The 3 steps to accurate and trustworthy enterprise AI

It’s been over two years since OpenAI released ChatGPT, kicking off the ultimate generative artificial intelligence (GenAI) hype cycle. But what do enterprises really think about the potential of GenAI? And how are they using it today?

According to a recent survey of 1,100 technical executives and technologists by Economist Impact, 73% of businesses believe GenAI is critical to their long-term strategic goals.

In fact, 85% are already using GenAI in at least one function. Despite this enthusiasm, accuracy continues to be a significant challenge. Just 37% of executives believe their GenAI projects are ready for production.

To turn hype into actual results, businesses must address what’s undermining GenAI adoption and trust: accuracy, risk and access. With specialized systems built on proprietary data, coupled with a robust governance strategy, organizations can solve AI’s performance problems and begin using the technology cost-effectively to drive overall business growth.

Here are three steps every leader should take for more accurate and trustworthy enterprise AI in 2025.

Whether a bank is trying to detect fraud or a pharmaceutical company is attempting to accelerate drug discovery, each enterprise workload has unique demands that can’t be addressed through a single, commercial large language model. Instead, organizations are quickly learning that better data leads to smarter AI.

A company’s proprietary data holds vital information about customer, employee, partner, and supplier interactions. These corporate assets can train models that understand an organization’s unique culture, jargon, and internal processes. These systems provide the deeper intelligence and advanced automation employees need to do their jobs better, faster, and smarter.

Using multiple AI models and systems then helps improve overall performance. This type of architecture can create an enterprise AI agent system that lets users tackle more complex and dynamic inquiries than traditional chatbots and even take actions on behalf of employees. For example, Mastercard recently launched an AI-powered agent to simplify customer onboarding.

The system automates routine tasks, answers critical customer questions and continuously improves through expert feedback. Trained on Mastercard’s proprietary data, it ensures accurate, up-to-date and Mastercard-specific responses. Pairing agents with human-in-the-loop design in this way helps organizations avoid risk from “bad AI.”

As data is used more extensively across the organization, enterprises need to provide and control access to sensitive information. Effective unified governance gives enterprises the confidence to selectively broaden access to data while maintaining responsible AI tools that facilitate improved organizational efficiency and growth.

Data governance extends beyond establishing the proper permissions. With the right controls, governance can help organizations enhance trust in AI systems. Establishing standards around the data, policies on how the assets can be used and controls over the outputs all help improve the overall quality and reliability of the data.

For example, with the right governance protocols, Chevy might have prevented its AI agent from recommending the rival Ford F-150 to potential customers.

An AI system that works well in a test environment may not perform the same when used by thousands of employees or millions of consumers. While companies may be able to keep costs low while running the technology in a controlled setting, for example, that becomes much more challenging when the whole workforce or customer base has access to it.

As a result, many AI projects remain stuck in the experimental stage, where they’re likely producing little to no business value. When AI is effectively deployed at scale, it can help the organization enhance its current business strategy and more quickly pursue new market opportunities to drive overall growth.

For example, Grainger built a custom AI system to help the service teams navigate and manage the industrial manufacturing equipment distributor’s over 2.5 million products. Now, support agents can use natural language queries to quickly surface critical information, helping to reduce service times and improve customer satisfaction.

Importantly, the system can instantly process the nearly 400,000 product changes daily, ensuring users always get the most accurate, up-to-date information.

To implement these use cases in the real world, companies must be able to track how the systems are performing across the larger user base. Data governance can now help organizations monitor model usage to control costs.

Meanwhile, continual training ensures that GenAI systems' performance doesn't degrade over time. Importantly, employee feedback loops can help improve the systems' overall accuracy.

While every enterprise will ultimately have its own unique adoption journey, these steps are fundamental to building and scaling enterprise AI systems that are accurate, cost-effective, and actually help employees and customers. This is how organizations build trust in transformational technology.