<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Martin Ross on Medium]]></title>
        <description><![CDATA[Stories by Martin Ross on Medium]]></description>
        <link>https://medium.com/@rossmrtn?source=rss-af1f2925b040------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*lsq-z_zGB2mqhk03mn6I8A.jpeg</url>
            <title>Stories by Martin Ross on Medium</title>
            <link>https://medium.com/@rossmrtn?source=rss-af1f2925b040------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:02:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rossmrtn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Approaches in using Generative AI for Application Integration]]></title>
            <link>https://medium.com/@rossmrtn/approaches-in-using-generative-ai-for-application-integration-07b04aae907a?source=rss-af1f2925b040------2</link>
            <guid isPermaLink="false">https://medium.com/p/07b04aae907a</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[integration]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Martin Ross]]></dc:creator>
            <pubDate>Fri, 19 Apr 2024 17:13:26 GMT</pubDate>
            <atom:updated>2024-04-19T17:13:26.089Z</atom:updated>
            <content:encoded><![CDATA[<p>Co-author: <a href="https://medium.com/u/501e352a2fdd">Allen Chan</a></p><p>See other related stories in the “Approaches in using Generative AI” series:</p><ol><li><a href="https://achan2013.medium.com/approaches-in-using-generative-ai-for-enterprise-content-management-9dbb7a3affe1">Enterprise Content Management</a></li><li><a href="https://medium.com/@pierrefeillet/approaches-in-using-generative-ai-for-business-automation-the-path-to-comprehensive-decision-3dd91c57e38f">Decision Automation</a></li><li><a href="https://medium.com/@achan2013/approaches-in-using-generative-ai-for-workflow-automation-part-1-15aea74a3d21">Workflow Automation Part 1</a></li><li><a href="https://achan2013.medium.com/approaches-in-using-generative-ai-for-workflow-automation-part-2-14ccb3e7e6e0">Workflow Automation Part 2</a></li></ol><p>In the last 4 articles, we talked about using Generative AI in Enterprise <a href="https://medium.com/@achan2013/approaches-in-using-generative-ai-for-enterprise-content-management-9dbb7a3affe1">Content Management</a>, <a href="https://medium.com/@pierrefeillet/approaches-in-using-generative-ai-for-business-automation-the-path-to-comprehensive-decision-3dd91c57e38f">Decision Automation</a>, and <a href="https://medium.com/@achan2013/approaches-in-using-generative-ai-for-workflow-automation-part-1-15aea74a3d21">Workflow Automation</a>. In this article, we are going to tackle the topic of using Generative AI as part of an application integration solution.</p><p>Before we talk about the solution, let’s first examine what Generative AI with Large Language Models (LLMs) allows us to do today.</p><h3>Language-Based Integration — Generative AI and LLM</h3><p>Application integration traditionally revolves around well-defined interfaces to allow for integrations to be built, with interfaces defined using standard formats like OpenAPI, WSDL and various specific industry standards. With the emergence of generative AI we are seeing new interfaces based around natural language that are more probabilistic and require knowledge in prompt engineering to leverage effectively — you no longer call a REST API, you ask a question. Different language models need different ways to ask the same question, and there is no standard format as has traditionally been the case for application integration — but you can ask the question in a way that you can use more traditional tools to interpret — such as getting back a list or a JSON object instead of a paragraph — to start to enforce a more consistent structure that you can build an integration around.</p><p>There are many considerations that need to be taken into account when looking at leveraging Generative AI within application integration scenarios. The risk of hallucination is one that needs to be understood, depending on the use-case this may or may not be so much of an issue, but if you use an LLM at runtime and use the response directly then the risk of hallucination can be high. There are methods to reduce the risk — such as with RAG or using a different model to confirm accuracy — but you can not be risk free. Cost is another factor that needs to be understood, many models charge based on prompt and response — and several have token limits etc. so need to consider the constraints and how often these integrations will be running to understand the cost implications.</p><h3>Use Cases</h3><p>In this article, we will cover 3 possible use cases and describe how we can utilize language-based integration together with traditional API-based integration.</p><h4>1. Building an integration that generates LLM requests</h4><p><strong>Description: </strong>In this use case, we leverage transformation and integration capabilities within an integration runtime to build a request to an LLM and use the generative AI response to provide additional information for the integration solution. We construct an appropriate prompt to provide a response that we can use within the integration — such as human-readable summaries or additional context to accompany business data. Usually the generated results are intended for human interpretation and it is less important in these cases to create a result that matches a certain machine-readable format.