Why Executives Should Take Note of AI Agents

In the last 18 months, as Executives took stock of the ‘AI situation’, some of them concluded that the technology had potential but the risk-reward equation was not favorable. Or that it was not favorable enough to justify a major investment or a major change in their business strategy. In their opinion, a small experimental investment was adequate but for the most part, the conclusion was “wait and watch because its just too risky to take the plunge”. Not only were the risks too high with no answers but there wasn’t even a clear line of sight to those answers. Well, that’s the precise point on which there are major industry updates which Executives should pause to note. This article aims to bring clarity to those industry updates. More specifically, there is now a very promising approach which is being employed to mitigate the risks down to an acceptable level for mass application. At the center of this mitigation path are AI Agents. Let’s jump right in to see how it all comes together.

What were the risks and gaps?

Hallucination: AI Models are known to provide inaccurate and made up answers (called hallucination in AI lingo) which are not based on any authoritative source. For example when you ask ChatGPT to help you troubleshoot an issue with a humidifier, there is no guarantee that the answer you get is based on that specific humidifier’s owner manual but you’ll get a response, nevertheless. 2) Bias: LLMs are known to respond with bias because they were trained with data from the world wide web, which inherently has bias in it. 3) Mistakes on simple tasks: Though LLMs have tremendous intelligence in them, they often make mistakes on some simple math questions or can’t function with basic common sense. For example, if a child asks an AI model how to set fire to a car, it might actually make suggestions. 4) Struggles with multi-step tasks: Let’s say you want to automate your process of determining inventory level for a product to place an order with your vendor. Your current process may involve multiple steps like capturing current inventory level, competitor pricing, social media trends, marketing plans etc. Until now, AI models have struggled significantly with such tasks where you complete one step and then provide the outcome of that task as input to the next step and so on. Most business processes are multi-step, so this has been a major gap.

So, what’s the new solution?

Instead of a technical breakdown of those issues and their technical fixes, let’s understand the solution approach with the help of a simple analogy. Think of the AI models as a freakishly talented soccer player. This player has a brain that can make critical decisions using factors no one else has been able to. For example, this player has, in his brain, stored the average speed of every single player on the field based on the weather, the location, the opposition, their mental state, the state of the match, their age and a multitude of factors. He can predict with a high degree of accuracy the direction and speed which will give him the best chance of getting to the goal post. In essence, imagine a player with a brain that is several times more intelligent from a football perspective than that of the best current players. If this soccer player’s every single kick is based on such intelligence, insight and calculation, then naturally that becomes a significant advantage for that team. However, there is a catch. Despite being such a soccer genius, he sometimes gets the most basic things wrong. For example, he just never remembers the times for team meetings. He forgets to wear the team uniform before a match and shows up in his pjs. When you ask him for a specific soccer statistic, his response could be completely made up. If you put him in a press conference and answer questions, he could make discriminatory comments on fellow players based on social statistics. Lastly, if you try to make him the team captain because of his incredible abilities, he doesn’t do well because coaching and captaining requires taking input from different sources or other players and crafting a strategy. He is simply not good at that. He gets lost. So, how do you get this player on the team but without all the issues above? By making the player do what he’s good at but making him do ONLY what he’s good at. We don’t need him to do the other tasks like — planning his day so he’s on time for the meetings, speaking to the press, or getting historical statistics or even captaining/coaching his team. We can have the support staff take care of all that. That way, we can leverage this player’s freakish talent while mitigating these other risks which come with him.

That’s what AI Agents do to an AI Model

Building on our analogy, AI model is that freakishly talented soccer player. The support staff for that player is what AI Agents are to the AI model. They take care of these tasks which we need to complete while automating a process, but AI models struggle to accomplish. From an AI perspective, there are three categories of agents —

1. Tools: This is to complete small sub tasks which an AI model is not reliable at, and we have standard APIs available to accomplish these tasks— like APIs to perform math calculations, getting facts from Wikipedia or completing an action (e.g. turn on a switch/bulb, send an email etc.) or responding from a proprietary data source. Instead of letting the AI Model respond based on its internal prediction logic (which is what basically an AI model is), we direct it to leverage our APIs when such small tasks are encountered in user requests. That way, the risk of ‘hallucination or making stuff up’ goes dramatically down.
2. Reflection to reduce bias: This is an interesting approach where we are pitting an AI model against an AI model. When an AI model responds to a user request, instead of providing that response to the user, we first ask another AI model to evaluate if that response is appropriate. If this second AI model finds anything inappropriate then it gives that feedback to the first model, so it comes back with a better response. This continues till our bar is met. This process is called ‘reflection’ because we’re asking the AI model to reflect on its response.
3. Planning: This aspect is still not reliable enough but it’s a matter of time. The concept is very logical. It involves teaching the AI model how to break down a task into a ‘plan’ of multiple steps and then completing that plan, one step at a time. The key aspect to note is that some sub-steps may fetch a new piece of information which may mean we need to refresh the original ‘plan’. For example, if we need the AI model to automate the process to define pricing for a product, then the ‘plan’ may include steps like a) check competitor prices, b) check inventory, c) check upcoming marketing campaigns and so on. At each of those sub steps, the AI model may uncover new data which may need it to refresh its plan.

Together, these ‘AI Agents’ are chipping away at the major risks and limitations of AI. When ChatGPT came out, it showed a lot of promise and people expected it to do everything, even aspects we didn’t really need ChatGPT (or AI in general) to do. Yes, AI was mocked for getting basic math wrong, some scientists responded by saying — why are we even asking AI to do the math when we already have the tools to take care of that. Let’s use AI for aspects which were traditionally not possible and supplement it with capabilities (i.e. Agents) which are already available. That’s what this ‘agentic workflow (as it’s called in the tech circles)’ is all about.

Conclusion

This Agent based approach to get the maximum value of AI is not just promising but inevitable. Very soon, an AI agent will be a term as common as App is. An agent can be as simple as a calculator or a Cooking Assistant (Consumer level) or a Pricing Agent (for a business use case). The concept is very foundational and has validation from all the key pioneers and the companies leading the way in AI. The most important message here for executives and leaders is — there is now a clear line of sight for risk mitigation — not for all applications of AI but for a large volume of applications, especially business processes. The planning ability does need more reliability but there is lot that the AI models are already ready for. Several companies will recognize this development and accelerate their adoption plans for AI. It’s worth noting that while some of the generic agents (e.g. calculator, search engine, Wikipedia validation etc.) will be available in the market, companies will have to build several APIs themselves (e.g. APIs which expose their proprietary data sets or capabilities in a manner that that AI models can consume). All of that will take investment, work and most importantly time. To be on the right side of that time, executives who decided to ‘wait and watch’ or just do light experiments in AI should reassess their conclusions in light of these developments.

If you found this article useful, please clap and / or subscribe. Thank you!