Don’t Let AI Make Bad Analytics Worse

Most companies are building AI analytics as an access problem: How do we let more people ask more questions of more data? But that’s the wrong starting point. The better question is: How do we help people make better decisions with the answers they receive? I’m an economist who has spent three decades studying how organizations use data to make decisions—as a professor at UC Berkeley, in senior leadership roles at eBay and Amazon, and through consulting with companies across a range of industries. In the course of that work, I’ve identified an important way of answering that question. The key is to make a distinction between self-service analytics and decision-service analytics. Self-service analytics starts with the user’s question: What do you want to know? Decision-service analytics starts one step earlier: What decision are you trying to improve? That difference matters. If a manager asks, “Which campaign had the highest conversion rate?” AI can answer quickly. But if the real decision is how to allocate next quarter’s marketing budget, the better question may be, “Which campaign generated the most incremental profit among customers who would not otherwise have purchased?” That may require experimentation, holdout groups, or a different measurement design altogether. AI can be extremely useful in this model. It can reduce rote work, help managers explore data, surface anomalies, generate hypotheses, write code, document workflows, and communicate findings. It can also free data teams to spend less time answering repetitive questions and more time on causal modeling, forecasting, experimentation, and strategic analysis. But that upside will not arrive automatically. Many companies will use AI analytics for efficiency (more automated queries, faster turnaround, fewer bottlenecks); better companies will use it for reliability (semantic layers, documentation, data provenance, validation workflows, expert review); and the best companies will use it for decision quality (redesigning analytics around the decisions that matter most). Ajay Agrawal, Joshua Gans, and Avi Goldfarb have argued that cognitive tools such as computers and AI often substitute for implementation skill—the ability to execute a task—while increasing the value of judgment. That’s right. But in analytics, the key managerial question is more specific: What organizational routines turn judgment into decision-grade evidence? Build Decision Discipline into AI Analytics AI analytics should not be treated as a plug-in for generating more answers. It should be embedded in an operating model for “decision-grade” analytics: the practices that determine which questions are worth asking, which answers can be trusted, and which decisions should follow. That operating model has six parts, which I’ll lay out below. Keep in mind that this model demands engagement not just from the data team but also leaders (who must own the decisions), analytics teams (who must own the measurement discipline), and senior executives (who must own the incentives that determine whether inconvenient evidence is welcomed or ignored). 1. Start with the decision. Before asking AI to analyze data, leaders should ask: What decision is this analysis meant to improve? What action might we take? What evidence would change our mind? Amazon’s Working Backwards process offers a useful analogy. Before building a product, teams often write a mock press release and FAQ that work backward from the customer problem and intended outcome. The point is not that this answers the analytics question. It is that it forces the team to clarify the decision before debating the evidence. AI analytics needs a similar discipline: Before asking the system to analyze data, leaders should first specify the decision that the analysis is meant to improve. Without that discipline, analytics becomes an exercise in producing interesting facts—or worse, motivated reasoning with better charts. A dashboard may show that one customer segment has higher retention than another, but this fact does not tell leaders what to do. The relevant analysis depends on the decision: whether to change pricing, adjust onboarding, alter product features, or shift marketing spend. Each choice requires a different comparison, a different standard of evidence, and often a different causal question. Decision-first framing forces a simple discipline: Before asking what the data says, specify what decision the evidence is supposed to improve. 2. Label the question. Many analytics failures begin because people do not know what kind of question they’re asking. A descriptive question asks what happened; a predictive question asks what is likely to happen; and causal question asks what will happen if we intervene. AI systems can blur these distinctions. A user asks whether a promotion “worked,” and the system reports that customers who received it bought more. That’s a descriptive comparison, not a causal answer. A model predicts which customers are likely to return, and a manager concludes that sending them an incentive will bring them back. That’s prediction, not evidence of incrementality. I’ve witnessed this confusion in many organizations. A team builds a model to predict which customers are likely to churn, and managers treat that as an answer to who should receive a retention offer. But those are different questions. “Who is at risk?” is predictive. “Whose behavior will change if we intervene?” is causal. Important analytics outputs should label the question explicitly: Is this analysis describing a pattern, predicting an outcome, or estimating the effect of an action? The answer determines the standard of evidence. 