AI “illusions” (those convincing but wrong answers) have attracted a lot of media attention, and like the recent New York Times article, AI has become increasingly powerful, but its hallucinations are getting worse and worse. When you deal with consumer chatbots, hallucinations are a real danger. This is a more serious problem in the context of AI’s business applications. Fortunately, as a business technology leader, I also have more control over this. I can make sure the proxy has the correct data to produce meaningful answers.
Because that’s the real problem. In business, there are no excuses for AI illusions. Stop blaming AI. Blame yourself for using AI incorrectly.
When generating the illusion of AI tools, they are doing the work they design – providing the best answer based on the available data. When they form something, when an answer that is not based on reality is produced, This is because they lack relevant data, can’t find or understand the problem. Yes, new models like Openai’s O3 and O4-Mini are becoming more hallucinated, and their behavior is more “creative” when they don’t have a good answer to the questions they ask. Yes, more powerful tools can hallucinate more – but they can also produce more powerful and valuable results if we are ready for success.
If you don’t want your AI hallucinations, don’t starve it to get the data. Provide the best, most relevant data for the problems you want to solve, and don’t go astray.
Even so, when using any AI tool, I recommend that you maintain your critical thinking skills intact. The results provided by AI agents can be productive and enjoyable, but the key is to unplug the brain and let the software do all the thinking for you. Continuously ask questions. When an AI agent gives you an answer, ask the answer to make sure it makes sense and is supported by the data. If so, it should be an encouraging signal that deserves your follow-up questions.
The more questions you have, the better the insights you will get.
Why hallucinations occur
This is not a mystery. AI is not trying to lie to you. Each large language model (LLM) AI predicts the next word or number based on probability.
At a high level, what’s happening here is that LLMS strings sentences and paragraphs together in a word, predicting the next word that should occur in the billions of examples in its training data. The ancestors of LLMS (except Clippy) are automatic complete prompts for text messages and computer code, human language translation tools, and other probabilistic language systems. With the increase in brute force computing power, coupled with the training of the amount of data on the Internet-scale, these systems became so “smart” that they could have a full conversation through chats, as the world learned when introducing Chatgpt.
AI Naysayers like to point out that this is different from true “intelligence”, only software that can refine and reflect on human intelligence that has been incorporated into it. It is required to summarize the data in a written report and to mimic the way other writers summarize similar data.
As long as the data is correct and the analysis is useful, this makes me an academic argument.
What happens if AI has no data? It’s full of blank space. Sometimes it’s very interesting. Sometimes it’s a total mess.
This is 10 times the risk when setting up an AI agent. Agents should provide actionable insights, but they make more decisions along the way. They performed a multi-step task, where the result of step 1 informs the steps of steps 2, 3, 4, 5, …10…20. If the result of step 1 is incorrect, the error will be amplified, making the output of step 20 worse. In particular, because the agent can make a decision and skip the steps.
Do it right, agents do more to deploy their business. However, as AI product managers, we must recognize the greater risks of carrying greater rewards.
This is what our team does. We saw the risk and solved it. We don’t just build a fancy robot. We make sure it runs on the correct data. Here is what I think we are doing right:
- Build the agent to ask the right questions and verify that it has the right data. Make sure the agent’s initial data entry process is actually more certain and less mindful. When do you want the agent to have no correct data, instead of continuing to perform the next step, rather than constituting the data.
- Building scripts for your agents – make sure you don’t invent new plans every time, but instead take a semi-structured approach. During the data collection and analysis phase, structure and environment are very important. When an agent has facts and is ready to write a summary, you can let the agent relax and be more “creative”, but first, get the facts done correctly.
- Build a high-quality tool to extract data. It’s more than just an API call. Take a moment to write code (people still do it) to make the right data and all kinds of data gathered and incorporate quality checks into the process.
- Let the agent show its work. The proxy should reference its source and link to where the user can verify the data from the original source and explore it further. Slight hands are not allowed!
- Guardrail: Think about what might be wrong and establish protection in errors that you absolutely cannot allow. In our case, this means that when the agents analyzing the market have no data, I mean our network-like data, rather than some random data source extracted from the network – making sure that it does not constitute something is an essential guardrail. For an agent, it is better to be unable to answer than to provide false or misleading answers.
We incorporate these principles into our recent launches of three new agents, and more. For example, we prepare agents for salespeople’s AI meetings not only ask for the name of the target company, but also provide detailed information on the goals of the meeting and who to be with to provide better answers. It doesn’t have to guess, as it uses a lot of company data, digital data and execution materials to inform its recommendations.
Is our agent perfect? No. No one has created the perfect AI, not even the largest company in the world. But facing problems is much better than ignoring it.
Want less hallucinations? Give your AI a lot of high-quality data.
If hallucinations create hallucinations, maybe it isn’t the AI that needs to be fixed. Maybe this is your way to take advantage of these powerful new features without spending time and effort getting them right.
