In 1996, a derivatives trader at Sumitomo Corporation lost $2.6 billion in unauthorized copper trades. The technology that enabled it was not particularly sophisticated — electronic trading terminals and a fax machine were enough. What made the loss possible was not the technology. It was the absence of oversight and the assumption that a profitable employee did not need watching.

Thirty years later, we are watching the same lesson unfold with artificial intelligence. Companies are handing AI tools to employees and measuring only the output. Revenue up, costs down, reports generated faster. Nobody is measuring what those tools are doing with the data, what shortcuts they are enabling, or what decisions are being made without human review. The metrics look excellent right up until the moment they do not.

Every major technology shift in business has followed the same arc: enthusiasm, delegation, crisis, regulation. AI is currently somewhere between delegation and crisis, and most executives do not realize which phase they are in.

The internet arrived in offices in the mid-1990s. Within five years, companies had email policies, acceptable use agreements, and firewalls. But those policies came after the damage — after the first employee emailed a client list to a competitor, after the first malware infection wiped a server, after the first lawsuit over workplace browsing habits. The technology moved fast. The governance moved after the ambulance.

Cloud computing followed the same trajectory. Companies migrated data to platforms they did not fully understand, attracted by cost savings and scalability. Then came the breaches. Equifax lost 147 million customer records in 2017 because a single server had an unpatched vulnerability. The cloud was not the problem. The assumption that someone else was managing security was the problem.

Mobile devices created the same cycle again. Bring-your-own-device policies emerged only after companies discovered that employees were accessing corporate email on phones with no encryption, no remote wipe capability, and apps that harvested contact lists.

AI compresses the timeline because it compresses the work. A task that took a human analyst three days can now be completed in ten minutes. That is genuinely useful — but it also means that errors, biases, and data leaks that would have taken three days to produce now arrive in ten minutes. The speed of output has increased. The speed of oversight has not.

Consider a financial analyst who uses AI to generate quarterly projections. Previously, the three-day process included natural checkpoints — conversations with colleagues, manual data verification, a review meeting before the numbers went to the board. Now the analyst can generate the same report in an afternoon, alone, with no checkpoints. The quality of the model might be excellent. The quality of the process has degraded.

The central problem is not that AI makes mistakes. Humans make mistakes too. The problem is that AI makes mistakes at scale and speed, and organizations have not built review processes that match that pace.

A human copywriter who produces two press releases a day will be reviewed by an editor who reads both. An AI system that produces forty press releases a day will overwhelm the same editor. The editor will start skimming. Then sampling. Then trusting. And the first fabricated statistic that reaches a journalist will arrive in the press release the editor did not read.

This is not a prediction. It is already happening.

The lesson from thirty years of technology adoption is straightforward: the tool is never the risk. The risk is the gap between how fast the tool operates and how fast the organization can verify its output. Close that gap and the tool is safe. Ignore it and the tool will eventually produce a loss that exceeds every efficiency gain it ever delivered.

Companies that learn this now will build verification processes alongside their AI deployments. Companies that learn it later will build them after the lawsuit.