Every scale up wants AI analytics now, and I can understand why. The promise is very attractive. A CEO asks a question in plain English and gets an answer in seconds. A CMO can investigate performance directly. A CPO can explore product behaviour without waiting for an analyst. A CFO can test a hypothesis without opening a reporting request. After years of companies talking about self-service analytics, AI finally makes the experience feel genuinely self-service.
I think that direction is real. Conversational analytics will become normal, executives will interact with data more directly, and a lot of routine analysis will move away from dashboards and towards natural language. Analysts will spend less time answering repetitive questions and more time on work that requires judgement and deeper business context. Where I think many scale ups are getting this wrong is in the sequence. They are trying to jump straight to the AI layer while the foundations underneath are still incomplete, inconsistent or simply messy.
I have seen this pattern many times. A company grows quickly and the data setup grows with it, usually in a fairly organic way. Finance develops one view of the business, Sales another, Product has its own definitions, and Marketing builds another reporting layer again. Over time the Data team ends up maintaining a large collection of dashboards, models, SQL logic, spreadsheets and undocumented knowledge. Eventually someone looks at all of this complexity and asks a very reasonable question: why not put AI on top of the warehouse and let everyone ask questions directly?
The difficulty is that AI makes access easier without resolving the inconsistencies underneath. If the company has three definitions of ARR, the model still has three definitions to choose from. If customer identity is broken across several systems, the model still has to work with that broken mapping. If Product calls someone an active user after one event and Finance applies another rule, the model needs to know which definition matters. If half of the business logic lives inside old dashboards and the rest lives in the heads of a few experienced analysts, the model has no automatic way to recover that context.
This matters because AI can return a polished answer even when the underlying logic is weak. In some ways, that makes the problem more dangerous. A messy dashboard at least looks messy. A fluent answer sounds authoritative. People are naturally more likely to trust something that arrives quickly, clearly and confidently, especially when it comes from a system they already associate with intelligence.
One of the most important things companies underestimate is how much interpretation happens today between a business question and the final answer. Imagine a CEO asking, "What was churn for enterprise customers last quarter?" An experienced analyst may know that the enterprise segment changed definition six months ago. They may know Finance excludes a certain class of accounts. They may know one historical table has a known issue. They may also know that the CEO probably means logo churn rather than revenue churn. None of that may be properly documented. It is simply part of the analyst's understanding of how the company works.
That analyst is effectively carrying a lot of the company's semantic model in their head. Once AI becomes the interface, that knowledge needs to exist somewhere explicit. The system needs to understand what the company means by customer, account, subscription, revenue, churn, conversion, active user, market, segment and product. It also needs to understand how those concepts relate to each other and which definitions are official. Without that structure, the model is being asked to reconstruct business meaning from technical tables and fields that were never designed for that purpose.
I have seen this very clearly in practice. In one scale up I worked with, a key revenue metric had several different versions depending on which team you asked. None of them were absurd. One used a particular FX treatment, another reflected how the commercial team viewed performance, and another was closer to the Finance definition. The company had plenty of data and plenty of reporting. What it lacked was one clearly owned definition that everyone accepted as the official view.
In that environment, an AI assistant can absolutely answer the question, "What was ARR last month?" It can retrieve a number quickly and explain it beautifully. The difficulty is deciding which version of ARR it should use. Unless that choice has already been resolved in the business logic, the answer can sound much more certain than the underlying data deserves.
I have seen the same thing with customer identity. In many scale ups, customer information is spread across CRM, billing, product, support and marketing systems. Users belong to accounts, accounts belong to legal entities, names change, companies merge, people use different email addresses and historical mappings are incomplete. Then someone asks an apparently simple question such as, "Which customer segments generate the most support volume?" The AI can return an answer very quickly, but the quality of that answer still depends on whether the underlying customer mapping is reliable.
This is why I think the fundamentals matter even more in the AI era. Core metrics need clear definitions. Customer identity needs to be reliable. Business concepts need a shared semantic model. Metadata needs to be usable. Data quality needs to be good enough that people can trust the answers. Ownership also needs to be clear, because many of the inconsistencies that appear in analytics are really unresolved business decisions.
That last point is especially important. A surprising amount of data inconsistency comes from different parts of the company having different, perfectly reasonable interpretations of the same concept. Finance prefers one definition, Sales another, Product a third. Nobody makes the final call, so the disagreement gets embedded into the data stack and survives for years. Eventually leadership asks why the numbers do not match. At some point, someone has to decide which rule the company will use. Once that rule exists, AI can apply it consistently.
I remain very positive about AI analytics. I think it will reduce the number of dashboards companies need and make data much easier to access. I expect executives to explore performance more directly and I expect analysts to spend less time acting as human APIs for questions the business should be able to answer itself. The opportunity is significant. The companies that benefit most, though, will usually be the ones that have already done the less exciting work underneath.
That work means clear metrics, reliable models, good customer identity, strong data quality, shared definitions, useful metadata, clear ownership and a semantic layer that reflects how the business actually operates. Once those foundations are in place, AI becomes genuinely powerful because the system has enough structure and context to return answers people can trust.
For me, the goal is simple. A CEO should be able to ask an important question about the business and get a correct, trusted answer within seconds. AI is likely to become one of the best interfaces for delivering that experience. The company still needs to create the conditions that make the answer dependable.
So before asking which AI analytics tool to buy, I would ask a more basic question: can your leadership team already get a reliable answer to an important business question today? If the answer is yes, AI can probably make that experience dramatically faster and easier. If the answer is no, the first job is to understand why.
The AI layer is the easy part. Fixing the data mess underneath is where the real work begins.
