In previous posts, we’ve talked about Agentic AI, how it’s different from Generative AI and other types of models. Plus, we covered RAGs, Agents, and chatbots.
Part One: Initial Training Phase
Like many people playing with the technology, we’re curious and impressed by what we can do with agentic AI. AI agents use systems that are trained on specialized data sets, and are capable of independently solving multistep problems in service of a particular goal. They’re AI that has a specific job – essentially, they’re workers.
But as with any new “worker,” you need to train your agents, and it’s both different and similar to training a human.
How Does AI Learn?
How Is Agentic AI Learning Different From That?
Agentic AI takes an LLM, with all of its abilities, and trains it to use business specific data (provided by RAGs) as the primary context the agent uses to perform a certain action (or “job”). And it can adapt over time and learn to solve more complex versions of that job. That’s why they’re called AI “agents” – they function with a degree of agency.
But they have to be trained to perform the role you’re asking them to do, so that when they’re asked a question, they’re able to serve up relevant answers that are immediately useful to the person who is asking.
Training an Agent vs Training a Human
We recently hired a data analyst. Around the same time, we started training an AI agent to perform some data analysis tasks with the intention of making our analysts’ lives easier. The AI agent is incredible at some things – it can scale and do calculations exponentially faster, for example. But in other ways it is initially a lot faster and easier to train the employee, and here’s why.
Business Common Sense
When some metrics (like number of sales) steadily go up, it’s obviously positive. But when other metrics (like ad spend, CPA) go up, it’s generally bad news! Similarly, several metrics fluctuate in the course of business (again, sales and spend are good examples), and these minute fluctations are often not cause for alarm. And then there are metrics like Total Subscribers or Total Lifetime Customers whose numbers you want to always increase, and a decrease is concerning. This is something that a human analyst intuitively understands, but many LLMs don’t know the difference between “good” patterns and “bad” patterns out of the box.
To teach an LLM, you can list these details out in your context instructions. Or you can teach it to read a metrics schema that has helpful context embedded in it for the metadata for each metric. (This schema should be accessed by RAG before the agent analyzes data).
Know Your Audience
Before she started her first assignment, our new human analyst asked important questions like, “How much does this stakeholder know about the details? How much detail do they want in their answer? Are they a subject matter expert?” But our AI agent treated everyone who asked a question the same way.
So to help the AI agent better understand who was asking the questions, we created personas and roles, and fed those into the context data during training. This way, it could determine whether to respond with granular information, or keep its responses high-level.
Secondly, we taught our agent how to structure analysis in a way that their audience was ready to receive. If a user asks “how are things going?,” we didn’t want the agent to return all of the data it has available, or spout off a lot of nitty-gritty details about specific campaigns. Instead, we wanted the agent to be able to walk the user through a 50,000 foot view of how sales or new subscribers (or whatever business-specific metric they care about) are doing compared to recent, average, and year-over-year.
So when you’re first training an AI agent, you need to begin by explaining the high-level information first, before getting to the lower-level data. This means teaching it how the business is doing, what patterns and trends are positive (and which ones aren’t so positive), where it operates and what markets are significant, brand information – all the big stuff that someone who works at the company might know and take for granted. Then, once all that’s in place, you can work down to explaining all the lower-level information, like campaign performance.
Remember Context Is Key
Because our human analyst is smart and attuned to the world, she intuitively knows that there’s a question behind every question she’s asked. For example: if someone were to ask how an ad performed during the Super Bowl, she wouldn’t compare it to ad performance on any given Sunday — she’d compare it to ad performance from previous Super Bowls.
AI agents, on the other hand, have to be taught that contextual knowledge. They require a little more work in the set-up process to learn about business contexts and other background information. When training your AI, think about what your end user is actually asking, and what they need to know. Trying to put yourself in the shoes of your audience and figuring out what your end user cares about is (always) a crucial part of creating a useful tool.
Keep It Bite-Size
This is true for both humans and AI agents: don’t overwhelm them. When you’re directing a person, if you throw 17 notes at them at one time, you’ll end up confusing and overtaxing them. While an AI agent has the capability to implement many notes at once, doing so may similarly make training more difficult. If you go slowly and give the AI agent one note at a time, you can better track which notes are effective and how you should adjust your next prompt. Staying gradual is a good way to make sure your training is methodical, measurable, and if necessary, repeatable.
Your approach to training an AI agent is just as important as your approach to training any new hire. While AI (obviously) learns differently than people do, setting it up for success with a solid foundation is crucial to making it an effective part of your team.
Part Two: Reinforcement Learning
This post has focused on training a model and human with initial context to start doing their job. The next step is reinforcement learning. In our next post, we’ll cover how that works differently: it’s harder to do with a machine because you need a higher volume of data, but easier in that you’re providing falsifiable examples.
