Not Every Pricing Problem Needs AI. Knowing Which Ones Do Is the Hard Part.
The instinct in recent years has been to reach for AI-first solutions and treat language models as a general-purpose analytics engine. For pricing in specialty materials and industrial businesses, that instinct is mostly wrong. The more important design question is not whether to use AI, but which parts of a pricing analytics system genuinely benefit from it, and which are better served by clean deterministic logic running on top of existing BI infrastructure.
Getting that distinction right is the difference between a system that produces auditable, actionable pricing decisions and one that produces outputs nobody can defend in a customer meeting.
The gap that BI alone does not close
The specialty materials and industrial sectors have made real progress on pricing analytics over the past decade. Investment in ERP modernization, business intelligence platforms, and data warehousing has given pricing teams tools that simply did not exist fifteen years ago. Power BI dashboards refresh nightly. Salesforce tracks discount patterns. SAP generates margin reports by product line. The infrastructure is real, and the investment has been substantial.
And yet a specific gap persists. Not in data collection, but in the translation of that data into account-level pricing decisions.
In their landmark work The Price Advantage (Wiley, 2004), McKinsey's Michael Marn, Eric Roegner, and Craig Zawada introduced the concept of the pocket price waterfall: the cascade of discounts, allowances, and cost-to-serve elements that separates a published list price from what a transaction contributes to the bottom line. Their research found that in most industrial businesses, the spread between the highest and lowest pocket prices in a customer portfolio is enormous, often 20 to 30 percentage points, and that this variation is rarely visible to the people making pricing decisions.
Today's BI platforms are excellent at the descriptive layer: margin by product, revenue by region, discount trends by customer segment, variance to budget. Gartner's research on analytics and BI platforms consistently shows strong adoption in manufacturing and distribution, with most mid-market firms now running some form of automated dashboard refresh. Where they stop is at two specific layers above the descriptive.
The first is analytical: applying a structured pricing framework like the pocket price waterfall at the individual customer-application level rather than at product line or segment level and doing so with enough granularity to allocate cost-to-serve correctly. This is a calculation problem, not an AI problem. It requires clean logic applied consistently across all accounts. Automating it is entirely achievable with modern orchestration tools without involving a language model at all.
The second is inferential: dealing with incomplete and unstructured data. In practice, B2B manufacturers rarely have precise share-of-wallet data for more than a fraction of their customer base. What they do have is often unstructured: estimates recorded in call notes, impressions from RFQ conversations, fragments captured in CRM free-text fields. Hermann Simon, in Confessions of a Pricing Man (Springer, 2015), notes that pricing power is fundamentally about understanding the relationship between price and customer value, and that this relationship is almost always partially opaque in industrial markets. This is where language models genuinely earn their place: structured inference from incomplete and unstructured information, with appropriate uncertainty expressed in plain language.
The right architecture is a hybrid
The two layers above suggest a clear division of responsibility. Deterministic rule engines generate pricing recommendations, not language models. Automated data pipelines compute margin and cost-to-serve, not AI. Language models are used selectively for the tasks where they genuinely outperform rule-based logic: estimating unknown inputs from partial evidence, synthesizing quantitative outputs into readable executive narratives, and surfacing segment-level strategic assessments that would otherwise require a skilled analyst to produce.
The reason this distinction matters is auditability. A recommendation to increase a strategic account's price by four percent needs to be traceable to a specific logic: margin gap versus application target, competitive position, share of wallet, customer tier. A sales rep presenting that recommendation in a quarterly business review cannot say "the AI suggested it." A CFO signing off on a portfolio-wide price action needs to be able to challenge the rule, not the model.
Anderson and Narus, in Business Market Management (Pearson, 2004), argue that superior customer value understanding is the foundation of pricing power in industrial markets. Building that understanding systematically, through continuously updated analytics rather than annual consulting reviews, is what the right tooling enables. The AI design decision is ultimately a question of where in that system human-legible logic ends and probabilistic inference needs to begin.
Why the decision matters more now than it did before
What makes the current moment genuinely different from the one in which The Price Advantage was written is the accessibility of the tooling required to build the hybrid system described above. A pricing intelligence platform covering the full pocket price waterfall, share-of-wallet estimation, a deterministic recommendation engine, and executive narrative generation can now be assembled from composable, low-cost infrastructure and deployed in weeks rather than quarters. The barriers that once made this the exclusive domain of large enterprises with enterprise software budgets have largely dissolved.
That accessibility makes the design decision the central challenge, not the technical one. Organizations that reach for AI indiscriminately will build systems that are opaque and difficult to trust. Organizations that apply AI where the problem is genuinely inferential, and deterministic logic everywhere else, will build systems that pricing managers, sales teams, and leadership can act on with confidence.
The question is not whether the technology is ready. It is whether the organization is clear on which problem it is solving.
References
Marn, M., Roegner, E., and Zawada, C. The Price Advantage. McKinsey & Company / Wiley, 2004.
Simon, H. Confessions of a Pricing Man. Springer, 2015.
Anderson, J.C. and Narus, J.A. Business Market Management: Understanding, Creating, and Delivering Value.Pearson, 2004.
Gartner Magic Quadrant for Analytics and Business Intelligence Platforms (annual).
