The Six Months Before AI… Or Do They Actually Exist?

We asked two data and AI leaders what comes before launch. 

This article began with a simple question: What happens inside an organisation before an AI tool goes live? 

The obvious assumption is that there’s a preparation phase. Months spent cleaning data, settling governance questions, aligning teams, and building technical foundations before a product reaches customers or employees. But does it actually work that way? We put that assumption to two experts, one in data engineering, the other in enterprise data and AI architecture.  

“I don’t see that, at least within our organisation, as one big preparation beforehand”, says Artur Yatsenko, Director of Data Engineering at Urban Sports Club, a Germany-based sports and wellness platform operating across Europe. “It’s mostly a continuous discovery of the things we can do better and the capabilities we currently have. And those capabilities could be either resources, people, or expertise.” 

“AI is not approached differently”, he adds. “It is approached the same way as the product would be approached.” 

Naveen Kanneganti is a data, analytics and AI leader who has spent more than two decades working across consulting, implementation, enterprise architecture, and global platform strategy. For him, the preparation phase is less a checklist than a leadership discipline. 

“As the MVPs (minimum viable products) run in parallel, the conversation should shift from ‘Can we build this?’ to ‘Can we operate this reliably, repeatedly and economically at scale?’”, he says. “And that’s where many organisations underestimate the real work.” 

Often, that underestimation starts with something as basic as the data itself. 

Clean Data Isn’t the Same Thing as Ready Data 

“The data should always be accurate, because we don’t make the data accurate for the sake of AI”, Yatsenko says. “It has to be accurate for the sake of making decisions and taking the business forward. The data being accurate for AI is the bonus.” 

Urban Sports Club found out how easily that principle gets tested. The team built a conversational agent to answer questions based on company data. The problem: “We figured out that the data was not fully ready because it couldn’t understand some of the relationships in the data. We basically needed to recreate quite a lot of datasets before including them in the agent”, Yatsenko says. 

For him, the answer was ownership. Put the people who understand the data in charge of it. “If you’re owning the data end to end, this is the best scenario.” Teams move faster, he says, when they control how the data is defined, produced, and maintained, rather than depending on another team to correct underlying problems. 

“Having data is not the same as having AI-ready data”, Kanneganti says. In his view, AI-ready data is fresh, discoverable, accessible, trusted and structured in a way that reflects how the business works. 

That difference matters because AI doesn’t fill in gaps the way a human analyst can. A person may be able to interpret messy metadata or ambiguous terms. An AI system needs the relationship between customers, assets, and policies spelled out clearly enough for a machine to read. 

“That is where metadata, semantics, ontologies, and knowledge graphs become important”, Kanneganti explains. “They help AI understand relationships and constraints, not just retrieve isolated pieces of information. Without that context, AI may produce an answer that looks plausible but still recommend – or trigger – the wrong action.” 

A June 2026 industry position paper from Germany’s KI Bundesverband presents knowledge graphs as one way to give AI systems more explicit information about entities and their relationships. 

Governance Can’t Be a Last-minute Meeting 

Which data can we use? What rules apply? Where is human judgement required? Yatsenko is against leaving those questions until the end. “Those questions are better to be asked earlier in order to not have surprises in the last-mile stage.”  

Kanneganti says AI preparation works best when organisations move from sequential handovers to joint decision-making. “Engineering, data, security, and business teams need to shape value, risk, and feasibility together from the beginning – not review them separately at the end”, he argues. 

Being informed matters here as much as being involved, according to Yatsenko. “Everybody should be on board. They don’t necessarily need to make the decisions, but they definitely need to be informed.” 

Even when everyone’s informed, Kanneganti says some organisations still make one common mistake: confusing access with readiness. “Giving people tools is easy; making AI reliable, governed, cost-aware, and useful in real business workflows is the harder part.” 

Finished Doesn’t Mean Done 

None of that groundwork guarantees the rest is simple. “You can build things rather fast but integrating them into the main product is challenging”, Yatsenko explains. “The gap between the MVP and production is still very high, and it’s high for pretty much every organisation.”  

A recommendation algorithm might work perfectly behind the scenes and still stall at the point of going live on the homepage. At that stage, design, maintenance, and release planning all come into play. 

“That last 20% of the work often takes longer than expected”, Yatsenko says, because that’s where the smallest details have to be right: legal reviews, compliance checks and questions about what data was used and how the model was trained. 

And the Bill? 

That’s the other blind spot. “AI cost is not only token consumption”, Kanneganti says. “It includes the wider operating ecosystem: compute, data movement, orchestration, monitoring, integration, security, testing, and human review.” 

Much of that is easy to overlook during a prototype. But it becomes harder to ignore at scale. “If those costs are not visible early, organisations risk treating AI as a cheap experiment and only discovering the real economics when usage scales”, he explains. 

So Does the Six-month Runway Actually Exist? 

Not really, not as a distinct phase anyone could point to on a calendar. Yatsenko rejects the idea of a separate preparation period. The work is real, he says, but it doesn’t happen in a clearly defined sequence. Kanneganti is slightly less absolute, arguing that some of it still has to happen upfront, whether or not anyone calls it a phase. 

Either way, the work doesn’t stop once the product is live. “There is no particular readiness”, Yatsenko says. “It doesn’t really work like that.” 

“Think big but start small”, he suggests. “Take the MVP approach in order to validate the idea and then get more traction within the organisation for more resources and more dedication to that particular problem.” 

Kanneganti sees readiness as an ongoing discipline. “Leaders should ask not only whether AI can do something, but whether it can be operated sustainably, securely and economically at scale,” he says. “The organisations that lead will not simply be those running the most pilots, but those that combine ambition with discipline.” 

Editor’s note: Naveen Kanneganti contributed to this article in a personal capacity. His comments draw on 20 years of experience in data, analytics and AI and do not represent the views of his employer.