The Judgment Gap

Nearly every financial firm now runs on AI. Almost none can prove it is paying off. The distance between those two facts is where the next advantage will be won — and it is a human distance, not a technical one.

Read the headline surveys and finance has already crossed into the AI era. NVIDIA's 2026 State of AI in Financial Services reports that 65% of firms are actively using AI, up from 45% the year before; 42% are using or assessing autonomous "agentic" systems; and 89% say AI is both lifting revenue and cutting costs. A separate NVIDIA-cited poll has 89% of finance executives crediting AI with higher revenues and 73% calling it crucial to their future. If adoption were the scoreboard, the game would be over.

It isn't, and the more sober data explains why. A Cambridge Centre for Alternative Finance survey of 628 firms and regulators across 151 jurisdictions — one of the largest of its kind — found that while four in five firms now deploy AI, the use is overwhelmingly back-office: software engineering, data management, document processing. Crucially, 76% of large financial institutions and 55% of all industry respondents say they struggle to measure the value of what they have deployed. Only 40% report any profit boost; 43% report no change at all. Parallel surveys from McKinsey, the IIF, EY and others echo the same gap between rhetoric and result.

The Cambridge authors call this a "deep execution gap" between experimentation and institution-wide integration. We would put it more bluntly: most firms have bought the tool and not yet built the judgment to use it.

Machines do not deploy themselves

The most instructive item in the whole stack is an anecdote, not a statistic. Writing in the Financial Times in May 2026, Gillian Tett relayed a New York financier's verdict on his first cohort of genuinely "AI-native" interns: dazzling on the surface, "alarmingly shallow" once their thinking was probed. The firm's response was telling — fewer return offers, and a deliberate tilt toward hiring humanities graduates over pure STEM. "We want critical thinking, not just AI."

That instinct is the thesis. AI is becoming a commodity input: the same Cambridge survey shows most firms build on a handful of external models — OpenAI first, then Google and Anthropic — rather than training their own. When everyone draws intelligence from the same few taps, owning the tool stops being a differentiator. What separates winners is the human layer wrapped around it: the discernment to choose where AI is actually decisive, the discipline to measure honestly rather than cheerlead, and the strategy to integrate it instead of bolting it on. Intelligent machines, as Tett puts it, do not deploy themselves for good or ill. Human strategy is the scarce input.

The same scarcity governs the downside

This matters more because the risks scale with the hype. The Financial Stability Board warned in May 2026 that the roughly $2tn private-credit market — some of it now financing the very AI data centres driving the boom — is opaque, increasingly leveraged at five-to-seven times earnings, newly crowded with retail money, and "untested" in a real downturn; the warning landed a day after HSBC disclosed a $400mn hit on a single private-credit exposure. The IMF, writing the same week, argued that AI is now actively fuelling cyberattacks, and that resilience, supervision and international coordination are the only adequate response.

The technology itself is dual-edged in a way that demands judgment rather than enthusiasm. Federal Reserve Vice-Chair for Supervision Michelle Bowman, in May 2026 remarks, described a frontier model capable of rapidly detecting cyber vulnerabilities — useful for defenders patching their own systems, dangerous in the hands of attackers — serious enough that the Treasury Secretary and Fed Chair convened the largest banks to discuss it. Her supervisory conclusion was not to clamp down but to stay adaptable, even narrowing model-risk guidance so it no longer misapplies to generative and agentic AI.

Two details sharpen the point. The Cambridge data shows AI vendors consistently rank adversarial-AI and resilience risks lower than either industry or regulators do — the people selling the tool are the least alarmed by it. And regulators' own AI adoption runs at roughly half the level of the firms they supervise. In other words, the judgment required to govern AI is unevenly distributed, and thinnest precisely where it is sold and where it is policed.

What BridgeUp is doing about it

AI amplifies whatever operating logic an institution already has. Coherent firms compound their advantage; disorganised ones automate their disorder faster. The execution gap between AI experimentation and AI value is not waiting on a better model — it is waiting on better human and governance scaffolding.

That scaffolding is the work. BridgeUp helps institutions cross the judgment gap: moving from adoption metrics to evidenced value, building the measurement and governance that tells you whether a deployment is actually working, and keeping the critical-thinking layer — in hiring, in oversight, in strategy — that no foundation model supplies. The firms that win the AI decade will not be the ones that bought the most intelligence. They will be the ones that brought the most judgment to it.

Sources: NVIDIA, "State of AI in Financial Services: 2026 Trends"; Cambridge Centre for Alternative Finance, "2026 Global AI in Financial Services Report"; Gillian Tett, Financial Times (May 2026); Financial Stability Board private-credit report (May 2026); Michelle W. Bowman, Federal Reserve remarks (May 2026); International Monetary Fund, "Financial Stability Risks Mount as Artificial Intelligence Fuels Cyberattacks" (May 2026).