Open-weights AI models let SMEs automate legacy software locally

That legacy accounts package your finance team has used for fifteen years finally has a path to automation, and it does not require ripping out the software or paying enterprise licensing fees.

On 28 September 2026, French AI developer H released Holo 4, a family of open-weights computer-use models that can see a screen and operate it the way a human does. For operations managers and finance managers who have spent years manually processing data through software that has no API, this is worth paying attention to.

What Holo 4 actually does

Holo 4 is a family of AI models built specifically to understand graphical user interfaces, the buttons, menus, forms, and screens that make up desktop software. The main models are built on Alibaba's Qwen and released with open weights, meaning the underlying model can be downloaded and run on your own hardware.

The key shift here is that these models do not just generate text. They observe a screen, interpret what they see, and decide what to click, type, or scroll next. That is the same capability that expensive Robotic Process Automation platforms have offered for years, but delivered through models that, in compressed form, can run on a single high-end workstation GPU.

H released the models alongside open-source agent harnesses, which are the scaffolding frameworks that connect the model to a live desktop environment and let it execute multi-step tasks from start to finish.

Why legacy software has been so hard to automate

Most modern automation tools work through APIs. An API is a structured data connection that lets one piece of software talk to another without a human in the middle. Cloud accounting platforms, modern CRMs, and recent ERP systems all offer them.

But a significant portion of UK SMEs run desktop software that pre-dates this era. Older industry-specific tools, such as the job management system at a fire protection maintainer, locally installed accounting packages, manufacturing scheduling software, and similar applications were built before APIs became standard. They have no integration layer.

The only way to automate them has been through screen-scraping or RPA tools. RPA works by recording and replaying mouse clicks and keystrokes. It is effective but brittle. A minor interface update breaks the automation. And enterprise RPA platforms carry licence costs that most SMEs cannot justify for a small number of processes.

The result is that many operations managers and finance managers are still completing repetitive screen-based tasks by hand, or delegating them to team members who could be doing more valuable work.

Why this matters for operations, IT, and finance teams

For an operations manager, the immediate relevance is screen-based process ownership. If your team re-enters data between two desktop systems each week, a computer-use agent could own that task end to end. It navigates the first application, extracts the information it needs, opens the second, and enters the data without a human touching the keyboard.

For a finance manager, the value is in repetitive, rules-based tasks. Purchase order matching, invoice data entry into a locally installed accounting package, or extracting figures from a legacy reporting tool are the kinds of processes that consume hours of skilled time each week for no analytical return.

For an IT manager, the significance is data residency. Because Holo 4 runs locally on your own hardware, no screen content is sent to a US cloud provider. For businesses handling sensitive client data, financial records, or information subject to internal governance policies, that matters. You retain full control over what the model sees and processes.

The 80% of this challenge that is not technical is worth naming clearly. Most SMEs do not fail to automate because the technology does not exist. They fail because no one has mapped the process, documented the decision rules, or thought through what happens when the model encounters an unexpected screen state. The technology becoming more accessible does not remove the need for that groundwork.

Three practical steps to take this week

1. Map one screen-based process that your team completes more than three times a week. Write down every step as if you were training a new starter. Note every screen, every field, and every decision point. This document is the foundation for any automation attempt, regardless of which tool you use.

2. Ask your IT manager whether your business has a machine with a dedicated GPU of 24GB or more. You do not need to buy new hardware immediately. The question is whether the infrastructure to run a local model already exists in your organisation. This takes under thirty minutes to check.

3. Identify your single highest-volume legacy software task and estimate the weekly hours spent on it. A simple tally across the team over one week gives you a baseline. If it exceeds four hours per week, the business case for exploring automation is worth building.

The honest limitations

Holo 4 is genuinely capable, but deployment is not a matter of downloading a model and pressing run. Setting up the open-source agent harnesses requires developer-level technical knowledge. Someone needs to configure the environment, connect the model to the desktop, write the task instructions, and test the agent across enough real-world scenarios to trust it in production.

For most SMEs, that means bringing in external technical support for the setup phase, even if ongoing operation is straightforward. The models are open-weights, which means no vendor support contract exists. Check the licence on each model before commercial use: some Holo 4 weights are released for non-commercial use only. If something breaks, the path to resolution runs through developer forums and documentation rather than a helpdesk.

Accuracy is also not guaranteed on first attempt. Computer-use models can misread unusual interface states, particularly in older software with non-standard layouts. A quality-checking step in any agent workflow is not optional; it is the difference between a time-saving automation and a data entry error at scale.

Where to go next

If your organisation runs legacy desktop software and you are exploring what AI agents could realistically take off your team's plate, gecco's AI Agents combine specialist assistants and event-triggered automations into end-to-end workflows with quality gates and structured handoffs, designed around your existing setup.

For operations managers and finance managers weighing up where to start, the AI Readiness survey is built to surface exactly this kind of question. Taking the survey gives you access to 65+ free resources and a custom AI Readiness report, followed by a free 30-minute showcase to walk through the results with a member of the gecco team.

Website · LinkedIn · Case Studies · Newsletter