The AI adoption paradox:
Why trust, not capability, will define what happens next

The rise in AI adoption has been nothing short of extraordinary. ChatGPT reached one million users within a week of its launch in November 2022 and generative AI has scaled faster than any comparable technology.

Investment continues to accelerate. Orgvue’s research shows that 88% of companies plan to spend more on AI this year than in 2025 or 2024.

But outcomes tell a very different story.

MIT has reported that over 95% of corporate AI initiatives fail. Similarly, Orgvue found that 78% of corporate AI projects stall or fail, while Anthropic’s research shows that usage of AI in business domains remains far behind the capability of frontier models.

So, adoption is accelerating, investment is rising, and capabilities are improving. Yet impact remains limited. This is not just a contradiction; it’s a paradox. To understand why, we spoke with Paul Coleman, VP of Information Security and Data Protection, and Robert Rollings, Senior Principal Product Manager at Orgvue.

Part of the answer seems to be about trust. Across governance, regulation, and real-world testing, a consistent pattern emerges. The issue is not AI capability; it’s whether organizations can trust it when decisions matter.

You may trust a generative AI model to create a meme or draft an email. But do you trust it to respect your intellectual property and data? Do you trust it to provide the right answers to questions that shape your business and the future of your employees? In most organizations, the answer is no.

That trust gap becomes most visible in transformation. And that’s not surprising. Transforming a business is one of the most complex and high-stakes challenges organizations face. Orgvue’s research found that two in five CEOs would rather resign than lead another transformation.

If AI can’t be trusted in that context, then can it be trusted at all? To understand why AI struggles to deliver outcomes, it helps to break the trust gap into two parts: governance and credibility.

The governance gap

Paul Coleman is clear: “The next battleground in workforce software is not AI capability; it’s whether AI is embedded inside a governed decision environment.”

In workforce decisions, AI should support judgment, evidence, and workflow. It should not become the decision-maker. Because the closer that software gets to decisions about people, the more important oversight, traceability, accountability, and transparency become.

As Paul puts it, the real question is not whether AI is intelligent, but whether it is governed. And this is not just a hypothesis. An OECD report shows that software is already being used to instruct, monitor, and evaluate workers. Managers highlight concerns around questionable accountability, lack of explainability, and inadequate protection of workers’ wellbeing.

An AI-only approach may be acceptable for drafting or summarizing text in emails and reports, but it’s not a reliable model for workforce decisions.

Once software starts influencing promotion, termination, task allocation, monitoring, or evaluation, the benchmark changes. It’s no longer about convenience; it’s about governance.

Governance is not an add-on; it’s the product

Regulation is moving in the same direction as the governance debate. As an example, the EU AI Act identifies certain employment and workforce-management uses of AI as higher-risk scenarios, such as recruitment and systems used in decisions affecting promotion and termination.

In practical terms, this moves the conversation away from AI capability and toward questions of governance, oversight, transparency, documentation, accuracy, and control.

The closer AI gets to decisions about people, the more organizations need evidence of governance, auditability, human oversight, and control over data and change.

The regulatory question is no longer whether organizations use AI; it’s whether their systems are documented, interpretable, overseen by humans, and governed.

But this creates a divide. AI-only workflows are fast, flexible, and language-driven. But when they influence workforce decisions, they are hard to evidence, reproduce, or defend. AI struggles with clear ownership of decisions, consistent governance, transparent assumptions, and auditable change control.

By contrast, governed AI operates inside a structured environment. As Paul puts it: “In this category, governance is not a compliance wrapper around the product. Governance is part of the product.”

This idea also reframes the concept of “human in the loop.” According to an ICO paper, the key question is whether a person has real influence over AI decisions, with the authority and competence to change them. If not, there’s no meaningful human involvement.

Trust, in this context, depends on systems where human oversight is actively engaged, not just a rubber stamp.

Governed AI is different from AI-only

This distinction becomes clearer when you look at how different approaches operate in practice. An AI-only model may be fast, flexible, language-driven, but it’s also difficult to evidence, reproduce, govern, or defend.

A governed model takes a different approach. Take Orgvue for example.

Inside Orgvue, AI operates in a controlled environment of structured data, governance rules, human review, and auditability. Changes to organizational data are not saved without human review, and outputs are trackable and reversible.

This is supported by formal governance structures, including an AI management system, risk management oversight, and certifications such as SOC 2 Type 2, ISO 27001, ISO 27018, and CSA STAR, with ISO 42001 certification in progress.

This does not eliminate bias or guarantee fairness, but it does make fairness more governable.

Decisions can be structured, reviewed, logged, audited, and tested against agreed rules and analytical methods. Accuracy is not just about model performance, it’s about whether outputs are generated from controlled data, within a known framework, and whether they can be challenged.

The credibility gap

Yet even with governance in place, another issue emerges, and that’s credibility. Robert Rollings has been testing how leading AI models perform on real transformation tasks, such as organizational baseline analysis, and the findings are clear. Frontier models can produce accurate, repeatable, and consistent answers, but only under very specific conditions:

1. A skill, instructions, or equivalent that provides a governed framework, including a data dictionary and explicit assumptions
2. Structured prompting to control scope, expectations, and outputs
3. Clearly defined scope within the task

When these conditions were removed, performance broke down.

A simple approach, such as uploading a dataset and asking for analysis, produced inaccurate, inconsistent, and incomplete outputs. Results varied depending on how prompts were phrased. Assumptions were often missing or misinterpreted.

But when a governed framework was introduced in the form of a well-defined and scoped “skill,” performance improved significantly. The implication of this is clear:

AI capability alone does not create reliable outcomes. Without structure and governance, analysis is not decision ready.

This also reinforces a broader point. General-purpose tools like ChatGPT, Claude, or Copilot will not replace specialist platforms in areas that require precision, consistency, and defensibility. Reliability comes from how AI is applied, not just from the model itself.

AI can accelerate analysis, but it does not guarantee decision readiness

Across both perspectives, a consistent pattern emerges. AI can analyze data at speed. It can surface patterns and generate options. But without governed assumptions, structured context, and human oversight, those outputs don’t hold up to scrutiny when decisions need to be made.

This is where many organizations are getting stuck. They’re investing in AI capability but not building the conditions required to use it effectively. The result is a gap between insight and action. And in transformation, that gap matters.

These are decisions that affect jobs, cost millions, and shape the future of organizations. They require consistency, traceability, and accountability. AI can support those decisions, but it can’t replace the systems needed to govern them.

Trust is the real constraint on AI adoption

The AI adoption paradox becomes clearer through this lens. Organizations are not failing because AI doesn’t work. They’re failing because they can’t trust it in the moments that matter. Governance addresses whether AI can be used safely and responsibly, while credibility addresses whether organizations can rely on its outputs.

Without both, adoption will continue to stall at the point where decisions need to be made. And that’s where the real value is.

Organizational design

Use Orgvue to restructure your organization.