Most AI pricing tools have one mode: produce a recommendation. Feed in the data, receive a number, apply it. The question of whether the underlying analysis is strong enough to act on is usually not surfaced. The number comes out regardless.
We built Pricing Intelligence with a different default. In our four-station pilot, seven of eight product lines produced statistically validated pricing recommendations. The eighth — Site 2, Diesel — did not. So we held it back.
What the data showed
The statistical test for Site 2's Diesel produced a p-value of 0.41. In plain terms: the data could not reliably distinguish the price effect from random noise. The measured sensitivity was slightly negative (-0.69), suggesting modest volume response to price, but the confidence interval included zero — meaning we could not rule out that there was no real price effect at all.
Site 2 runs a fleet-heavy customer base. Fleet accounts tend to be contractually locked in, price-insensitive by nature. That context is consistent with the statistical result. But context alone is not evidence.
The data said we could not prove it. So we did not price it.
Why this is worth knowing about
For an operator, a held-back recommendation is not a failure. It is information. Site 2's Diesel is price-insensitive, or at least the current data cannot prove otherwise. That changes how you think about that product line — it is not a candidate for margin optimisation through price, and treating it as one would be guesswork.
For us, the held-back result is also a design principle. A pricing system that produces a number regardless of evidential quality is not a safer system — it is a system that hides its own uncertainty inside the output. The manager sees a recommendation. The manager does not see whether the underlying analysis passed or failed a significance test.
Pricing Intelligence separates those two things. Seven product lines cleared the bar; one did not. The seven get a daily recommendation. The one gets a flag and an explanation.
Broader implications
Statistical significance thresholds are the same standard used to validate A/B tests across major technology and e-commerce businesses. The bar — a p-value below 0.05, a 95% confidence interval that excludes zero — is not unique to academic research. It is the convention used by companies making pricing and product decisions every day.
Applying that standard to fuel pricing, per station and fuel type, means that every recommendation Pricing Intelligence produces has been tested against it. And it means that when the test fails, the recommendation is withheld — not hidden, not smoothed, not estimated.
From initially eight opportunities, seven proven but one held back. That is what evidence-based pricing looks like in practice.
About Backwell Tech
Backwell Tech is a Berlin-based high-tech company specializing in predictive AI solutions. The platform offers companies scalable AI models for profit maximization by utilizing historical and real-time data and ensuring data integrity. Since its founding in 2019, Backwell Tech has combined cutting-edge research with practical innovation in explainable algorithms. The company focuses on ethical AI development and delivers reliable, interpretable forecasts that enable informed business decisions. More information at www.backwelltechcorp.com.
Backwell Tech Corp contact:
Maximilian Gismondi
hello@backwelltechcorp.com
Sources:
- Pilot data: Backwell Tech internal pilot, 4-station network, 110-week backtest. Results indicative of opportunity, not a guarantee of future performance.
- Statistical framework: Wasserstein, R. & Lazar, N. (2016). The ASA Statement on p-Values. The American Statistician 70(2), 129-133.
- A/B testing convention: Kohavi, R., Tang, D. & Xu, Y. (2020). Trustworthy Online Controlled Experiments. Cambridge University Press.