Enterprise AI in action

The case for responsible AI

How to build responsible AI as an organisational capability, in conversation with experts from the United Nations International Computing Centre and the Frugal AI Hub at Cambridge University

Enterprise AI in action Image collage Pillar 3 hero UNICC hero

Every large business depends on delegation. Shareholders rely on boards, boards on executives, executives on managers. But delegation brings risk. Those who act on a firm’s behalf do not always act in its interests. Economists call it the principal-agent problem. To align the behaviour of agents with the goals of principals, firms provide incentives and build systems of control, from audits and oversight committees to the threat of dismissal.

As firms delegate more and more tasks to artificial intelligence (AI), they face a renewed principal-agent problem. AI systems are unlikely to embezzle or poach clients like an unfaithful employee might, but they can hallucinate, be biased, leak data and reach conclusions that even their makers cannot fully explain. Just as firms learned to manage human agency, they now need governance frameworks to keep enterprise AI systems in line with their corporate goals and social obligations. 

Most firms are learning this the messy way. A global survey of nearly 2,000 firms conducted in 2025 by McKinsey, a consultancy, finds that about half of those deploying AI have seen at least one negative consequence from using the technology, ranging from inaccuracy to breaches of cybersecurity or intellectual-property theft. The survey also finds that more firms try to curb AI risks in 2025 than in 2022, but their approach remains lopsided. Most say they worry about the (in)ability to understand how AI reaches its conclusions, yet “explainability”, as the issue is known in AI speak, is not among the risks firms say they are trying to mitigate.

In conversation with Economist Enterprise, Anusha Dandapani, chief AI officer at the United Nations International Computing Centre (UNICC), and Elizabeth Osta, co-founder of the Frugal AI Hub at the Judge Business School of Cambridge University, argue that responsible AI should not be overlooked. It is a business capability that can enable AI adoption, reduce its risk and waste, and make the technology fit for purpose. There are a few things firms can do to make their AI use responsible.

Make responsibility everyone’s job

An easy way to fail at responsible AI is to govern it like regular enterprise software. A policy written by legal, a model reviewed by IT and training videos watched at double speed will not suffice. Using AI responsibly, Ms Osta suggests, “is a muscle that an organisation needs to build”. Board endorsement and executive commitment are essential, she says, but responsibility must be “pervasive across the organisation”.

A collage illustration of graphics with male profile and paper and pen

To grasp “organisational capacity” in practice, look at how firms learned to manage other complex risks. Financial crime and workplace safety, for instance, did not yield to thicker rulebooks. What helped were measures like training people to notice concerns, making escalation easy rather than career-limiting and weaving checks into everyday work. Similarly, AI governance requires a shared understanding of risk that shapes routine decisions. This follows from AI’s general-purpose character; embedded across functions and tasks, it cannot be fully governed through specific points of control.

The need to build organisational capacity for responsible AI use becomes more pressing as more and more people use the technology. A survey from 2025 by EY, a consultancy, finds that two-thirds of firms allow “citizen developers” to build or deploy AI agents (systems that can perform tasks with minimal or no human intervention), but only 60% provide organisation-wide frameworks to guide their use, and half admit to limited visibility into how employees deploy them. That gap matters. As far as it exists, delegating to AI is not empowering but negligent.

“Responsible AI is a muscle that an organisation needs to build. It needs to be endorsed by the board and leadership, but pervasive across the organisation.”

Elizabeth Osta

Co-founder, Frugal AI Hub at Cambridge University

Co-founderFrugal AI Hub at Cambridge University

Build trust through deeds, not words

Few notions can be more valuable to firms, or get invoked more carelessly, than trust. If firms build genuine trust around AI, people are more likely to adopt the technology meaningfully, supercharging innovation and improving decision-making (read more in this series). But if trust rests on airy value statements, expect employees to treat enterprise AI warily, even as they use unlicensed tools in the shadows. A survey by Ipsos, a pollster, and analysed by the Tony Blair Institute, a think tank, finds that 38% of British respondents cite a lack of trust in content generated by AI as a barrier to using the technology. 

Building genuine trust around AI includes a few essential steps. Ms Dandapani emphasises the one that tends to concentrate corporate minds fastest: “the protection of data”, which, in AI, looks different from traditional security. The main worry is often not theft by outsiders, but leakage: when staff paste, upload or route sensitive information into AI tools that may store it, share it or use it in ways the firm cannot fully see or control. An analysis by Harmonic, a cybersecurity firm, quantifies the risk by reviewing 1m prompts and 20,000 files used with over 300 AI-enabled applications. It finds that nearly 22% of files and 4.4% of prompts contained sensitive content, and that the average enterprise saw 23 previously unknown and unlicensed AI tools newly used in a single quarter of 2025.