3. Expose assumptions and counterfactuals. Good analysts do not merely produce numbers. They expose assumptions. Which customers were included? How was the metric defined? Was the campaign targeted? Did competitors act at the same time? Is the analysis based on correlation, an experiment, or a quasi-experimental design? What would have happened otherwise? AI analytics should make these assumptions visible by default. Every important answer should include a short “reason to trust this” and “reason to be cautious.” A useful system should not only answer “What did you find?” but also “What would make this conclusion wrong?” I helped make this happen at eBay. Standard paid-search metrics suggested that advertising on branded keywords was highly effective. But the crucial counterfactual was whether those customers would have come to eBay anyway. Once we tested that question with experiments and holdouts, the conclusion changed: Much of the apparent value was not incremental. After that, eBay made counterfactual testing a much more central part of how it evaluated marketing spend. 4. Govern the metrics. Many AI analytics errors will arise not because the model is weak but because the data environment is ambiguous. Revenue, for example, may be defined differently by finance and product. Churn may depend on whether inactive customers are counted after 30, 60, or 90 days. “Active user” may mean logged in, transacted, opened the app, or used a core feature. A model cannot resolve these conflicts on its own. Airbnb faced this problem as it scaled. Different teams were giving different answers to basic business questions because they were using different tables, metric definitions, and business logic. The company’s response was to build Minerva, a metrics platform designed to provide a single source of truth for analytics, reporting, and experimentation. The lesson for AI analytics is straightforward: Before companies democratize answers, they need to govern what the answers mean. Companies need clear ownership of important metrics: canonical definitions, approved datasets, documentation, known limitations, and changes over time. Self-service analytics does not reduce the need for governance. It increases it. 5. Validate high-stakes and causal claims. Not every answer needs a committee. But some do. Organizations should identify high-stakes analytics in advance: major budget allocations, pricing changes, compensation decisions, regulatory claims, investor communications, large campaigns, personnel decisions, and strategic pivots. Financial institutions offer a useful analogy. In banking, model-risk-management frameworks commonly use risk-based prioritization: models with greater complexity or materiality receive deeper and more frequent validation. Companies using AI analytics need a similar triage system. Low-risk questions can move quickly, but high-stakes claims —about budgets, pricing, personnel, regulatory filings, or strategic pivots—should require review before they influence major decisions. For high-stakes claims, AI-generated analysis should be reviewed by people with relevant statistical, institutional, and business judgment. Validation should not mean checking the math after the conclusion is already attached to a decision. It should mean asking whether the analysis answers the right question, uses the right comparison, makes its assumptions explicit, tests plausible alternatives, and reflects the cost of error. This is especially important for causal claims. When feasible, leaders should ask for experiments. When experiments are not feasible, they should ask what assumptions are needed to support a causal conclusion. Before acting, they should ask what evidence would change their mind. 6. Reward truth-seeking, not persuasion. The biggest barrier to better analytics is often not technical. It’s organizational. If teams are rewarded for proving that their initiatives worked, they’ll find evidence that they worked. If leaders punish bad news, analysts will learn to soften it. If dashboards are used to justify decisions already made, AI will become another instrument of internal persuasion. Leaders need to create incentives for honest measurement. That means celebrating disconfirming evidence, funding holdout tests, separating measurement from campaign ownership when necessary, and asking teams to define success before the results arrive. The cultural message should be clear: We do not run analysis to win arguments. We run analysis to make better decisions. “Have Backbone; Disagree and Commit,” one of Amazon’s leadership principles, captures this cultural discipline. It requires leaders to challenge decisions respectfully when they disagree, even when doing so is uncomfortable, and to stand up for ideas using data rather than opinion. In my experience at the company, people were not merely allowed to challenge a leader’s interpretation of the evidence; they were expected to do so. That norm matters for AI analytics, because the danger is not only that systems will produce bad answers but also that organizations will reward the answers that make existing plans easier to defend. . . . The phrase “data-driven” has always been slightly misleading. Data does not drive decisions. People do. Data can discipline judgment, challenge intuition, and reveal patterns we would otherwise miss. But data can also mislead when organizations confuse measurement with understanding. AI will give leaders more answers than ever before. Some will be valuable. Some will be wrong. Some will be technically accurate but useless for the decision at hand. Some will be seductive because they confirm what powerful people already believe. AI will transform analytics by making answers abundant. But answers are not the same as evidence, and evidence is not the same as a better decision. The managerial task, then, is not to slow AI down everywhere—it’s to add friction where the cost of being wrong is high, while letting low-risk analysis move faster. Before acting on an AI-generated answer, leaders should ask: What decision will this analysis change? What comparison are we making? What would have happened otherwise? What would convince us we are wrong? The winners here will not be the companies that ask the most questions. They will be the companies that know which answers are good enough to act on.

Don’t Let AI Make Bad Analytics Worse

Summary.

Partner Center