AI is finding its way into the ordinary rhythm of work. Employees use it to draft reports, analyse information, respond to customers and make sense of large volumes of data. In some organisations, AI systems are also beginning to take action, moving beyond recommendations to interact with applications and execute tasks.
For enterprises, trust goes beyond receiving a convincing response. If that response is inaccurate, biased or draws on information it should never have accessed, it can create serious risks.
Adoption is moving faster than readiness
The pace of adoption makes organisational readiness worth examining. Deloitte’s State of AI in the Enterprise 2026 report finds that Indian enterprises are moving beyond experimentation, with nearly 40 percent reporting significant or full use of AI, compared with a global average of 28 percent.
According to industry research, many organisations are yet to successfully scale AI across multiple business units or adopt an AI-first approach, despite plans to increase AI spending in 2026.
These findings show that enthusiasm for AI is moving faster than many organisations’ ability to use it reliably in day-to-day operations. The challenges often begin with the basics. Data may be incomplete, outdated or unsuitable for the task. Employees may be comfortable using AI without fully understanding where it can go wrong. They may not know when to trust an output or when to check it more closely. Processes also need to make clear how outputs are reviewed and what happens when errors occur. Security teams should be able to see what data the system uses, which systems it connects to and what actions it can take.
Readiness comes down to four areas working well together, which include data, people, processes and security. Data must be accurate, relevant and used responsibly. People need the right knowledge and judgement. Processes should account for mistakes and unexpected situations. Security controls must prevent unauthorised access, protect sensitive data and flag unintended actions. If one area falls short, the reliability of the whole system can suffer.
Turning Trustworthy AI into everyday practice
Trustworthy AI principles need to be supported by practical controls throughout the AI lifecycle. Organisations can build these controls into their existing delivery and risk management processes, starting at the design stage and continuing after deployment.
During design, teams should define the system’s purpose, boundaries, data access requirements and the possible consequences of failure. Privacy, bias, security and regulatory risks should be assessed early, with a business owner accountable for its use and outcomes.
During development, access to data, system instructions, APIs and connected tools should be restricted. Secure development practices, data minimisation and appropriate logging should be applied. AI agents should only have the permissions needed to perform their role.
The system needs to be tested before it goes live. The tests should check whether it gives incorrect or unfair results, exposes sensitive information or can be manipulated to ignore its safeguards. They should also reflect situations that may arise in daily work. Any limitations should be clearly recorded, including when a person needs to step in.
Review should not stop after launch. Important outputs need to be checked, and high-impact actions should not proceed without human approval. Users should have an easy way to report issues. Any remaining risks must be accepted by the person responsible before the system is rolled out more widely.
Human oversight must also be meaningful. A person asked to approve an AI-generated recommendation needs sufficient context, authority and time to challenge it, rather than simply endorsing the system’s output.
Deloitte Trustworthy AI Framework
Adopting a Trustworthy AI (TWAI) framework built around nine core attributes, namely Fair and Impartial, Robust and Reliable, Transparent and Explainable, Responsible and Accountable, Respectful of Privacy, Safe and Secure, Professional and Competent, Environmentally and Socially Responsible, and Inclusive and Accessible, enables organisations to operationalise responsible AI at scale.
Collectively, these attributes provide a common language for business, technology, risk, legal, and compliance teams, enabling consistent decision-making, stronger governance, reduced risk exposure, and greater trust among customers, employees, regulators and other stakeholders. By following a clear TWAI framework, organisations can achieve the Apex Advantage, enabling them to innovate with confidence while keeping their systems ethical, reliable, secure and aligned with their goals and values.
Trust needs attention after deployment
Passing pre-deployment testing provides a useful starting point, but it does not guarantee that the system will remain safe over time. AI behaviour may change as models are updated, data shifts, users adopt new ways of working or additional tools are connected. Regular monitoring is therefore essential.
Organisations should monitor four areas. The first is output and data behaviour, including accuracy, hallucinations, bias, data quality and changes in performance. The second is security and access, covering prompt injection, exposure of sensitive data, unusual access patterns, permission changes and unauthorised tool use. For AI agents, records should show what actions were taken and the identities or permissions used.
The third area is human intervention and business impact. This involves reviewing complaints, overruled recommendations, repeated corrections and decisions that lead to poor outcomes, as these may point to a wider problem. The fourth is changes to the system. Significant updates to models, prompts, data sources, integrations or permissions should be reviewed and tested before they are introduced.
Trustworthy AI is ultimately a continuing operational responsibility. The objective is to show that the system remains within acceptable boundaries over time. Any change should be detected and addressed before it affects the wider business or its customers. Regular oversight helps protect the business and its customers, makes accountability clear and builds confidence in how AI is used. Trust is maintained through responsible management at every stage of the system’s lifecycle.
Authored by Ashish Sharma, Partner, Deloitte India