Is AI in the enterprise ready for primetime? Not yet.

Although bullish on the prospects for AI to automate many work activities, McKinsey acknowledges it’ll take several decades for this to happen at any scale. CIOs and other executive leaders should keep this in mind amid the hype and wild claims made by many vendors and consultants.

There are a number of reasons why meaningful AI deployments within the enterprise will take longer than many imagine.

Complexity of human work

It’s been estimated that the average person makes 2,000 decisions every hour. While many of these decisions are routine and require little thought, others are far more complex and nuanced. At work, we’re efficient at processing multiple inputs in rapid time to take into account issues of safety, social norms, the needs of our colleagues and employer, as well as accuracy and strategic goals. At the same time, we can communicate these decisions orally, in writing, and through gestures using multiple systems and workflows.

While computing technologies and enhanced access to data may have helped businesses make better routine, low-value decisions, anything more complex still requires human input and oversight. An organization’s reputation lives or dies by the decisions made within it and once lost, is difficult, and often impossible, to regain. While chatbots will take over many functions currently performed by human-powered call centers, these will operate within tightly defined parameters including their data inputs and the answers they can give.

AI hallucinations

The problem of AI hallucinations, where a large language model (LLM) presents authentic-looking but made-up results, shouldn’t be underestimated for enterprise AI deployments. It’s been estimated the hallucination rate for ChatGPT is between 15% and 20%, an unacceptable figure for business-critical decision making.

Hallucinations can be reduced within enterprise deployments by fine-tuning LLMs through training them on private data that’s been verified. Further improvements can be made by restricting queries to proven prompts as well as incorporating open source tools such as Langkit and Guardrails, or proprietary products like Galileo. These tools and frameworks are still in the early stages of development, and users will need to experiment with multiple approaches and solutions. It’ll be several years at least before established and trusted methods for reducing hallucinations to acceptable levels are widely available.