An agentic AI reality check
No one can deny the tech sector’s level of excitement when something new starts to gain traction. Rather like the car mechanic or builder who urges their customers to buy more expensive components or materials because “they are better”, agentic AI is the new piece of wonder-tech.
Analyst firm Forrester says agentic AI systems are able to plan, decide, and act autonomously, orchestrating complex workflows with minimal human intervention. A few years ago, a similar description could have been applied to robotic process automation (RPA). Some industry experts even used the term “intelligent automation”, because it seems robots aren’t that intelligent.
Nevertheless, if RPA in all its guises, actually did what it is supposed to, surely by now, we’d have discovered all the broken, manually-intensive workflows that sap business efficiency.
And these will have been automated with logic that reroutes business processes based on simple decision-making: “If I see an exception to what I’m expecting, then I take this particular action, otherwise I pass the job onto the next piece of automation.” The exception is where a human or some intelligence in the workflow is used. But the industry has a new solution: agentic AI.
Random results
Demis Hassabis, the CEO and founder of Deepmind, has a vision of what agentic AI will eventually look like. It’s early days, but what we are starting to see, to quote Hassabis, is a world where AI agents spend more time thinking and planning before they act. And when they do act, they draw on the expertise of multiple agents to solve a problem. DeepMind has used to play the strategy game StarCraft II, where a league of agents compete or cooperate.
Games have prescribed rules, which makes it relatively easy for agents to decide on the best outcome, even when the AI may produce erroneous results – it’s probably a bit like being called out for cheating. But the real world is not a strategy game, no matter how many cards one country leader may think his opponent is holding. And cheating is not on the cards, when important, real world decisions, must be made with 100% reliability.
AI models can produce erroneous results and as Hassabis notes: “If your AI model has a 1% error rate and you plan over 5000 steps, that 1% compounds like compound interest.” By the time the 5000 steps have been worked through, the compounded error, according to Hassabis, means that the possibility of the answer being correct is effectively random. This simply is not progress. Instead of buying into the agentic AI hype, IT leaders should first look introspectively at what needs improving in their own organisations.