Artificial intelligence has slipped quietly into the command centres of modern life. It no longer merely provides recommendations; it increasingly stands in the gap between citizens and the services they rely on. From evaluating mortgage applications and monitoring workplace productivity to determining eligibility for public services, AI is being woven into decision-making across global markets.
Yet as the technology becomes more assertive, our trust in its judgment is stagnating. The shift from generative AI – which summarises information – to agentic AI – which plans and acts independently – has sharpened this unease. While the promise of unparalleled efficiency is real, it comes with a drawback: our governance frameworks are static, but the AI systems they now oversee are not. In the competition between intelligent systems, trust is becoming the decisive advantage.
The trust paradox
According to a 2025 study by Harvard Business Review Analytic Services, 86% of corporate leaders plan to increase their investment in agentic AI over the next 24 months. There is a palpable fear among executives of being ‘left behind’ in the race for autonomous efficiency.
However, beneath this optimism lies a profound lack of confidence. Only 6% of those same executives trust AI agents to handle core business operations without human intervention. Public sentiment tells a similar story. The 2025 Edelman Trust Barometer uncovered that only 49% of the global population trusts AI as a technology, and only 44% feel comfortable with businesses using AI in their operations.
This ‘trust gap’ has significant consequences, as it means organisations are deploying powerful systems on shaky foundations. We are witnessing a world that is eager to use AI but worried about depending on it, which ultimately calls for stronger governance. Indeed, Gartner found that only 12% of organisations have a formal AI governance framework in place, while research from the Ada Lovelace Institute and Alan Turing Institute suggests that 58% of the public now favours the intervention of an independent regulator to ensure these systems are used safely.
When governance fails
The risks of unmanaged AI are already coming to the fore. For example, some banks have deployed AI-driven credit scoring models, seeking to accelerate loan approvals and reduce operational overhead. For a brief period, systems like these can deliver on all their promises.
However, as the systems scale and process increasingly diverse datasets, anomalies can begin to surface. Customers with strong credit profiles, many from specific regional or immigrant backgrounds, may be unexpectedly rejected as the model starts to misinterpret regional dialects, irregular (but stable) income patterns, and the use of subsidised housing as indicators of financial risk. Fortunately, GDPR Article 9 prohibits this type of behaviour, but does not protect those in other non-UK/EU jurisdictions.
Failures like these are not one of technical capability; the model is processing data exactly as it had ‘learned’. Instead, it is a failure of governance, as firms often rely on periodic, static audits rather than continuous fairness monitoring. But if there is no mechanism for ‘drift detection’ to alert the team, a model’s behaviour can shift away from its original parameters. Crucially, it requires a human-in-the-loop review for ‘edge cases’ so that systems don’t operate in a vacuum of accountability; another reason for models to be able to explain their reasoning.
The governance mosaic
In recent years, many regulators have developed their approaches to AI, although the resulting landscape remains a fragmented mosaic that poses its own risks. The EU has adopted a risk-tiered approach through the AI Act, while the UK favours a principles-led model that leverages existing regulators. In the US, reliance on executive orders and transparency testing creates a different set of obligations. For multinational organisations, this fragmentation makes compliance a moving target; Some have addressed it by delaying or withholding their AI tools from certain countries. For example, Google’s Gemini was unavailable in Europe from its launch in December 2023 until June 2024.
Inside the workplace, the challenge is cultural. Research from Harvard Business Review highlights that frontline employees frequently resist AI tools they do not understand or trust, regardless of the efficiency gains promised by management. This insight is reinforced by WalkMe’s 2025 research, which highlights a severe ‘adoption gap’: while AI tools represent 28% of enterprise applications, only 32% of employees know how to use their company’s AI applications effectively.
When users feel that a machine is making decisions for them rather than assisting them, adoption stalls. The result is a ‘shadow AI’ culture, where employees hide their use of unvetted tools to bypass oversight, leading to the accidental leakage of sensitive corporate data into public models.
A practical path to governance
To bridge the trust gap, organisations need to move from abstract ethical principles to practical, human-centric mechanisms. A credible governance strategy must be built on several foundational pillars.
First, and already implemented through GDPR in the EU, human checks should still be required for high-stakes decisions. That means that AI should not normally be the final arbiter in decisions that affect human lives or financial fates. Particularly in sectors like lending, healthcare, or public services, and as we pass through a period where AI starts to equal or even exceed human performance in certain areas, all AI outputs should be treated as recommendations that require verification by a trained professional.
Where bias is spotted, it should not be treated as a one-time flaw that can be ‘fixed’ immediately. It is a moving target that shifts as data and social contexts change. Organisations must implement real-time fairness dashboards that identify potential demographic disparities and trigger automatic alerts when drift occurs.
One example here is LinkedIn, which introduced the LinkedIn Fairness Toolkit (LiFT). This framework allows engineers to integrate automated fairness testing directly into their machine learning pipelines, enabling the system to measure and monitor demographic disparities across large-scale applications (e.g. recruitment systems) as a part of the standard scoring and training process.
This is particularly important in the age of agentic AI, where incident response cannot rely on human vigilance alone. Systems must be architected with ‘drift detection’ and automated error reporting that flags anomalies immediately.
Ultimately, AI is too consequential to be left solely to the IT department. Effective oversight requires a ‘responsible AI’ view of perspectives, bringing together legal, operational, technical, and user experience to design safeguards that work in the real world. This calls for multidisciplinary governance teams.
A case in point is Microsoft, often cited as the gold standard for “Responsible AI” infrastructure because they don’t just leave it to developers; they have a tiered governance structure. Microsoft implemented the Aether Committee (AI, Ethics, and Effects in Engineering and Research). This multidisciplinary group brings together leaders from legal, sociology, and engineering to vet high-risk AI deployments, ensuring that technical capabilities are balanced with ethical and legal safeguards.
Finally, traditional technical KPIs – like accuracy or speed – do not truly capture the health of an AI system. Organisations should track ‘trust metrics,’ including override rates (how often a human disagrees with the AI), appeal patterns, and employee sentiment surveys. These provide a more authentic picture of whether the system is being accepted or merely tolerated.
For example, Intuit, the financial software company, implemented a ‘Trust Dashboard’ for its AI products. Rather than just measuring accuracy, they track ‘Explainability’ and ‘Human-in-the-loop’ metrics, specifically monitoring how often expert accountants override AI-suggested tax categories to ensure the system remains a helpful tool rather than a source of frustration.
The competitive advantage
As we enter the next era of intelligence, the technologies that endure will not necessarily be the most sophisticated, but the most trusted. Organisations that invest early in robust, human-centric governance will earn legitimacy, the most valuable currency in an autonomous economy.
By aligning intelligent systems with human expectations and real-world working patterns, businesses can move beyond pilots and ‘shadow AI’ to create truly intelligent enterprises. Trust is no longer a nice-to-have; it is the engine behind AI adoption and the only sustainable path to long-term innovation.
