Why AI can only ever be as trustworthy as the data behind it

Two years ago, most conversations about generative AI centred on what the technology could do. Organisations were experimenting with large language models (LLMs), building proofs of concept, and marvelling at the speed with which GenAI could generate text, summarise reports, or answer complex questions.

 

Today, the question has shifted from what AI can do to whether we can trust it—and, more importantly, how businesses can make AI worthy of that trust. Because however powerful and flexible an LLM may be, it does not operate at the same level of predictability as traditional ERP platforms. For businesses, that introduces a very different challenge. Here’s why.

 

Trust begins with information

Most enterprise AI failures won’t occur because a language model lacks intelligence. They’ll happen because the information feeding that model is incomplete, inaccurate, poorly governed, out of date, or impossible to verify.

 

AI can reason over enterprise information at remarkable speed, but it cannot compensate for weak information management. In fact, left to its own devices, it can do something far worse, amplifying the existing weaknesses in our data and assumptions.

 

We used to talk—perhaps less so now—about the ethics of AI. But as enterprises move beyond isolated AI pilots and toward enterprise-wide adoption, trust is becoming a technical architectural challenge, not merely an ethical aspiration. That challenge will only grow as employees move away from navigating document repositories or searching through shared drives, and increasingly expect to ask questions in natural language and receive authoritative answers.

 

For sure, this is an evolution of how organisations work with business information that promises enormous productivity gains, but those gains will only come only if those answers are grounded in information that can be trusted.

 

The problem predates AI. Enterprise information is rarely as organised as we’d like to assume. Documents are duplicated across repositories, policies become outdated, records lack consistent metadata, and sensitive information is stored alongside public content. Governance varies across departments, while critical knowledge often exists in formats that even today’s systems, let alone tomorrow’s, struggle to interpret consistently.

 

These are not new information management challenges. So, what can be done to move us forward?

 

The Seven ‘Guardians’ of Trust

We’ve seen customers coalesce around seven approaches or ‘guardians’ that help fledgling AIs deliver reliable business outcomes.

 

The first is content quality. Unless you establish clear controls over the quality of the content entering the AI’s knowledge ecosystem, AI cannot reliably distinguish between accurate information and outdated documentation. Poor-quality information inevitably produces poor-quality answers.

 

The second is governance. Enterprise AI requires clear rules defining who owns information, who can modify it, and how authoritative sources are maintained. Strong governance provides the structure that prevents AI from concluding answers that might have big financial or strategic ramifications from sources that should never have been considered reliable.

 

Third is metadata and context. Business documents rarely exist in isolation; their value depends on understanding when they were created, who approved them, which business process they support, and how they relate to other information. The right metadata allows AI to interpret enterprise knowledge with significantly greater accuracy while helping users understand why a particular answer was generated.

 

The fourth guardian is security. As organisations embed AI into business processes, access controls become even more important. AI should respect the same permissions as the people using it, ensuring sensitive financial, legal, or customer information is only surfaced to authorised employees. Trust leaches away quickly when AI reveals information it should never have accessed.

 

The fifth guardian is compliance. Regulations governing privacy, records retention, and industry-specific requirements continue to evolve. AI must always operate within those frameworks and never be permitted to bypass them. The CIO needs confidence that automated responses remain consistent with the legal and regulatory obligations of every market in which they operate.

 

The sixth guardian is provenance. Increasingly, employees don’t simply want an answer, they want to understand where it came from. Can the AI cite the original policy? Was the information approved? Has it been superseded? By allowing users to trace responses back to authoritative sources, provenance transforms AI from a black box into an accountable decision-support system.

 

Finally comes lifecycle management. Information changes constantly; contracts come up for renewal, procedures evolve, and policies are replaced. AI systems that continue relying on obsolete content gradually become less trustworthy over time. Effective lifecycle management ensures knowledge remains current throughout its useful life while removing information that should no longer influence business decisions.

 

Taken together, these seven disciplines create the conditions for enterprise AI to earn, and critically maintain, organisational trust. AI doesn’t replace the solid information management disciplines of yesterday; it ends up dependent on them. The more organisations rely on AI for decision-making, the more valuable mature practices around content management, records, taxonomy, governance and security become. AI ultimately increases the return on those investments because they determine the quality of information it can access.

 

In many organisations, enterprise content management has historically been viewed as operational infrastructure—important, but largely invisible. AI changes that. In the new enterprise architecture of trust, ECM is no longer simply a place where documents are stored, but the knowledge foundation from which intelligent systems retrieve evidence, establish context, and generate business recommendations.

 

AI cannot create trust on its own

Successful AI strategies should begin not with model selection but with an honest assessment of enterprise information readiness. Before expanding AI across the organisation, leaders need to understand whether information is governed consistently, whether authoritative content can be identified, permissions are reliably enforced, and whether AI-generated answers can be traced back to trusted sources.

 

Organisations deploying AI in customer service, legal operations, financial management, healthcare, or government cannot afford systems that produce answers without accountability. As AI becomes embedded within core business processes, explainability and trust matter as much as speed and productivity.

 

The good news is that many organisations already possess much of the infrastructure needed to support trustworthy AI. Enterprise content management, records management, workflow automation, metadata management, retention policies, and security controls have long addressed many of the information challenges AI now brings into sharper focus. The opportunity lies in connecting those proven capabilities to modern AI platforms rather than treating them as separate initiatives.

 

Ultimately, AI cannot create trust on its own—it can only inherit and curate it. For CIOs, that may be the most important lesson of all: the future of enterprise AI will be shaped not simply by advances in AI, but by the strength of the information architecture supporting it.

 

---

 

Dr John Bates is CEO of European-headquartered Doxis, the document intelligence company

 

Main image courtesy of iStockPhoto.com and showcake