AI features can be added to legacy systems through APIs, middleware, gradual migration, or a hybrid setup without rewriting the whole system.
2
Hakan Tek describes a textile system that scanned legacy Excel files, predicted fabric defects and production delays, and triggered alerts without changing the existing workflow.
3
Privacy and security may require private servers, restricted network access, middleware, firewalls, and carefully limited protocols rather than unrestricted endpoints.
Summary
Hakan Tek explains how companies can add AI to older systems that remain important to daily operations. He describes the main obstacles as outdated technology, rigid data formats, security risks, and resistance from staff and leadership. His proposed starting points are API integrations, middleware that translates between systems, incremental replacement of components, and hybrid deployments that keep sensitive or latency-sensitive services on premises while using cloud models elsewhere. Examples include fraud detection on transaction systems, predictive maintenance for factory equipment, healthcare insights connected to legacy records, and a textile application that scanned Excel files for defects and production delays. Tek also discusses low-cost services for language models, search, summaries, and messaging. In the question period, he is direct that third-party APIs cannot provide complete certainty about data security. He recommends private infrastructure or servers in jurisdictions with strict rules, along with network isolation and protected interfaces.
Legacy systems still carry the business, so replacement is often unrealistic
Hakan Tek describes legacy systems as older systems that companies still rely on, even when their technology is difficult to work with. They may have outdated technology, few APIs, and rigid data formats. Replacing them can disrupt the business. He compares them to an old office coffee machine that nobody wants to touch because it has worked for years. His argument is that AI integration should fit around these constraints instead of assuming a clean rebuild is possible.
The main barriers include data formats, security, and people
Tek says older systems often store structured data in rigid formats, while AI applications may use unstructured or semi-structured data. Connecting the two can require messy translation work. AI systems that call external servers or make autonomous decisions also create a new attack surface. He says resistance to change is often the largest obstacle. Workflow changes, staff training, and fears about job replacement can prevent a technically workable project from getting started.
APIs and middleware provide practical entry points
Tek presents API-based integration as the easiest place to start when a legacy system already exposes APIs or can be wrapped to expose them. An AI model can then handle a specific task through a stable, secure interface. Middleware can translate data formats, route requests, and manage communication between the old system and AI tools. He also mentions gradual migration, where individual components are replaced over time instead of rewriting the whole application at once.
Hybrid deployments keep some workloads on premises
Tek prefers a hybrid approach when a company needs to keep services on premises because of latency or privacy requirements, while still using cloud-based AI models where appropriate. He says the right choice depends on the existing architecture, the team's capacity, and the business goal. In his explanation of agent and model protocols, he distinguishes coordination between agents, model deployment and querying, and the exchange of structured messages.
AI can add value to financial, industrial, healthcare, and textile systems
Tek gives several examples of AI layered onto existing systems. Financial services can add fraud-pattern detection to transaction systems without replacing them. Manufacturers can use predictive maintenance with existing machinery-monitoring platforms to identify likely failures. Healthcare records can connect to AI diagnostic tools while the core record system stays in place. In a textile project, his team scanned legacy Excel files to predict fabric defects and production delays, then triggered alerts without disrupting the company's workflow.
Small teams can test integrations with inexpensive external services
Tek argues that an AI project does not always need a large infrastructure budget. He mentions Groq for fast language-model access, SerpAPI for real-time Google search results, a service that summarizes web pages with language models, and Twilio for SMS, WhatsApp, or voice features. He says these services can often connect to legacy applications through API calls, scripts, or Excel macros. He also notes that Excel now supports TypeScript alongside its earlier Visual Basic support.
During the discussion, Tek says a company cannot be completely certain about data security when it sends data to a third-party service. His cleanest option is to run AI on the company's own servers, although he acknowledges that the required hardware may be difficult to obtain and that Germany has strict legal requirements. He also suggests using a server provider with suitable rules and protections as an intermediate option.
Developers should restrict interfaces and protect against injection
When asked about security features, Tek recommends private servers without outside connections for sensitive systems. He says developers still need to protect their servers against injection attacks. Middleware and firewalls can help, and carefully chosen protocols can limit access to approved operations instead of exposing every endpoint. His advice is to start with a small integration, test it quickly, and expand only after the approach works.
"Instead of replacing entire system, we built a solution that scan their legacy Excel files and triggers alerts enabling the company to take proactive action without distributing their existing workflow."08:36
Who should watch
You maintain an older enterprise application and need to add AI without committing to a full rewrite.
Your team is deciding between API calls, private infrastructure, or a hybrid deployment for an AI feature.
You need a practical starting point for connecting AI to Excel files, transaction systems, machinery monitoring, or other existing workflows.
