Yuki
Making a complex migration viable with AI
Executive Summary
The challenge
To migrate a large, complex bookkeeping platform where the scale of the system and the difficulty of testing made modernisation slow, costly and high risk.
What we delivered
A simpler migration approach, supported by AI-enabled workflows that improved testing, reduced manual effort and created a repeatable path to modernisation.
The duration
12 weeks
The outcomes
We helped turn a difficult migration into a viable programme, reduced delivery risk and left the team with reusable approaches they could apply at scale.
Key results
The full story
Yuki set out to modernise a complex bookkeeping platform, but the scale of the system and the difficulty of testing its behaviour made migration highly challenging. By simplifying how the application could be migrated and introducing repeatable AI‑enabled workflows, we helped unlock a viable, scalable path to modernisation.
Yuki provides a highly automated, cloud-based accounting platform to businesses and financial professionals. The company was modernising its core application, but this was a daunting challenge. Over time, the platform had grown in both scale and complexity, while knowledge of its behaviour had become harder to maintain, making it increasingly difficult to test and safely evolve. In this context, migrating even a single page carried significant risk, and the modernisation process was looking prohibitively time-consuming and costly.
Scott Logic was brought in to help Yuki use agentic AI to tackle this migration challenge. By following our highly collaborative approach to Agentic Engineering Enablement, we helped redesign how the system was tested and validated, using AI agents to understand behaviour, structure the work, and embed repeatable workflows. This unlocked the migration and gave the Yuki team confidence to accelerate progress, while also significantly reducing the effort required for future work. As a result, the programme was reforecast from 15 months to nine, delivering faster outcomes and an estimated €380k reduction in cost.
Scott Logic didn’t just help us make a complex migration viable. They worked closely with our team to develop the skills, practices and confidence to continue independently. They left us with a trained team, a proven approach and a clear path to scaling it across the business.
Ricardo van Marion, Software Development Manager, Visma Yuki
Assessing challenges and building foundations
Our Agentic Engineering Enablement process begins with an Assess phase, which examines the entire software development lifecycle to identify where AI can add value and where constraints may limit its impact. Our workshops in this phase were critical in diagnosing the core technical and delivery challenges the migration faced, and in identifying where to focus our efforts.
Through these discussions, we gained a clear snapshot of the system and pinpointed a back-office entry page that offered the ideal proving ground for the new approach. The page was of high value to Yuki but also extremely complex. Demonstrating that agentic AI could migrate the page safely would instil confidence and unblock the migration’s progress.
As the first step towards this, we needed to enable agentic AI to work effectively within the constrained technical environment. That became the focus of the Foundation phase. To unlock real value, AI agents need to run ‘agentic loops’: building and running the system in their own environment, testing and refining changes against the existing system, and iterating autonomously.
In the context of Yuki’s system, this was very challenging for two reasons. First of all, the size and complexity of the codebase made it difficult for the AI to locate and work on the right parts of the system. Second, the team had to make AI effective in a constrained, enterprise-scale legacy environment, where it could not simply run the full system locally.
To tackle the first challenge, the team refined how the AI indexed and accessed the system, and provided more focused guidance to enable the AI to navigate the code reliably. The second challenge required a combination of solutions. The team packaged key supporting components into self-contained environments, allowing the AI to run and test changes independently of the live system. Alongside this, they gave the AI access to the test environment of the live system by making it an authorised user with login credentials. This allowed the AI to observe how the system behaved, providing a benchmark against which it could build and test the new version.
From experimentation to repeatable delivery
This combination of controlled environments and real-world reference points gave the team a reliable way to explore different approaches.
Early in the Accelerate phase, they began experimenting with how best to migrate the back-office entry page, testing several approaches to translating it into the new architecture. Some of these focused on translating larger sections of the page in one pass, while others attempted to create comprehensive test coverage up front. While these ideas were promising, they proved difficult to make reliable in practice.
