The AI agent promise vs reality
Almost every large business is trying AI agents right now. Cisco’s data shows 85% of enterprises are running pilot programmes. But here is the number that matters more: only 5% have actually shipped these agents to production.
At the VB Transform 2026 conference in San Francisco, Bryan Silverthorn — Amazon’s Director of AGI Autonomy — explained why that gap exists. His answer was not what most people expect. The problem is not that AI agents are not capable enough. The problem is that they are not reliable enough to trust in real business operations.
Silverthorn, who joined Amazon through its acquisition of Adept AI, broke reliability down into four dimensions: consistency, robustness, predictability, and safety. His point was that most teams only measure one or two of these, and they miss the ones that cause failures in production.
Why AI agents pass tests but fail in the real world
Silverthorn described a customer that deployed an AI agent for software quality assurance. The agent extracted serial numbers from screenshots automatically. It worked perfectly for two months — then started reading the wrong numbers intermittently.
The cause was subtle. The underlying vision encoder behaved differently depending on where the serial number appeared on screen. A software change that was invisible to humans triggered the failure. The agent passed every internal test but collapsed in the real world when conditions shifted slightly.
This pattern is common. VentureBeat’s own research found that half of surveyed companies shipped agents that passed internal evaluations but failed with real customers. Most companies track uptime — is the agent running? — but ignore accuracy. Is it actually doing the right thing?
That is like checking your car’s engine is running but never looking at the road.
Think of AI agents as interns, not experts
Silverthorn’s most useful advice was not technical. Inside Amazon’s AGI lab, researchers call their AI agents “interns.” The joke is that “I will have my intern talk to your intern” — agents interacting with each other. But the philosophy behind it is serious.
Interns are powerful but occasionally clueless. They can do amazing work and spectacular damage, often without realising the difference. Managing them requires management skills, not just software skills. You ask: what could go wrong? What backups do you have? How do you undo a mistake? What level of risk can you accept?
For Irish small businesses, this is the right way to think about AI agents. Do not ask whether an AI tool can do an impressive demo. Ask whether it can do the same task correctly a thousand times in a row, under varying conditions, without supervision.
How to escape pilot purgatory
If you are testing AI tools in your business but hesitating to put them into real use, you are not alone — you are in the 95% majority. Here is how to move forward safely:
- Start with low-risk tasks. Do not let your first AI agent handle payments or customer complaints. Start with internal tasks where mistakes cost time, not money.
- Measure what matters. Track accuracy, not just uptime. An agent that is running but producing wrong results is worse than one that is offline.
- Build in human oversight. Have a person review the agent’s work before it takes action. Gradually reduce oversight as you build confidence.
- Test in production carefully. Your controlled test environment will never match the real world. Run limited live tests with monitoring and rollback plans.
The businesses that eventually deploy AI agents successfully will not be the ones with the smartest technology. They will be the ones with the best management practices. Start building those practices now, and you will be ready when the reliability catches up.