sequencing failure – oversight was supposed to arrive after the funding, and for a lot of organizations, it just never showed up.

That's a very different failure mode than the one most AI coverage focuses on. The dominant narrative is still capability-centric – is the model good enough, hallucinating too much, missing context. Grant Thornton's data says that's not where the money is actually being lost. It's being lost in the gap between "we approved this" and "we could explain, on short notice, exactly how this is governed" – a gap that has nothing to do with which model sits underneath.

Why "whose sign-off" beats "which model" as the diagnostic question

If governance failure outranks technical failure as a cause of underperformance, then a lot of AI risk conversations are pointed at the wrong target. Asking "is this model accurate enough" is a reasonable question, but on this data it's not the question most likely to explain why a given AI initiative isn't delivering. The better diagnostic is closer to: who actually owns the decision to deploy this, what standard did they check it against, and could a stranger reconstruct that chain of accountability on demand. Most of the advisory-firm content I've seen on AI governance treats it as a checklist – a maturity model, a readiness framework, something you complete once and file away. This data argues governance isn't a compliance tax sitting alongside the real work. It's the leading indicator for whether the real work pays off at all.

That's a genuinely uncomfortable reframe for a lot of organizations, because it means the fix isn't a better model, a bigger budget, or more capable vendors. Those are the levers most executives are used to pulling when a project underperforms, and none of them touch the actual gap this survey identifies. Fixing a governance gap means someone has to be assigned ownership of a decision they may not have realized was unassigned, and that's a much less comfortable conversation to have in a leadership meeting than "let's evaluate another vendor."

That reframing matters most for regulated or risk-averse organizations, and for the advisors who serve them. A firm doing competitive intelligence, M&A diligence, or regulatory-adjacent advisory work is exactly the kind of client for whom "can you pass an audit in 90 days" isn't hypothetical – it's the actual question a regulator, an acquirer, or a board member is going to ask, on a timeline you don't control. Diligence timelines have already been stretching, not just deepening, and I'd bet a meaningful share of that stretch is exactly this: buyers and their counsel probing governance chains that don't hold up to a direct question, the same gap this survey is measuring at 78%.

This connects to a pattern I keep seeing

This is the same underlying failure I've written about from a couple of different angles. AI-washing has already become a legal liability, not just a credibility one – and a company that can't produce a clean governance trail for its AI use is exhibiting exactly the symptom regulators and litigators are learning to look for. Due diligence has had to sharpen specifically to catch AI-washing in acquisition targets, and the mechanism is the same one Grant Thornton's survey is describing from the inside: a board approves the spend, nobody sets the standard for overseeing it, and the gap sits there until someone with the authority to ask a hard question finally does.

What I'd actually check

If you're advising a client on AI adoption, or running the adoption yourself, the three questions this data suggests are worth asking before you ask anything about the model: who specifically owns sign-off on this deployment, what standard or policy did they check it against, and could you produce that chain of reasoning for an outside auditor with 90 days' notice. If the honest answer to any of those is a shrug, you're not looking at a technology risk. You're looking at the exact governance gap 46% of surveyed leaders already say is the real reason these projects stall – and at 78% odds, per this same data, that you couldn't prove otherwise on short notice either.