Strategy

AI Predicts What Customers Might Buy — and That's the Problem

Mission Field CEO Jonathan Tofel says AI consumer research creates 'convincing false positives' — sophisticated predictions of demand that no one has validated with a real purchase.

By Olivia Hart

4 min read

AI Can Predict What Customers Might Do
AI Can Predict What Customers Might Doathree23 / pixabay

What's News

  • Jonathan Tofel, founder and CEO of Mission Field, argues AI predictions of consumer behavior are not proof of purchase
  • Tofel warns AI can generate 'convincing false positives' before any customer spends money
  • He prescribes staged evidence: concepts, prototypes, realistic shopping environments, then real purchases
  • His pricing example contrasts stated willingness to pay at $3.69 or $4.25 with observed store purchases

By Olivia Hart

AI-powered consumer research can now test product concepts faster and cheaper than ever — and that speed is precisely the risk, according to Jonathan Tofel, founder and CEO of Mission Field. His warning: even increasingly accurate predictions aren't proof.

Tofel, writing in an op-ed, says he is bullish on the technology. "I use it throughout my business, and I think we're still in the early innings of understanding how much it can help companies make better decisions." Synthetic consumer panels can evaluate concepts, pressure-test ideas and identify promising directions, he notes. His caveat follows immediately: as prediction gets cheaper, companies should raise their standards for proof before committing significant resources.

Why saying isn't buying?

The core of Tofel's argument is what researchers call the say-do gap. Telling a researcher you would buy a new snack costs nothing. Standing in a grocery aisle, choosing that snack over the brand you've bought for years, and paying your own money is a different decision entirely.

Consumers are honest, Tofel insists. They may genuinely love an idea in a test setting. The problem is structural: "the conditions of the question are different from the conditions of the decision." In research, a shopper focuses only on the product. In a store, they face competing brands, prices, promotions, habits and limited attention.

AI doesn't remove those conditions. Predicting a purchase and observing one remain different standards of evidence — a distinction that matters when millions of dollars and months of organizational effort ride on the outcome. In plain words: the better the forecast looks, the more skepticism it deserves.

What makes false positives more convincing?

Tofel's sharpest point concerns failure modes. As AI grows more powerful, companies can generate and evaluate more ideas, build detailed business cases and assemble impressive supporting analysis. That capability, he argues, creates the potential for more convincing false positives — innovations that test well but flop in the real world.

He describes a plausible sequence. A company uses AI to screen dozens of concepts. The tool flags a promising candidate, recommends improvements, predicts consumer interest and builds a compelling launch case. The presentation is terrific. The analysis is sophisticated. The team feels confident.

But has anyone bought the product? "The danger is that increasingly sophisticated predictions can create a level of confidence that the underlying evidence hasn't earned," Tofel writes. Positive feedback builds momentum; additional research reinforces it; an idea feels like a winner before consumers make a single meaningful purchase. Team attachment, not customer demand, becomes the evidence.

How should companies stage their proof?

Tofel's proposed fix is progressive realism: the evidence should grow more demanding as an idea approaches market and the stakes rise. Early on, cheap AI tools generate concepts, identify potential customers and eliminate weak candidates — there is no reason to spend heavily validating every idea. As confidence grows, so should the realism:

  • Move from concepts to prototypes
  • Place prototypes in realistic shopping environments, with competitors, prices and distractions
  • When stakes justify it, observe real purchase behavior

Each stage answers questions the previous one couldn't, he argues. A concept test shows whether an idea appeals. A prototype reveals reaction to the physical product. A simulated shelf shows whether it gets noticed against alternatives.

His pricing example is the most concrete. Ask consumers whether they'd pay $3.69 or $4.25, and AI can recommend an attractive price point. But place the same product in different stores at those prices and watch what shoppers actually buy. Now you measure a real tradeoff with customers spending their own money — evidence that, in Tofel's words, deserves greater weight. Not every decision requires a full-scale market test, he concedes; the level of proof should match the decision's size and risk.

Do better predictions mean better decisions?

Tofel frames AI as lowering "the cost of curiosity": companies can ask more questions, challenge assumptions and analyze what once required enormous time and effort. He wants discipline attached to that bargain.

Interest in an idea is valuable evidence. A model's forecast of strong demand is valuable. But evidence that customers sacrifice their own money to choose a product over competitors is something else entirely. AI, he concludes, should free up time and resources to seek stronger evidence before the biggest bets — "because ultimately, the final vote is what customers choose to do." The operational takeaway: treat every AI forecast as a hypothesis, not a verdict, until real transactions confirm it.

Source: inc.com

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,378 articles

Related articles

« Previous articleNext article »