“2025 will be the year of the AI agent.” Wait — didn’t we already say that in 2023? Back then, I talked about chaining AI models in my podcast. The excitement around AI was huge, and by 2024, many companies rushed to try generative AI, with some even using chained models. But many of these projects failed to scale. Why? Because they were test projects that were not integrated into the system infrastructure or were missing guardrails and quality controls. This will change in 2025, but the product challenges will remain the same. To see what AI agents can do in 2025, let’s consider a simple example: an email-answering tool. Imagine a system that drafts replies to emails automatically. This example shows the opportunities and challenges businesses face with AI agents.
Why GPT Wrappers Are Not Enterprise AI
The simplest way to build an email-answering tool is with a GPT wrapper. I saw many of these in 2024. These are basic setups where you connect AI to a small interface. For our use case, this means getting a ChatGPT API key, writing some code to take an email as input, adding a prompt telling the AI what to do, and displaying the response in a clean interface.
- System Prompt: Answer as if you were Lutz. Lutz always ends his messages with “Cheers!”
- Input: Hi Lutz, you are invited for dinner on Thursday. Can you come? Cheers, Tim.
- Output: Thanks, Tim! Let me check my schedule and I’ll get back to you soon. Cheers!
Even in this simple example, several key challenges emerge for enterprise AI today:
- No System Integration: The tool can’t check my calendar to see if I’m free.
- No Context: It doesn’t know if I like Tim or avoid these types of events.
- No Security: What if the email asks for private details, like my social security number?
- No Guardrails: How would it handle a controversial question, like asking for political views?
- No User Control: How much input does the user have in shaping the AI’s response?
- Hallucinations: AI sometimes makes things up, as the warnings on ChatGPT pages constantly remind us: “ChatGPT can make mistakes.”
Large language models are excellent at tasks like summarization or acting as interfaces, but alone they are not enough. As I emphasize in my eCornell certificate program, each of these challenges can be addressed. Let’s improve our tool by building AI agents within a workflow.
The Workflow of AI Agents: More Than Generative AI
AI models can be connected or “chained” to build workflows where the output of one model becomes the input for the next. Think of tools like Zapier or IFTTT, but powered by AI. Instead of fixed steps, the process is dynamic and adapts to each situation. These workflows don’t always rely on generative AI like ChatGPT. In fact, they often don’t—generative AI can be too slow and expensive.
Here’s how this could work for our email tool:
- Input: Hi Lutz, you are invited for dinner on Thursday. Can you come? Cheers, Tim.
- System Prompt: Analyze the email and figure out the steps needed to respond.
- System Output: (1) Check the calendar to see if the user has time. (2) Look up past emails with Tim and previous dinner invites. (3) Predict the chances the user would want to attend based on past behavior. (4) Create three draft replies: one accepting, one declining, and one asking for more details.
- System Execution: Carry out the outlined steps.
- Output: Present these three responses to the user.
This is how a chained AI model works. It overcomes many of the earlier issues. But the large language model is just one of many tools.
All tools need to be integrated and quality-checked. For example, our setup touches on:
- System Integration: Checking calendars and pulling data isn’t AI but requires system connections.
- Context Search: Using AI to retrieve relevant past interactions (called RAG or Retrieval-Augmented Generation).
- Traditional AI: Predicting the likelihood of attending uses classic data analysis.
- User Design: Offering multiple options improves usability and control.
In short, it’s about product work—creating solutions that are reliable and valuable.
2025 – AI Agents for the Enterprise
2025 will be the year of AI agents. AI will simplify, enhance, or automate workflows across industries. But there won’t be one “killer app.” If such an app exists, it would likely be Search—see my prediction on Search. Instead, we’ll see smaller workflows in areas like customer care (see my investment in ultimate.ai), legal support (see flank.io), or sales (see my company, r2decide.com).
To build these solutions, engineers and product managers will need to focus on creating value. In my course, Designing and Building AI Solutions, I teach a practical framework for delivering value to enterprise customers in media, finance, healthcare, eCommerce, and other industries:
- Define the business objective.
- Collect and clean the data.
- Develop the agentic workflow.
- Test with users.
- Create a feedback loop.
This sounds simple, but there are many considerations, including addressing risks such as bias and ethical issues. Most importantly, the focus must remain on value creation. Let’s build together.
Follow me here on Forbes or on LinkedIn for more of my 2025 AI predictions.
