KEHA Centre's Mylly platform was already mature when the Agile Data Agent pilot started. The foundation was solid: Data Vault methodology, Agile Data Engine, a governed delivery model built over several years of public-sector data work. The question was whether agentic development could improve how the work on top of that foundation gets done, without compromising what made it trustworthy in the first place.
KEHA is a national organization whose data supports employment statistics, political decision-making, and information needs across municipalities, EVK Centres (former ELY Centres), and central government. When that data is late, inconsistent, or wrong, the people making decisions across Finnish employment services are working with less than they need. The governance bar here is not a formality. It is the operating condition.
That context matters for what the Agile Data Agent pilot actually shows. This is not a story about a team that just moved fast. It is a story about a team that moved with awareness and intention, in a demanding environment, and still got meaningful results.
We spoke to two key people in the project: Teemu Väisänen, Data Warehouse Architect at KEHA Centre, who described the pilot from the platform side, and Miima Kuusela, Data Architect at Solita, who has worked on the Mylly platform for three years and was involved from the early pilot phase through everyday delivery work.
In Mylly, a lot of the routine work had already been standardized into reusable patterns. Staging, standard load patterns, ADE package handling — solved. What stayed manual and expensive was Data Vault modeling: hubs, links, satellites, the business rules on top of them. Generic AI tools had been tried, and while they are useful in their own context, for Data Vault work in a specific ADE environment they are too general to trust with the packages.
ADA is built for that work. It operates on ADE packages directly and generates metadata rather than raw code, which means the output lands inside an already governed workflow rather than outside it. That is why the pilot made sense: ADA fit the environment instead of requiring the environment to adapt around it.
Data Vault work that previously took around ten days was coming down to one or two days. Teemu described the shift directly: "Going from ten days of work to one or two days — that is a massive change." Miima's experience from the delivery side puts a sharper point on it. In one concrete case, the staging layer and a first version of the Data Vault layer for a new source system were built in a single day. As she emphasized, this is not the outcome every time, but it illustrates the kind of impact the approach is already capable of.
When the structural foundation arrives in a day, the team's attention moves to the parts that were always the real work: business logic, fact and dimension design, the publish layer, the reporting stakeholders consume. The time saving is real, but what it frees up is even more interesting than the saving itself.
Rather than building the first structural draft from nothing, developers could start from a generated foundation and move directly into review and refinement. The base quality was strong enough to work from. That changes the nature of the first hour of a task more than it changes the total time on a project.
But the gain does not stop at the first draft, as iteration becomes considerably lighter too. Once the basic structure exists, refining the model and steering the output is far easier than doing the same work by hand. The productivity effect compounds through the whole implementation, not just at the start.
Miima's review process reflects that shift. Once a Data Vault layer is generated, she checks whether the essential building blocks are in place. "And they usually are," she said.
Junior developers benefited particularly. Custom skills steer the agent toward a team's conventions, which means the output reflects an established way of working. Consistency holds whether the work is done by internal staff, external consultants, or someone early in their career.
Earlier, a junior might spend hours or days manually entering Data Vault content into ADE as a way of learning the basics. Now, the team can walk through the stage layer together and the junior can start building the Data Vault alongside ADA from the beginning. "It can grow developers' capabilities much faster," Miima said. "If you are good at explaining the problem and telling ADA what you want to do, you can create brilliant solutions with it."
What has struck those closest to the pilot is not a single dramatic turning point but something more durable. A continuous, compounding improvement in how much the tool can handle and how much it shortens the work. ADA has started to become part of the team's everyday rhythm, and its absence would now be felt.
A person still needs to read the output before it goes anywhere. Teemu was clear: "A human still has to be in between and read the end product before pushing it forward."
That sounds straightforward. But people learn to prompt faster than they learn to question what the prompt produced. If you have never written the Data Vault yourself, can you actually challenge what the agent handed you? The scarce skill is not just prompting, but also reading generated output critically.
Miima noted that ADA still does not understand everything, particularly when it cannot inspect the actual data. The output rarely looks obviously wrong — the risk is that it misses team-specific or data-specific nuances that only show up under careful review. The team's response is worth paying attention to: rather than correcting individual outputs repeatedly, they improved the shared setup around ADA so the same gaps would not recur. The agent is only as good as the context it receives. When the context is right, the results follow.
Moving from exploration to actual production use in KEHA's environment required clearing real questions: what can the tool see, where can data go, who approved what, how is generated output reviewed. Those are not one-time setup tasks in public-sector data work. They are part of how work runs.
ADA could be brought in under the same approval model KEHA had already used for other AI tools, because it does not access data differently than tools they had already cleared. That made adoption more practical than it might have been.
The next phase — giving the agent richer context about what tables and views actually contain, what the data means, and what rules apply — is a larger task. Changing legislation and shifting source systems make it a moving target. But it is also the most important work ahead. An agent that understands the environment it is operating in produces better output, requires less correction, and can be trusted with more. Building that context layer is not a technical overhead. It is what turns a governed data platform into something an agent can genuinely work with. The discipline Mylly was built on is what makes that next step possible.
Teemu's advice for anyone considering a similar move: just start prompting and start with something small and well-scoped. Validation is easier on a narrow entity than a broad one, and you learn what the agent can do without betting a delivery on it.
From the delivery side, Miima's advice starts one step earlier: before running a single prompt, build the shared repository. Common practices, custom skills, and team-specific instructions are what separate consistent results from one-off wins. Once that foundation is in place, the tool scales naturally across the team.
She sees ADA's role continuing to expand with deeper integration with DevOps workflows, specification meetings, and more automation of the handoffs between conversation and delivery. "ADA will be at the center of everything we do," she said. "It's absolutely fantastic that AI came into ADE in exactly this way. It changes the whole nature of the work."
For the people receiving the data Mylly delivers — the municipalities, the employment regions, the authorities making decisions on current numbers — faster and more governed delivery means better information arriving more reliably. The platform was already trusted. ADA is making it faster to build on top of that trust.
ADA didn't replace the discipline Mylly was built on. It's the reason the discipline now pays off faster.