A collage illustration of graphic hands and globe

Bias is another harm from bad data practices with AI. When training data are uneven, AI systems tend to be, too. Algorithms start producing errors that make them less useful or, when people are involved, fall more heavily on some groups than others. Biased AI can erode legitimacy in the technology in a way that is hard to repair even if the model is patched.

What restores confidence in both privacy and fairness is not lofty language but visible discipline through good practices. This includes putting boundaries around sensitive data, approving tools once they can be monitored and building mechanisms to audit AI outputs when someone challenges them. These are operational choices, but they also carry moral weight that can be the difference between AI that is trusted and one that is set aside. They signal that a firm expects mistakes and has planned for them, so employees feel confident to use AI tools.

Two other ingredients of trust in AI are reliability and intelligibility, and they, too, depend on clarity. Ms Dandapani’s suggested approach is direct: systems should be “understandable” and “explainable for their operations”, with transparency going “hand in hand with explainability”. A firm cannot sensibly ask workers to rely on tools they cannot question, especially at a moment when, according to the Work Trend Index by Microsoft, a technology firm, 52% of employees and 57% of leaders say job security is no longer a given in their industry.

“Responsible AI is not just about rules; it’s about mitigating risks like bias and privacy issues, as well as building trust so AI is fair, reliable and friendly for adoption.”

Anusha Dandapani

Chief data and AI services officer, United Nations International Computing Centre

Chief data and AI services officerUnited Nations International Computing Centre

Do not automate away accountability

The use of AI agents adds a twist to the principal-agent problem. Employees become principals to algorithms while remaining agents of senior management. Delegating to AI agents is a good way of getting more done but can be a messy way to multiply risk. Unless firms make human-AI relationships explicit—determining who may delegate what, under which conditions, on what evidence and with what scope for challenge—they run a familiar management risk.

One common response is to insert a “human in the loop”, as corporates say. It is a good strategy, but firms should be careful what sort of loop they build. In high-stakes settings, Ms Dandapani argues, AI should advise rather than decide, and humans should retain responsibility. Some organisations now ask staff to “sign off” on AI outputs. But if a model’s inputs are unclear, its limits unstated and its reasoning opaque, a human signature should not provide much comfort. The rule: do not ask staff to own what they cannot inspect.

A collage illustration of graphics

Real accountability has two parts. First, clear boundaries. Firms must decide which decisions must remain human-led, which can be automated and which require human review, supported with meaningful evidence. Second, legibility. Ms Dandapani emphasises that AI systems need to be “understandable” and “explainable for their operations” primarily as a control mechanism. Ms Osta adds the operational corollary: because models drift and providers update them, firms should monitor performance before, during and after launch.

The task has grown harder as AI models have become more sophisticated. In the past, firms trained simpler machine-learning models on their own data, making outputs easier to trace. Now they use large language models trained by hyperscalers on billions of parameters. “Models tended to be more stable,” Ms Osta says, but today “they are drifting,” and when providers upgrade, “everything changes”. 

Responsible AI in action

As companies hand more decisions to AI, using the technology in a responsible way matters as much as firms’ ambition. Business leaders can avoid AI risks by spreading responsibility across the organisation, grounding trust in clear safeguards, and keeping humans firmly in charge. These principles underpin the third pillar of our Enterprise AI in action series. The instalment also features an article on how E.on is hardwiring strict limits on its AI systems and another article on how DHL is making the technology work at scale through careful risk assessments and constant feedback."

Sources

  1. McKinsey & Company (QuantumBlack, AI by McKinsey). The state of AI in 2025: Agents, innovation, and transformation. November 2025. Available at: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

  2. EY (Ernst & Young). EY survey: companies advancing responsible AI governance linked to better business outcomes. October 2025. Available at: https://www.ey.com/en_gl/newsroom/2025/10/ey-survey-companies-advancing-responsible-ai-governance-linked-to-better-business-outcomes

  3. Tony Blair Institute for Global Change. What the UK Thinks About AI: Building Public Trust to Accelerate Adoption. September 2025. Available at: https://institute.global/insights/tech-and-digitalisation/what-the-uk-thinks-about-ai-building-public-trust-to-accelerate-adoption

  4. Harmonic Security. GenAI Data Exposure: What GenAI Usage Is Really Costing Enterprises. July 2025. Available at: https://www.harmonic.security/resources/genai-data-exposure-report-fa6wt

  5. Microsoft WorkLab (Microsoft). Work Trend Index — Annual Report: “2025: The year the Frontier Firm is born”. April 2025. Available at: https://www.microsoft.com/en-us/worklab/work-trend-index

More from the series