We are moving from time of LLMs without the context or with the context of a prompt to AI Agents which should be self-reliant and acting without peoples’ input.
If you have worked in any company in a white-collar job, you might start to think: “Most of the work I do is based on one or two sentences... How will AI Agent get the context to be able to gauge the right course of action?“
That’s spot on! Thank you for thinking that!
You: “AI Agent, fix the production, because my boss needs it fixed!”
AI Agent: “Wait, what issue? What production? What has happened? Where do I find more information for this?”
You: “I don’t know, just get it fixed.”
AI Agent: “Oh, ok… Here is some answer that is average of Reddit and LinkedIn posts, hope it helps.”
You: “But my boss…. senior management… clients… they NEED it!”
AI Agent: *Shrugs virtually*
In a lot of cases we just assume there is the context somewhere… unwritten, unspoken. Just cultural expectation that you don’t ask the “Why?” repeatedly because no one wants to say “Because I messed it up and I don’t want to be fired!”.
I believe we can all put ourselves into recent scenario when we were doing something and we had no understanding why. On the other hand, we will remember a task where we understood what the outcome is, how it fits into wider picture and what are the nuanced issues. I don’t think there is a LinkedIn influencer that wouldn’t agree that latter is better.
It is clear that having wider context is better to get to desired outcome and to keep the staff engaged and help them grow. And there is a similarity with how to get more out AI Agents. Just imagine AI Agent doesn’t have access to your codebase and you want it to fix something in it. Ridiculous, isn’t it? But how does this scenario translate to some other operations?
Let’s take client relationship management.
You are a new account manager, you are a week in the job and you get agitated email from a client. First thing you might give them canned response that you are new, looking after their account and that you will look into it.
So far, so good. AI Agent can do the same.
Now you need to figure out why they are agitated and how to manage it. This is where the fun starts, so let’s go into two extremes for the sake of argument.
1) There is no CRM system, no conversations were recorded, the contracts are nowhere to be found. The client was managed solely be a person who left and no one really knows anything about them. You are not set up for a success, are you? You basically need to get the information out of client without them realising you don’t have it and start to build up that database of what’s going on.
What do you expect AI Agent will do in this case? You don’t have any information to make any decision, how could it do anything sensible?
2) You have a CRM system, where you have all contracts, emails, transcripts from calls, continuous sentiment, your promises and their expectations. You have roadmaps from other teams, SLAs and historic services and order data. You have written down business objectives, desired client segments to support and clients you don’t want to lose.
You can start to immediately look into the details and to be able to unwind what has happened and what has to happen next. In this scenario an AI Agent has strings it can follow.
In reality, in most of the companies, it’s quite a mess (Schrödinger’s data, Managing data is a pain).
Why is it relevant to culture of transparency? Just imagine the scenario 2) you have EVERYTHING written down explicitly. What went well, what went poorly, what the business objectives are and what the goals are. It’s a massive commitment to write all things down in a way they should be objective. To keep conversation open, honest and your house clean at all times so to speak.
And there will be a lot of companies not solving this challenge, but implying they do. Which means that you will hit a wall with your AI Agents no matter what. There will be a limit of what you will allow them to do when you don’t yourself trust that you have all the context, information, data up to date.
Where the open challenge is that the most (perhaps all) the companies that now “bring context to agent” are effectively doing a lookup to a system. It solves problem of being able to access data, but not if the data describes well enough the real situation and in a way that it sets AI Agents up for guessing the right course of action.
Just above average 