Most conversations about AI and data analysis get stuck on the tools. Can it write SQL, build a chart or find a pattern faster than an analyst? Usually it can, and that is the easy part.
The harder question is whether the pattern means anything, whether the recommendation fits the business and whether acting on it is likely to produce a better outcome. Those decisions depend on context that rarely exists in one dataset or prompt.
I use AI regularly to explore problems, write and check code, generate possible explanations and document methods. It can contribute to the analysis, but I still own the question, the interpretation and the recommendation.
AI can see the dataset, but it cannot see the organisation
A dataset is only one layer of a business problem. Around it sits domain knowledge, internal events, external conditions, stakeholder relationships and organisational constraints. Those layers often determine whether an analytical answer is useful, misleading or impossible to act on.
This is what gets lost when analysis is described as a neat sequence of data in and insight out. The same number can mean different things depending on what changed inside the business, what customers are doing, who owns the decision and what happened in the market last week.
1. Domain knowledge determines what deserves attention
AI is good at generating plausible hypotheses, while domain knowledge helps determine which hypotheses are commercially realistic and worth testing.
In retail and FMCG, a sales decline might reflect distribution loss, pricing, promotional timing, poor availability, competitor activity, seasonality, range changes, packaging or a change in how the data is classified. A model can list those possibilities, but it does not automatically know which signals usually move first or which explanations make sense for that category.
Domain knowledge also helps identify results that are technically correct but conceptually wrong. I have seen clean sales data understate a segment because a major brand had disappeared from the category hierarchy. Nothing was broken in the file, but the definition no longer represented the market.
That distinction matters because analysis begins before the calculation. Deciding what the data should represent and which explanations deserve investigation already requires judgement.
2. Internal context explains numbers the model cannot
Important business events often sit outside the data warehouse. A packaging change, retailer dispute, out-of-stock issue, new target, delayed campaign or quietly revised reporting definition can completely change the interpretation of a result.
A conversion rate may have fallen because customer demand weakened or because the application process gained another step. Sales may be down because the product is less appealing or because it disappeared from shelves for three weeks. The chart can look the same in either case.
At BIC, a key product range declined by 48% within one month. I eliminated price, product quality and distribution as likely causes, then combined scan data, consumer panel data and custom shopper research. The sales data showed where the decline occurred, the panel showed previous buyers switching away, and the research explained why: the new opaque packaging prevented shoppers from seeing what they were buying.
AI could have helped organise the data, write code or generate hypotheses, but it would not have known unless someone supplied the context that the packaging had recently changed, a retailer range review was approaching and global teams controlled the redesign. Those facts affected both the analysis and the case I needed to make.
The useful answer was not simply that packaging correlated with the decline. It was a recommendation that could survive internal review, retailer scrutiny and the delay before redesigned packaging reached the shelf.
3. External context tests whether the business caused the change
A business never operates inside its own dashboard. Competitor launches, retailer decisions, regulation, economic conditions, seasonality and changes in customer behaviour can all affect the result.
Internal data tends to make every problem look internal. When sales fall, the first instinct may be to review the product, campaign or sales team, even though the category could be shrinking or a retailer may have changed its ranging strategy.
AI can help identify external explanations and compare them with internal trends, but someone still has to judge the quality, timing and relevance of the evidence. A news report, competitor claim and reliable market dataset are not interchangeable simply because they appear in the same answer.
4. Stakeholder relationships shape the useful answer
Good analysis is not only about finding the right answer; it also requires understanding the person who needs to use it.
A finance leader, sales director, product manager and retailer buyer can examine the same finding and ask different questions. One may want confidence in the calculation, another may care about what can change this quarter, while someone else may need to defend the recommendation in a customer meeting.
At Nielsen, much of the work involved understanding what a client was really asking rather than producing another cut of the data. A request for a market-share chart could actually reflect concern about an approaching range review or a need to explain a missed target to leadership.
Relationships reveal where the trust gaps sit, how much background is required and which recommendations are realistic within the organisation. AI has the words supplied in a prompt, but it does not have the history behind the conversation.
5. Organisational constraints affect what can happen next
Politics in analytical work often means incentives, ownership, risk and influence rather than something more dramatic.
Who owns the metric, controls the budget or will be asked to change? Who carries the risk if the recommendation is accepted? Does the decision need local, global or customer approval, and is there enough time to act before the next planning cycle?
A technically correct recommendation can fail because it arrives too late, ignores a commercial commitment or asks someone to take a risk they cannot defend. Pretending those constraints do not exist does not make the analysis more objective; it makes the recommendation less useful.
I do not change a finding to suit organisational politics, but I may change how I build the case. That could mean involving a stakeholder earlier, making the assumptions more visible, distinguishing firm evidence from directional evidence or presenting a staged recommendation rather than one large change.
AI can help structure that argument, but it cannot take responsibility for the trade-off.
Where AI fits in my workflow
I define the decision and assemble the relevant context before asking AI to help. Once that boundary is clear, I use it to draft code, generate competing hypotheses, summarise qualitative responses, document the method and challenge my first interpretation.
I then validate the technical output independently and compare the result with what I know about the category, customer, process and market. If it conflicts with that context, I investigate whether the result is a valuable surprise or evidence that something important is missing.
The final recommendation remains mine because I am connecting the analytical evidence to the people, constraints and consequences around the decision.
The value of an analyst was never typing every formula by hand. It lies in knowing which problem is worth solving, whether the evidence deserves to be trusted, what the result means in context and how to help someone act on it.
As AI makes producing analysis cheaper, judgement becomes the scarce part. My rule is to use AI to widen the analysis, accelerate the technical work and challenge my reasoning, without outsourcing the decision.