Photo By: Michael Weidemann
For years, businesses have used artificial intelligence to predict what is likely to happen. James Sun, CEO of Kapnova, believes the next step is answering a more difficult question: what will happen because a company does something?
That distinction is at the center of what Sun calls causal AI.
Predictive AI identifies patterns in historical data. Causal AI attempts to determine whether an action actually produced an outcome. For a retailer, for example, a predictive model might show that customers who receive a 20% discount are more likely to buy. But those customers may already have been likely to purchase, meaning the discount did not necessarily cause the sale.
For executives, Sun argues, the more valuable question is whether changing a price, promotion or marketing budget will produce an incremental financial impact.
“I sometimes describe what we’re building at Kapnova very simply: AI finds the opportunities. Math determines the answer,” Sun said.
Kapnova uses AI and agents to sift through large amounts of information and identify business questions worth investigating. The company then applies causal inference, econometrics, forecasting, simulation and optimization to evaluate potential decisions.
The approach begins with uncertainty rather than an immediate answer. Businesses often make multiple changes at once, from advertising and promotions to pricing and inventory, making it difficult to determine what actually drove a change in sales.
Kapnova attempts to reconstruct the counterfactual: what would most likely have happened if the company had not made the decision?
Depending on the circumstances and available data, that can involve causal inference techniques, econometric models, natural experiments, matched groups, time-series analysis or controlled experiments. Sun also emphasizes that some questions simply cannot be answered reliably with existing evidence.
“A good decision system should sometimes say, ‘We don’t know yet. Here’s the experiment we need to run,’” he said.
Sun’s interest in applying quantitative methods to consumer businesses partly stems from his cofounder’s background as a hedge fund quant modeler at Point72. He sees a contrast between financial firms that may use sophisticated mathematical models before making multimillion-dollar investments and consumer companies that can make similarly consequential decisions using dashboards, historical comparisons and executive judgment.
He argues that the difference is less about sophistication than infrastructure. Financial markets developed extensive systems for quantitative decision-making, while consumer businesses built powerful tools for reporting and attribution. Those systems can explain what happened, but Sun believes they are less equipped to determine what a company should do next.
That is where he sees an opportunity for AI.
At Kapnova, the focus is deliberately centered on decisions affecting revenue, gross margin and contribution margin. Pricing, promotions and marketing are among the clearest applications.
A company might ask whether raising a price by 5% would reduce demand enough to outweigh the additional margin. A promotion might increase revenue substantially while generating little incremental demand if customers would have purchased anyway. Similarly, a marketing channel may receive attribution for sales without actually generating much incremental revenue.
Sun believes these questions can exist across dozens or even hundreds of decisions within a company.
His longer-term vision is for AI agents to continuously identify such opportunities and present executives with evidence, models and estimated financial outcomes.
The stakes can be substantial. Sun recalls a beauty, fashion or consumer packaged goods brand that spent roughly $5 million on a Tier 1 celebrity influencer campaign but generated only about $140,000 in incremental revenue, according to his account. While the campaign generated significant exposure, Sun argues that exposure alone did not establish causality.
“Reach and causality are not the same thing,” he said.
He sees that distinction as central to how companies should evaluate major investments. Kapnova’s approach includes backtesting decisions by using historical information to model what the company would have recommended at the time, then comparing those recommendations with actual outcomes.
Still, Sun does not present causal AI as a way to eliminate uncertainty. Poor data cannot produce reliable causal conclusions, and historical information may not reflect a dramatically changed business environment. In some cases, an experiment may be necessary.
Human judgment also remains important. A model might recommend raising prices to maximize near-term contribution profit, while an executive may deliberately accept lower margins to enter a new market or expand distribution.
For Sun, the future is therefore not about replacing executives with AI. Instead, he expects executives to use quantitative evidence to test their instincts before committing resources.
The shift could eventually extend beyond executives using AI tools to AI agents making operational decisions themselves. Marketing, commerce and inventory agents could propose changes, but Sun believes those systems will need an underlying decision layer capable of modeling economics, respecting business constraints and creating an auditable record of why a decision was made.
Kapnova’s broader ambition is to build that layer. The initial question may be whether a company should change a price. Ultimately, Sun envisions a system in which an AI agent can ask whether it is authorized to make that change and what the evidence indicates will happen if it does.
Show Comments