More businesses are transitioning from API-based software to autonomous AI agents in an effort to remain competitive, but doing so requires significant planning to mitigate problems like hallucinations and their consequences.

Until recently, most businesses relied on traditional platforms and APIs to support their operations, in large part because these tools generally produced predictable results. When users performed a defined action, the software followed established authentication, authorization, and transaction rules.

Contrast this with agentic AI, which, while potentially more productive than traditional APIs, introduces the issue of interpretation. Specifically, AI systems function by interpreting a user’s intent and independently determining which actions to take. This flexibility creates value, but it also introduces a sense of uncertainty because an agent may take an action the user did not anticipate.

AI’s reliance on interpretation to function prompts an important question for enterprises looking to adopt AI systems: how can organizations benefit from autonomous systems without losing accuracy, control, and accountability?

APIs and Platforms Remain the Foundation of Agentic AI

To suggest that transitioning from APIs to AI systems means getting rid of APIs entirely is, in a word, false. Agentic AI is not a replacement for APIs, commerce platforms, databases, or transaction systems, as agents depend on those technologies to retrieve information and perform actions.

APIs are a necessary part of how agents communicate with business applications. Platform architecture, meanwhile, determines which actions are available, how permissions are enforced, and how systems respond at scale.

One individual discussing why APIs remain important today is Ashay Satav, a technology and product leader who has built commerce and API infrastructure at Blackhawk Network, Intuit, and eBay, including eBay’s AI-powered magical listing, which reached over 10 million sellers and won the 2024 AI Breakthrough Award for Best Overall Generative AI Solution. Much of his work has shown how foundational API and platform work has contributed to the infrastructure supporting modern AI agents.

Satav argues that the platform layer, not just the underlying model, determines whether an agentic system can be trusted with real business consequences. “The model is doing the reasoning, but the platform is doing the governing,” he explained. “You can have the smartest model in the world, but if the API layer underneath it doesn’t enforce what actions are actually allowed, at what scale, and with what fallback when something goes wrong, you don’t have an agent; you have a liability.”

In his view, this is precisely the gap that caused so many early enterprise AI pilots to stall before reaching production: teams optimized the model’s intelligence without first building the platform architecture to safely contain what that intelligence was allowed to do.

The Business Cost of AI Hallucinations

One of the most pressing issues facing AI adoption is hallucination, a phenomenon wherein an agentic AI effectively “makes up” information in response to a prompt. In minor cases, this means producing an incorrect sentence, but in more severe instances, an autonomous agent may use incorrect information to perform or recommend a business action.

Ashay discusses this issue in the context of e-commerce listings, pointing out that an AI listing tool may correctly recognize that an image contains a white cup and identify its brand, but it may embellish the cup’s condition or features when writing the description. That exaggerated description could then persuade a customer to buy that cup, only to realize later on that their expectations were mishandled.

If allowed to persist, this technical issue can snowball into a series of negative business outcomes, including returns, customer complaints, buyer remorse, reduced trust, and negative seller-performance metrics.

In an interview, Ashay explained it this way: “The hallucination rate of the agents where they are trying to, you know, sell something because they have to. I think that needs to be controlled so that the buyer remorse is at a minimum.”

He also noted that this type of return is associated with the “Seriously Not As Described,” or SNAD, metric at e-commerce platforms like eBay. The fact that this metric exists suggests hallucination rates are as connected to technical problems as they are marketplace outcomes.

Grounding Agents With Retrieval-Augmented Generation

Of the various methods used to reduce hallucination rates, one of the most useful so far is Retrieval-Augmented Generation, or RAG. An RAG system retrieves relevant information from an approved knowledge source before an AI model generates an answer or takes an action; in an e-commerce context, that means the system could examine historical listings and verified product information rather than allowing the model to invent details from its general training.

RAG’s usefulness is dependent on a multitude of factors, however, including the accuracy and relevance of the source data, the use of metadata filters, and the number and type of retrieved records, to name just a few. RAG remains imperfect at the best of times, and given that poor retrieval can still provide irrelevant or misleading context, teams must test retrieval and generation as one connected system.

Since RAG relies on so much information to properly function, many RAG systems store embeddings in a vector database. These embeddings allow the system to identify information that is semantically related to a request. That information is divided into chunks, though chunks that are too large or small may include or miss out on important details.

Finding that midpoint can be difficult, so teams should consider experimenting with chunk size, metadata, similarity thresholds, retrieval depth, and ranking methods.

When and How to Use Large Language Models

Large language models (LLMs) have swept through many businesses across industries, but in truth, not every task or operation benefits from using one. Using a large model for a simple or repetitive request can increase latency, cost, and system complexity, so enterprises should consider routing simpler tasks through APIs or smaller specialized models.

When considering whether or not to use an LLM, it may help to approach the decision this way: when interpretation, reasoning, or natural-language generation adds meaningful value to a task, using an LLM may be pertinent. Outside of those qualifications, LLMs may cause more problems than they solve.

Note that many traditional AI quality programs often focus on whether an LLM or other agentic AI’s answer is accurate, and while that emphasis is useful, more complex AI programs usually require teams to evaluate both the answer and the action that follows.

In practice, that may look like defining which actions an agent is allowed to perform, applying authentication and authorization at the API level, requiring confirmation for sensitive transactions, or establishing limits for spending, publishing, deleting, or modifying data. User consent here should be specific and understandable; if a user asks an agent for assistance, that agent should not automatically authorize every possible action the system could perform.

Autonomy Requires Stronger Foundations

Technical metrics such as response accuracy, retrieval relevance, and hallucination rates are valuable when measuring agent performance, but they often provide an incomplete picture. In addition to looking at technical metrics, organizations should also monitor the effects of agent behavior on operational and customer outcomes.

That split between technical and operational outcomes is part of why APIs and platform engineering remain important, especially as more businesses adopt agentic AI. AI agents benefit from flexible AI reasoning, of course, but that reasoning is best supported with trusted data, carefully designed APIs, explicit permissions, cost-aware model routing, and measurable guardrails.

Hallucinations may appear to have limited consequences, but in truth, in an autonomous system, inaccurate output can become an inaccurate action with direct consequences for customers and the business.

This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program.