The turning point came when the team approached the problem in a different way. Instead of trying to migrate the highly complex page in one go, they broke it down into smaller components or ‘panels’. They created a simple skeleton of the page, with placeholders for each element, allowing individual panels to be built and tested in isolation. Each of these panels became a clearly defined task with a limited scope, which made errors easier to identify and reduced the risk of unintended side effects.
To guide how each panel should be migrated, the team introduced structured workflows or ‘playbooks’. These set out the sequence of steps the AI should follow, from analysing the behaviour of the existing system to building, testing and refining the new version. Each task was accompanied by a running checklist that captured progress and outstanding work. This gave the AI a kind of “working memory” allowing it to return to incomplete tasks and handle interactions between panels without losing context.
Agentic loops could then be enabled, allowing the AI to work in cycles of building, testing and iterating, with each iteration bringing the new system closer to the expected behaviour.
Alongside this, the team strengthened how work was tested and validated. With the panel approach having broken up the development and testing into smaller units, the team focused on improving test coverage. They introduced example-based testing, which defined realistic scenarios alongside core functionality to capture both expected and edge-case behaviour.
They also integrated specialist testing and UI tools, using Model Context Protocol (MCP) connectors to enable the AI to draw on design and testing expertise as it worked. In combination, these improvements made validation more systematic and reliable, ensuring that the AI’s output could be trusted as the work scaled.
Embedding the transformation
The first three phases of our Agentic Engineering Enablement approach had demonstrated a trusted and viable approach to the migration. In the Scale & Sustain phase, attention turned to embedding these new ways of working within the Yuki team. The structured workflows developed during Accelerate became the foundations of a repeatable process, one which the team would be able to apply independently and at scale after we left.
The Scott Logic and Yuki teams worked closely together throughout, pairing on tasks and learning how to apply the new approach in practice. This ensured that knowledge was grown through delivery, not transferred in isolation.
One confidence-building task we encouraged the Yuki team to do was to re-migrate pages that had already been translated manually. This gave them a known benchmark, helping them to understand how to guide the AI effectively and assess whether the outputs matched the expected standard.
The playbooks the team developed together during Accelerate would provide a durable reference point for the rest of the migration. In combination with the supporting documentation that captured how tasks should be approached, these artefacts would enable the Yuki team to apply the same methods consistently as the work scaled across new areas of the system.
The team also developed a more mature approach to tooling over the course of our collaboration. While early stages of the project had relied heavily on Devin, the team recognised that not every task required the capabilities of such a top-tier tool. Larger pieces of work benefited from Devin’s ability to build, test and iterate autonomously, but smaller changes and test creation could be handled just as effectively with lighter-weight tools.
This more nuanced tooling approach allowed the team to balance performance with cost, ensuring that the approach remained sustainable as it scaled.
Accelerating delivery while building lasting capability
The transformation delivered measurable improvements across both delivery speed and software quality. The team achieved 58% faster lead times and reduced bug resolution time by 60%, compared to other teams working on the same codebase. What had been estimated as a 15-month migration programme was reforecast to be completed in just nine months.
Importantly, this acceleration did not come at the expense of quality. This was thanks to the time the team had invested in setting up the right foundations and practices, which included creating structured playbooks, organising the application into manageable parts, and enabling continuous comparison with the live system. The result was that the team produced 37% fewer defects per unit of work.
The most transformative outcome was the change in how the team worked – and how this change propagated throughout the business. The combination of structured workflows, AI tooling and continuous validation established a repeatable approach that could be applied to the rest of the migration. As the team’s confidence grew in the new ways of working, they shared with their colleagues the new methods and the rationale behind them.
By the end of the engagement, the Yuki team owned its processes, had documented its approach, and could complete the migration without external support. A high-risk, multi-year programme had been transformed into a viable and accelerating delivery effort, placing Yuki ahead of the curve in adopting AI-enabled software engineering.
Ready to accelerate software delivery with AI?
Get in touch to book a two-day Assess workshop
