Integration of AI into Traditional Systems // Hakan Tek // Agents in Production 2025

Hakan Tek [00:00:00]: Today I show you how can we integrate into traditional system, even Lego legacy ones without starting from scratch. Who am I and my name is Hakan and I have been working as postech developer for over decades now and mostly focusing on web technologies, enterprise applications and system integrations. Over the last few years I've been special involved in helping companies modernize their legacy systems. Not by throwing everything away and starting from scratch, but by ending AI powered features right into their existing infrastructure. I am currently based in Germany and I collaborate with the business of various sites on digital transformation and AI adaptation projects. Today I want to some of lessons I have learned while working on practical low distribution ways to bring AI into traditional systems. Before we jump into how AI can be integrated, let's quickly define a few terms to make sure we are all on the same page. First, what's the AI? Simply put it when machines do things that normally require human intelligence like understanding language, making predictions or even recognize patterns without needing a coffee break.

Hakan Tek [00:01:47]: And what is legacy system? These are older, often mission critical systems that companies still rely on. They are not always pretty, but they run on business. Think of them, that old confirmation in the office, nobody want to touch it, but it's been working for 15 years and everyone secretly afraid to replace it. So why should we integrate AI into these systems? Because business need to be faster, smarter and more competitive. AI can automate, automate, replace task, make better decision using existing data and real value without needing a fully system overhaul. And best practice, you don't need to throw out all confirmation, you just need to teach it how to make smarter coffee and more effective coffee. And challenge of these integrations? First, legacy constraint. Many of these older systems weren't built with AI mind.

Hakan Tek [00:03:02]: They have outdated tech stacks, lack APIs and upgrading them isn't always an option. Then there is data compatibility. AI thrives on data, but legacy system often stores structured data in rigid formats. AI on the other hand often works with unstructured semi structural data. Bridging that gap can be messy and security is another concern. When you introduce AI, especially models that connect to external servers to make autonomous decisions, you are also potential opening up new attack surface that needs to be addressed carefully. And finally, resistance to change. Let's to be honest, the biggest blocker is often not the tech but the people changing workflow, training stuff or even just convincing leadership that AI won't replace everyone.

Hakan Tek [00:04:06]: That's real challenge. So while AI offers powerful opportunity, these are the reason why companies often hesitate or fail to get started. Now we know the challenges and let's talk about how we can actually integrate AI in the traditional system without ripping and everything out and starting over first. API based Integration this is usually the lowest hanging fruit. If your legacy system exposes APIs or can wrap up to expose them, you can connect AI models that handle specific tasks. Whether you are using agent to agent model control protocols or agent communication protocol, the key is establish a stable and security interface and second middleware platforms. These acts translators between old and new middleware can standardize data formats and handle routing and manage communication between your legacy system and AI tools. And third, gradual migration.

Hakan Tek [00:05:22]: Instead of rewriting everything at once, you can incrementally replace authored components with AI enabled modules. This reduces risk and lets you test what works and what doesn't before going all in. Finally, it's my favorite hybrid approach. This is especially useful when you need keep some services on permits due to latency or privacy requirements, but still want to leverage powerful cloud based AI models. You get the best of both worlds. In the end, the best strategy depends on your current architecture, team capacity and business goal. What are the most common effective starting points here? These are the standard protocols when you integrate AI. First AI to AI agent Agent is about enabling autonomous agents to communicate and coordinate.

Hakan Tek [00:06:28]: For example, one AI model might process data and send the result to another agent and handle alerts or decision making. An MCP model control protocol is a way to standardize how different models are deployed, queried and updated and useful when you have multiple AI models being orchestrated alongside your traditional stack and agent communication protocol. Focus on how agents exchange structured message it's more about semantic and meaning of communication rather than just sending data back on board. And let's look at some real world examples where AI has been successfully integrated into legal systems and in financial service AI models have been added on top existence transaction systems to detect Freud patterns in real time. This avoids costly system revives but still enhanced security and manufacturing companies use AI driven product estimation maintenance on top of their existing machinery monitoring platforms. This helps predict failures before they happen and it's nice to reduce the downtime and in healthcare legacy hospital record system have been connected with AI diagnostic tools helping Docstar with faster and more accurate insight all while keeping the core system inactive and form of my own experience. We develop a system for a textile company that predicts fabric defects and potential production delays. But instead of replacing entire system we built a solution that scan their legacy Excel files and triggers alerts enabling the the company to take proactive action without distributing their existing workflow.

