When ChatGPT launched in late 2022, the world changed. Forever more, the history of business technology would divide into before generative AI (genAI) and after.
Suddenly, AI was powerful enough to accept natural language prompts as input and generate plausible natural language outputs. The creators of ChatGPT and other early large language models (LLMs) trained their systems on content from the Internet itself, giving these applications an immensely broad but inherently flawed basis for generating output.
GenAI is also the driving force behind AI agents – autonomous programs that leverage AI to gather information and take action, despite ongoing issues with the underlying technology.
Today, businesses are finally coming to terms with the genAI tradeoff: its enormous power to understand human prompts and return convincing answers despite its flaws.
Given the importance of genAI for building AI agents, it’s essential for organizations to gain a clear idea of what problems genAI is well-suited to solve to extract maximum business value from the technology as their deployments of agentic AI mature.
GenAI’s Sweet Spot
Despite its growing popularity, genAI is basically little more than an engine for making plausible guesses based upon data from three sources:
- Training data that the model creator uses to build the model. Training can be a time-consuming and expensive process, so model creators typically conduct training before putting the model into use.
- The information that users put in the prompt. When users interact with an LLM, they can provide information as well as asking questions. How much information users can include in a prompt depends upon the model. We call this prompt limit the context window.
- Additional information that the genAI application feeds into the model at the time of the query. For example, a prompt might request a summary of several documents, which the model must fetch via APIs or other integrations in order to respond to the query. We call this process retrieval augmented generation, or RAG.
Based upon these data sources, the LLM does its best to respond with what it calculates is most likely to be the best answer to the query.
Since such responses are inherently probabilistic, it’s unlikely that they will be accurate every time. As a result, genAI is good at some things but falls short on others.
GenAI is particularly good at:
- Summarizing information. Taking large quantities of information and boiling it down to its salient points is an excellent use case for genAI, as it has to make judgment calls on which information to include in its response – a task it handles well.
- Matching responses to input data. For example, genAI excels at, say, searching for a new supplier that matches business criteria, either in the prompt or available via RAG.
- Uncovering specific information within large quantities of data. Finding needles in haystacks is right up genAI’s alley. This ability explains why people are using genAI as search engines – either for the Internet at large or targeting specific business-centric data sets.
- Making plausible conclusions about actions to take. To support AI agents, genAI is able to leverage available information from a variety of tools and data sources to support human decision making.
GenAI’s Weaknesses
Take heed, however, that genAI isn’t good at everything. Here are some notorious weaknesses of the technology:
- Arithmetic: Ask genAI what 2 + 2 equals, and it will probably say four – but that answer is not guaranteed. All it takes is the appearance of, say, ‘2 + 2 = 5’ in its training data to convince it to produce an incorrect response.
- Creativity: genAI cannot deliver answers that go beyond its source data. Responses may appear to be original (thus illustrating true creativity), but that appearance is simply a function of its ability to create plausible responses. If you want true creativity, you still need a human.
- Moral or ethical decisions: genAI has no way of knowing whether its training data are biased or favor unethical or immoral responses. If you feed it, say, several resumes of white men, it will be only too happy to tell you a white man is the right person for the job.
- Security issues: genAI opens up entirely new types of holes in an organization’s threat surface. From prompt injection attacks (submitting malicious information to generate a nefarious result) to shadow AI (unmonitored, unregulated use of genAI in an organization), security teams have a range of new challenges on their plates.
- Understanding when there are problems with input data: the old saying ‘garbage in, garbage out’ applies to genAI. If there are data quality or veracity issues with the source data, then genAI responses will reflect those issues.
Given these weaknesses, it is essential to put together a genAI strategy with your eyes open. The technology may appear to be magically transformative, but it’s just another tool – and you should always keep in mind the maxim ‘the right tool for the job.’
The Intellyx Take
The first step in harnessing the power of genAI is identifying the right use case for your organization and then mapping its capabilities to solutions you need.
GenAI alone excels at finding and summarizing information. However, when empowering agentic AI, the technology supports the automation of dynamic business processes, as organizations leverage tools like the Boomi Enterprise Platform to implement agentic workflows.
This article is the first in a series. As my colleague Eric Newcomer and I will explain in future installments, one of the most powerful uses of genAI today is solving complex automation challenges that demand the reasoning and autonomy only agentic AI can deliver.
Read Boomi’s Agentic Transformation Playbook to learn how to get your organization primed for AI success.
Copyright © Intellyx BV. Boomi is an Intellyx customer. Intellyx retains final editorial control of this article. No AI was used to write this article.