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/868/1*STlKPM-zF_9gPoBMUU7cfA.png" /><figcaption>Figure 1. Building and integration that generates LLM requests</figcaption></figure><p><strong>Pros: </strong>One can rely on the LLM to formulate a well structured human-readable response based on technical inputs. This is especially useful in cases where we want the responses in different languages or if the target system has a language based interface such as a chat-based personal assistant.</p><p><strong>Cons: </strong>There is no simple way to validate the generated output will always be correct and always be the same. As a result, the usefulness of the use case will be limited to situations where lack of precision can be tolerated. For runtime interactions the anticipated workloads for the integration need to be known to calculate and understand the associated costs of leveraging the LLM.</p><h4>2. Integrating with LLM to generate a response that can be used to interact with subsequent applications</h4><p><strong>Description: </strong>In this use case we use Generative AI with an appropriate prompt to generate a response that can be used to directly integrate with a subsequent system. We construct an appropriate prompt to generate data that can be used to invoke a subsequent application — such as generating a JSON request payload matching the schema of a target REST API that we need to integrate with.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/643/1*gbvGG00-78BjOPuqmMT9mQ.png" /><figcaption>Figure 2. Integrating with LLM to generate a response that can be used to interact with subsequent applications</figcaption></figure><p><strong>Pros:</strong> Integration solutions can be enhanced to provide additional business value, enabling integration with source systems using unstructured data and leveraging LLMs to extract the relevant information from the business information and available data to automate integrations with systems exposing interfaces requiring data conforming to a given schema. This can enable integrations to be built that could not previously without human task intervention.</p><p><strong>Cons: </strong>Although the prompt can be carefully engineered to instruct the LLM to respond in a given shape to invoke the subsequent API, there is no simple way to validate the accuracy of the request data generated — as a result, there is potential for hallucinations that may limit the use cases. As with #1, the anticipated workloads for the integration need to be known to calculate and understand the associated costs of leveraging the LLM.</p><h4>3. Integrating with systems to return useful information to inform the AI application</h4><p><strong>Description</strong>: In this use case we work alongside generative AI to provide integration with systems to enable appropriate information to be presented to a user. A user interacts with an AI assistant to ask questions relating to recent orders, this will be parsed by the foundation model and result in a request being generated to an API that integrates with an enterprise CRM to provide the information required to present back to the user.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/872/1*L02jfgUCxn2bONemm42-Bw.png" /><figcaption>Figure 3. Integrating with systems to return useful information to inform the AI application</figcaption></figure><p><strong>Pros</strong>: With an increasing number of natural language channels becoming available, similar to use case #2 LLMs can be leveraged to generate well structured requests to existing APIs that integrate with other sources of data to enhance the user experience and provide additional and more tailored information to the user.</p><p><strong>Cons</strong>: As with use case #2 there is no simple way to validate the accuracy of the request data generated, as such there is potential for hallucinations and invocation errors if the API is invoked with an invalid request that would need to be considered and handled. Access control and data privacy is another key consideration here that the solution would need to address, to ensure that a user only had access to data they were allowed to access and that appropriate controls were in place to ensure any sensitive data did not end up in the model.</p><h3>Summary</h3><p>Language-based systems are here to stay and will increasingly be part of an overall integration story. However, given service API are usually not defined with human conversation in mind, some adaptions are to be expected to ensure a smooth integration experience. With proper understanding of the benefits and limitations of Generative AI, we can potentially expand the list of things that can be integrated together and will also reduce the complexity of integration.</p><p>Ultimately we are living in a world where services are interconnected and integration is how we connect these services together. Leveraging language-based integration with AI is a natural extension to what we have today and will eventually become a common place in future integration solutions.</p><p>AI without integration is incomplete, and integration without AI is incomplete.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07b04aae907a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>