LegacyBridge Uses AI To Automate Data Entry On 5250 Screens

March 16, 2026 Alex Woodie

IBM i shops that are interested in eliminating manual data entry on 5250 screens may be interested in a new AI-based offering from a startup called LegacyBridge. According to the product’s developer, LegacyBridge uses computer vision and language AI technology not only to understand the information on source documents like PDFs, but also to navigate the 5250 screen and enter the pertinent data into the appropriate field.

You could say that Dmitrijs Tarasovs is an automation afficionado. “I’ve had a thing for automating boring, time-consuming tasks long before the AI boom,” he told IT Jungle in an email interview. “Finding repetitive tasks and making them disappear is what I do.”

Tarasovs had abundant opportunities to automate boring tasks while working as a full-stack engineer at Scandiweb, a Latvian Web development company that specializes in Adobe’s Magento e-commerce application. Tarasovs said he took a Magento 2 configuration task that previously took upwards of eight hours and turned it into a five-minute task for a junior developer.

Connecting Magento to Scandiweb’s proprietary systems like IBM i servers provided another opportunity for automation. Organizations often maintain their IBM i server as the system of record, and want to front-end it with e-commerce systems such as Magento. But moving data in and out of IBM i server into e-commerce platforms requires human effort.

As he explored the available AI tools, he expected to find that someone had built something to automate data entry for IBM i servers, as they had for other platforms. Plenty of companies were using computer vision AI and large language models (LLMs) to automate Web and desktop interfaces, he said, but nobody was applying it to terminal-based legacy systems. That was the kernel of opportunity that led him to found LegacyBridge.

“I was working on a document processing project and realized there were thousands of companies that still needed data entered manually into 5250 terminal screens, because there was no good way to bridge modern document formats into those workflows without a human in the middle,” Tarasovs said. “Once I started digging into the IBM i ecosystem, I realized the problem was far bigger than I expected, and that existing automation tools weren’t built to handle the complexity of real 5250 environments.”

There are existing robotic process automation (RPA) tools that use hard-coded screen coordinates to automate the movement of a mouse and keyboard on desktops, including 5250 or 3270 terminal emulators. But this technique never worked very well. And the recent revolution in computer vision and LLMs offered the potential for a brand-new take on an age-old problem.

The new generation of foundation models like Claude, Gemini, and GPT have made incredible progress in not only understanding vision and language, but combining them. Tarasovs says that computer vision is perfectly suited for solving data-entry automation in an emulator.

“Turns out, vision is extremely useful,” he said. “It lets the agent see green screens the way a human operator sees them. Reading field labels contextually, recognizing error messages, understanding screen layouts without hard-coded coordinates. And that same vision capability works for extracting data from source documents too. The two halves of the problem – reading documents and navigating terminals – turned out to use the same core technology.”

LegacyBridge learns how to navigate 5250 applications through a discovery process.

LegacyBridge has two modes. The first is extracting data from source documents, such as invoices, purchase orders, RX claims, etc. Tarasovs said this extraction utilizes “good old OCR with vision LLM for maximum accuracy.” It can also run in parallel for multiple documents.

The second mode is done through 5250. The software fires up a terminal emulation session and navigates through the green screens as part of a discovery process. The AI agent, which is based on Anthropic’s Claude, has been trained to understand how 5250 applications typically work, but every IBM i application is a little bit different. “It reads field labels, identifies where it is in a multi-screen workflow, enters data field-by-field, validates entries, and handles errors,” he said. According to Tarasovs’ YouTube video on the discovery process, it takes anywhere from 30 seconds to five minutes, depending on the complexity of the application.

When LegacyBridge is put into production and runs into something unexpected, such as confirmation screen it hasn’t seen before, an error message, or an ambiguous field — it uses the vision model to figure out what to do, Tarasovs said.

“For example, a duplicate PO confirmation prompt that appears for the first time isn’t a crash scenario – the agent reads the screen, recognizes what it’s being asked, and responds appropriately,” he wrote. “An invalid date format is fixed by reading the validation error and editing field value according to required format. It only escalates to a human operator when it genuinely can’t resolve the situation with high confidence.”

LegacyBridge’s core use case is document-to-terminal automation, but the underlying capability is broader, Tarasovs said. “The agent can navigate any 5250 workflow autonomously,” he said. “It’s not limited to data entry from documents. It can perform lookups, cross-reference data across screens, handle multi-screen workflows and execute conditional logic based on what it finds on the screen.”

LegacyBridge founder Dmitrijs Tarasovs.

LegacyBridge does not require any software to be loaded on the IBM i server. It works entirely through the terminal emulation software. (Tarasovs, frustrated at the lack of a free and open source 5250 emulator, took it upon himself to build one.) The AI model, which is currently Claude but could be any other AI model, runs in the cloud. LegacyBridge is based in Latvia, but it is offering the server component (i.e., AI model hosting) in the US, the EU, and the UK to comply with data residency requirements.

Tarasovs said it’s not important which model is being used, but what layers LegacyBridge has built around it. “We don’t send raw screen data to an LLM and hope for the best,” he said. “There’s a structured validation pipeline: field-level type checking, range validation, cross-reference verification, and confidence scoring. The LLM provides contextual understanding; the validation layers provide accuracy guarantees.”

The system also includes a complete audit and compliance layer that Tarasovs says is composed of a tamper-evident SHA-256 hash chain audit trail, role-based access control, MFA, and human-in-the-loop escalation for any situation that requires human judgment. Data at rest is encrypted with AES-256 and data in motion is encrypted with TLS 1.2 (1.3 where available).

While few IBM i shops have full-time positions for data entry clerks, it’s an activity that is still critical to getting business done. Plenty of folks who have other job responsibilities in a warehouse, a retail outlet, or a hospital are also responsible for getting data from various documents input into the IBM i host. Today, that work is done manually when the worker has the time. “LegacyBridge gives those people their time back to do the work they were hired for,” he said.

LegacyBridge is new and the company is currently working on pilot projects. Tarasovs has stress-tested the system on Holger Scherer’s PUB400 system, and he says it is ready for real-world deployments. For more information, check out the company’s website at www.legacybridge.software.