As 2025 draws to a close, the bill for the Artificial Intelligence boom has officially come due. While corporate roadmaps remain cluttered with generative pilots, the gap between “magic” and “margin” in AI revenue generation is widening.
Recent data paints a stark picture of this “ROI Gap.” According to a December 2025 study from MIT, nearly 95% of enterprise AI projects are currently failing to deliver measurable returns. Similarly, Forrester reports that only 15% of executives have seen any improvement in profit margins from their AI investments over the last year.
The uncomfortable silence in boardrooms is no longer about whether the technology works – it’s about why it isn’t paying.
Moving from a promising demo to a revenue-generating engine requires more than just clean data and good models; it requires a fundamental shift in strategy – one that bridges the divide between executive ambition and engineering reality.
To navigate this divide, we turn to Vladyslav Chekryzhov, Director of Data Science & AI at AUTODOC. Operating across 27 distinct European markets, Chekryzhov sits at the rare intersection of executive product ownership and hands-on system architecture. Unlike the theoretical futurists often dominating the headlines, his mandate is grounded in the high-stakes reality of major e-commerce: delivering production-grade systems that directly influence pricing, retention, and customer loyalty.
He represents a discipline we might call “Revenue Realism” – the understanding that an AI model is only as valuable as its ability to survive in the wild and deliver measurable commercial impact.
Here are five strategic pivots required to turn AI hype into P&L reality.
The “Utility Filter”: Ruthless Prioritization
The first trap many organizations fall into is the “solution in search of a problem.” With the barrier to entry for Generative AI lower than ever, the temptation to build “cool” features is high. However, revenue generation requires a disciplined refusal to chase trends that don’t move the needle.
For Chekryzhov, the distinction between a feature and a business driver is stark. It begins not with code, but with financial modeling.
“Ultimately, prioritizing any AI/ML initiatives comes down to the discipline of building assumptions. Don’t rely on intuition; model the impact first – make money in Excel before the code is even written.”
He categorizes initiatives into three levels: Optimizing current economics (Level 1), Unlocking new product economics (Level 2), and Remodeling the business ecosystem (Level 3). The danger zone, he notes, is usually Level 3, where strategic stories often mask weak assumptions.
“The common failure mode is building an expensive toy… I force a vendor test: would we pay for this capability at vendor rates (e.g., OpenAI) and still maintain margins? If there’s no defensible path to revenue growth or a step-change in operating expenses, it’s just a costly experiment.”
Balancing the Algorithm: Pricing vs. Retention
In e-commerce, AI is often tasked with optimization. But optimization is rarely zero-sum. A model designed to maximize immediate margin (Dynamic Pricing) might inadvertently punish long-term loyalty (Retention).
Chekryzhov argues that managing this tension isn’t about finding the perfect neural network architecture, but about establishing the proper organizational boundaries.
“The minimum that works surprisingly well is culture, not architecture: rigorous experimentation with the right guardrails. Every pricing or promo change is measured not only on immediate efficiency but also on the “halo effects”: how it shifts behavior across cohorts and segments… We define upfront which metrics are allowed to move, in which direction, and by how much. If a margin win comes with a retention or CLV hit outside those bounds, it’s not a win.”
To implement this technically, he suggests avoiding “black box” monoliths in favor of a layered approach that gives business leaders control without requiring a full model retrain.
“One practical way to do it is a cascade of models: a pricing model proposes candidate prices, then lightweight models predict user outcomes and act as a filter or a weighting reranker. The benefit is control: you can adjust business logic by changing the final configuration rather than retraining the heavy model every time priorities shift.”
The “Production Gap”: Where ROI Dies
A Proof of Concept (POC) is a controlled experiment; production is a war zone. Many revenue projections fail because they underestimate the engineering overhead required to keep a model running at scale.
Chekryzhov warns that AI introduces a specific type of technical debt that traditional software engineers often miss: non-determinism.
“The honest answer is that a successful PoC doesn’t prove you have a scalable product… The model is non-deterministic: a rerun can produce different outputs. That explodes debugging cost, makes incidents harder to reproduce, and raises the bar for monitoring. Technical debt shows up sooner in AI systems than in traditional software, becoming a tax on the entire team’s development speed.”
Strategically, this means your ROI calculation must include the cost of reliability. If you only budget for development and not for the “tax” of maintenance, your margins will evaporate.
“The best investments I’ve seen here aren’t exotic… I push for basic hygiene (MLOps culture and the continuous process of ML systems design), the parts that don’t go out of date: measurable quality, debuggability, and reversibility.”
Isolating the Signal: The Attribution Challenge
Perhaps the most complex strategic question to answer is: “Did the AI do that?” In a complex ecosystem involving dozens of markets, seasonality, and marketing spend, attributing revenue to specific sources is statistically messy. Yet, without clear attribution, continued investment is impossible to justify to the C-suite.
Chekryzhov approaches this with the rigor of a scientist, rejecting the idea that complex models generate trust. Instead, he relies on counterfactuals – proving what would have happened in the absence of the AI.
“The only way to claim ‘AI drove X’ with a straight face is to anchor on a credible counterfactual. I rely on two families of evidence: randomized experiments (A/B) when feasible, and quasi-experimental methods when not. If the decision matters beyond the test window, we add a global holdout to the A/B setup: a persistent control group that never sees the feature. It’s painful – you’re literally losing money. But it’s often the only reliable link to reality.”
“For the C-suite, the message is consistent: trust doesn’t come from a complex model. It comes from a transparent approach and a measurement design you can explain clearly.”
Safety Rails: Trusting the Machine
Finally, automating revenue decisions – such as bidding or pricing – carries inherent risks. A “hallucinating” chatbot is embarrassing; a pricing algorithm that sells inventory at a 90% loss is catastrophic.
Strategic implementation requires a “human-in-the-loop” philosophy that evolves into “human-over-the-loop” governance. Chekryzhov advises assessing the cost of error before granting autonomy.
“I start with ML/AI system design, and one artifact matters most here: the cost of error. If the downside is high and hard to reverse, I don’t chase full autonomy… When the risk profile is acceptable, I like an “autonomy slider.” Early iterations are human-validated. As you accumulate data and confidence, you move the slider toward automation in controlled steps.”
Even when a system is fully autonomous, it must operate within strict bounds defined by the business, not the model.
“Autonomy must be bounded by policy-as-code. The system should have explicit constraints, circuit breakers, and safe fallbacks… You’re not debating autonomy in theory; you’re earning it.”
AI Revenue Needs a Maturity Upgrade
The transition from AI experimentation to AI revenue is not a technological upgrade; it is a maturity upgrade. It requires moving away from the allure of novelty and embracing the rigor of engineering, the complexity of attribution, and the discipline of prioritization.
As Chekryzhov’s experience at AUTODOC demonstrates, the companies that will win are not necessarily those with the most advanced models, but those with the most robust bridges between data science and business strategy.