Hakan Tek [00:08:47]: And these examples show that with smart integration strategy, AI can bring significant business value without starting from scratch. And often companies aspired to start AI project because of budget constraints, lack of clear direction or fear that they need a huge infrastructure approach. But in reality, not all AI invites require massive investment. There are many tools out there with free trials that let developers and small teams start building and testing today without a problem. From financial months planning and here you see my popular third party APIs and GROQ offers fast LLM interface. It's great when you need fast response and can't afford host your own models and you can select a lot of different model as I know it supports Also ChatGPT or IT support deep seq R1 and with a limit it's free. You can easily start to develop your project and you can present it. Then you can directly book API key and SERP API gives you access to real time Google search results through an API.

Hakan Tek [00:10:27]: It's perfect for an AI tool that needs to be to look things up dynamically and before in hackathon I develop a market research API with this service so I mean normally reaching this kind of API is so costly but I directly without any cost I develop my application and this tail goes a step further and summarize web page using LLMs and I think it's ideal for AI agents or knowledge extraction tools and Twilio is a classical for adding communication features like SMS alerts or WhatsApp post or aid driven voice assistants without managing your own servers. And the best part you can integrate most of these with legacy systems using simple API calls scripts or even Excel macros in some cases and on last update I find an Excel macro before they develop these Microsoft Excel Visual Basic language but now they start to support TypeScript language so I think it's so nice experience for a developer so and as a result I can say AI and legacy system are not enemies and in fact they can work together very well if you approach the integration through truthfully. You don't need to fully rewrite to see real business value from AI whether it is Excel sheet ERP system or on prom database. There are smart log distribution ways to make them intelligent and start with APIs, use middleware if needed and don't under under under timemate the power of free tools like Grox, SERP API or the others and test your ideas quickly. In the end the hardest part is often just getting started. So instead of waiting for the perfect time or budget just kick off something small because even simple AI layer can have a huge impact on your Existing systems. And thank you for your time and if you have question, please feel free to type me on chats. Awesome.

Skylar Payne [00:13:22]: Is there a place where folks should connect with you if they have like other questions they want to follow up with?

Hakan Tek [00:13:29]: You can follow me on my LinkedIn. Should I write on the chat?

Skylar Payne [00:13:36]: Sure, yeah. Yeah. Put it in the chat and then I'll.

Hakan Tek [00:13:39]: Okay. Here.

Skylar Payne [00:13:43]: You know one thing I was interested in like at hearing your talk. Since the focus is really on like legacy systems like old enterprise. Just curious about to dig a little bit deeper into the point around resistance to change. One thing I feel like I've seen a lot with companies is almost a concern about like data privacy and so they don't want to use the APIs. Do you feel like you see that a lot and if so like what do you do to get around that?

Hakan Tek [00:14:17]: Or like I mean the. The clear solution is actually if you use third party service you can't be. You can never be sure about your data security. Not hundred percent. But in this case you can establish it in your own server. I am not sure if now it is easy to buy this kind of systems like these special Nvidia drivers. But this is the only one cleanest solution. But in Germany also it is so big challenge because there are strict legal rules about this.

Skylar Payne [00:15:03]: Yeah.

Hakan Tek [00:15:03]: So the. The maybe the other tools. I think what I believe is booking a server in a good quality company like in for example Germany there are some servers you can book it with strict rules. So maybe it can be a solution but in long term. Yeah. It's better to have your own Nvidia tools. Totally.

Skylar Payne [00:15:38]: Where did a question in the chat come in? From Thomas at Duplica. How can I get started into all the security features needed to add when implementing AI automation tools to these types.

Hakan Tek [00:15:50]: Of services.

Skylar Payne [00:15:53]: Resources to point.

Hakan Tek [00:15:55]: I mean actually the biggest challenge is your. I mean if you mean the legacy systems I mean you can close your network and. And have your own AI servers. So but apart from the. I mean actually it's a part of more security department so I'm. It's not directly my profession. But the best way is how in the private company having their own private servers without connection. So but in as a developer what can we do? We can.

Hakan Tek [00:16:45]: We have to protect our servers against injections. So and with correct middleware and firework we can solve it. So maybe the other point can be if we use the correct protocols. Like maybe it's better if I direct to show and if we implement some correct protocols. I think it can be also a solution. So instead of opening each endpoints, we can allow only some special protocols, so. Yeah, but I mean, as a developer, I can say the clear points. Mm, totally.

Hakan Tek [00:17:38]: Well, cool.

Skylar Payne [00:17:39]: We appreciate you coming by. Let's check the chat one more time. Yeah, I didn't see other questions, but yeah, we really appreciate you coming by and sharing your work with us and hope to see you again. Remember, connect with him on LinkedIn. His info's in the chat. And with that, we're entering a short break. Thanks, everyone.

Hakan Tek [00:18:01]: Thank you so much.