Businesses are spending months choosing models, copilots and platforms. Dariel’s Sasha Slankamenac says the bigger risk sits in the data underneath, which was never built to be read by a machine that asks its own questions.
Walk into most AI steering committees and the same questions come up. Which model, which vendor, which copilot licence, which platform to standardise on. They are fair questions, but Sasha Slankamenac, Architect in the Office of the CTO and Practice Lead: AI at Dariel, thinks they take up more energy than they deserve.
“I see organisations spend months comparing models when what decides whether the project works is the data underneath,” says Slankamenac. “A better model on bad data just gives you a more fluent, more convincing answer.”
In a February 2025 Q&A, Gartner predicted that through 2026 organisations would abandon 60% of AI projects that lack AI-ready data. The same research found that 63% of organisations either lacked the right data management practices for AI or were unsure whether they had them. With 2026 in its final quarter, that forecast is close to being tested.
For AI to be useful, Slankamenac says, the data feeding it has to be accessible, accurate enough for the decision at hand, governed and consistently defined. Definitions are where many roadmaps first come apart.
Take an insurer. The policy administration system counts a policy as active from inception. Finance counts it once the first premium clears. The claims platform has its own rules for lapses and reinstatements. Each definition made sense to the team that built it, and for years people who knew which report to trust managed the gaps. Ask a copilot connected to all three how many active policies the business has and it will give you a confident answer that may well be wrong.
“If ‘customer’ or ‘revenue’ or ‘active policy’ means three different things across your systems, you have a data maturity problem that has been filed under AI strategy,” says Slankamenac.
Old data problems, new behaviour
The problems themselves are familiar. What changes is what happens to them once an AI system, rather than a person reading a report, consumes the data. A report is designed. Someone chose the tables, the joins and the filters. An AI agent or copilot queries on the fly, pulls from sources that were never meant to sit side by side, and interprets what it finds.
“Nobody sat down and designed the path the agent just took through your data,” he says. “It crossed boundaries no report ever crossed, so what the data means, who is allowed to see it and how current it is all matter far more.”
Dashboards are no longer the only way in
For a long time, the dashboard was how business users met enterprise data. It also worked as a buffer, with definitions settled and edge cases handled before anyone saw a number. Conversational analytics removes much of that buffer.
Slankamenac expects dashboards to keep their place for tracking known metrics. Meanwhile, more questions will be asked in a chat window, and more actions will be taken by agents working inside business processes.
“Analytics is becoming something that sits inside the work you’re already doing,” he says. “That’s a real improvement for users, and it also puts the mess the dashboard used to hide much closer to the surface.”
Governance has to cover what AI can combine
Traditional data governance was largely about access: who can open this database, run this report or log into this application. It assumed a person was doing the retrieving and the interpreting.
A copilot can retrieve, combine and summarise across many sources on someone’s behalf, and the result may reveal something no single source would have exposed. In South Africa, where POPIA governs how personal information is processed, that deserves attention well before a copilot goes live.
A 2024 Gartner survey of 132 IT leaders, reported by Computerworld, found that concerns about data oversharing led 40% of respondents to delay their Microsoft 365 Copilot rollouts by at least three months. Nearly two-thirds said information governance and security risks took significant time and resources to deal with. “The question is now what the AI can assemble, infer and act on for this person, and whether we can explain where the answer came from,” says Slankamenac. “If you can’t trace an answer back to its sources, you can’t stand behind it.”
Where to start
Slankamenac does not suggest shelving AI plans until the data estate is perfect. He recommends starting from the decisions the business wants AI to support and fixing the data those decisions depend on.
In practice, that means agreeing definitions for the handful of business terms that matter most and recording them somewhere a machine can use them. It means knowing where key data comes from and how fresh it is. It also means extending access controls to cover what an AI can combine, as well as what a user can open.
“Choosing a model is the easy part, and it changes every few months anyway,” he says. “The organisations that get value from AI will have done the unglamorous work on their data first.”
