General-purpose AI or a specialized pricing solution: Who should decide your prices?
Comparing two implementations of the same general-purpose language model before deciding which one to use to guide pricing is an increasingly common practice among large retail groups: both variants fall into the same category of tool.
A general-purpose LLM lacks four key components required for pricing decisions: access to real-world data, explicit business rules, impact simulation, and explainable governance.
Generative AI projects that combine in-house expertise with that of a specialized partner have a success rate far higher than those developed solely in-house: 67% versus 22%.
Comparing two implementations of the same general-purpose language model before deciding which one to use to drive pricing has become standard practice among large retail groups. Structurally, however, this is a poor starting point: both implementations belong to the same category of tools and lack native connectivity to pricing data, business rules, and the impact simulation required for pricing decisions.
This article explains why and sets the stage for deciding between building in-house and using a specialized solution.
An increasingly common dilemma
In recent weeks, several retail contacts have described a similar situation to us. Internally, two teams are each developing their own version of the same general-purpose language model, before comparing them to decide which one will be selected for pricing. The clearest example comes from a major retail group: two competing implementations of the same Gemini model, built by two separate teams at the parent company, are now being evaluated to determine which one will be integrated into pricing decisions.
On paper, the approach appears rigorous: benchmarking, measuring, and selecting the best implementation. However, it rests on an assumption that deserves to be questioned even before the comparison begins—namely, that two implementations of the same general-purpose model, no matter how well-crafted they may be, fall into the right category of tools for determining a price.
This is not an isolated case. As major retail groups ramp up their internal AI initiatives, the “build or buy” dilemma is becoming increasingly acute in the area of pricing—one of the functions where an ill-founded recommendation immediately costs the company in margins or price reputation.
What an Internal Bake-Off Really Measures
A comparison between two implementations of the same language model generally measures three things: the quality of the prompt, the configuration of the context provided to the model, and, in some cases, fine-tuning on a proprietary dataset. These are real, measurable differences that justify a decision between the two teams.
But these differences all stem from the same underlying principle. Both implementations remain a model trained to predict the most plausible sequence of words, not to reason based on a product catalog, price elasticity, or a minimum margin constraint. The comparison distinguishes between two variants of the same tool, without ever verifying whether that tool, within its category, was the right choice for this specific problem.
Our feature article, “Why Generative AI Gets Prices Wrong, ” explains this mechanism in depth: a general-purpose language model does not verify anything—it makes predictions. When it comes to a price—a specific, time-stamped piece of data—this structural limitation does not disappear simply because we chose the better of the two implementations tested.
What a general-purpose LLM Lacks Structurally
Regardless of the team that built it or the prompt used, a general-purpose language model lacks four key components that are essential for any retail pricing decision:
Data: connection to actual prices. Without a link to a product database and up-to-date competitive data, the model responds based on its training data, not on the actual conditions in the store.
Rules: explicit domain-specific constraints. Minimum margin, product line consistency, flagship products: a general-purpose LLM is unaware of any of these rules unless they are restated to it with each query—and there is no guarantee that it will follow them.
Impact: Pre-deployment simulation. A language model does not test the effect of a price on sales or inventory before recommending it. It generates text; it does not simulate anything.
Governance: Explainability and Auditing. A price recommendation must remain justifiable in hindsight. A general-purpose LLM produces plausible reasoning, but rarely an actionable audit.
The True Cost of Doing Everything In-House
Having two internal teams handle this type of decision-making comes at a cost that far exceeds the time required for benchmarking. Each team commits data scientists, infrastructure resources, and several months of development—all for a result that, by design, remains a variation of the same general-purpose model.
Generative AI projects in companies have no measurable impact on the income statement, according to a study based on 52 interviews with executives, 153 respondents, and an analysis of 300 deployments (MIT NANDA, *The GenAI Divide: State of AI in Business 2025*).
This figure does not suggest that generative AI is generally ineffective. It highlights a problem with the approach: the majority of internal projects remain poorly defined, poorly integrated into existing workflows, and lack a defined business objective before development begins—exactly the risk of a bake-off between two in-house implementations without a common business specification.
Generative AI projects in enterprises will be abandoned after the proof of concept by the end of 2025, due to insufficient data quality, spiraling costs, or poorly defined business value (Gartner, July 2024).
Having two teams develop two in-house implementations in parallel doubles the risk of the project being abandoned—not just the budget already committed. If the bake-off selects one of the two, the effort invested in the other will have yielded no pricing decision.
What a Specialized Pricing Solution Can Change
The same MIT report offers a lesson that can be directly applied to the "build or buy" dilemma in pricing: successful generative AI projects are not the ones that rely entirely on in-house development.
This is the success rate for generative AI projects that combine internal expertise with that of a specialized partner, compared to just 22% for projects developed solely by internal IT teams (MIT NANDA, *The GenAI Divide: State of AI in Business 2025*).
A specialized pricing solution is not an enhanced language model. It is a different category of tool, built from the ground up around the four building blocks that are structurally missing from a general-purpose LLM (connected data, business rules, simulation, governance), and operated by teams whose expertise lies in retail pricing, not general AI engineering.
Companies that do not yet use AI in their pricing cite a lack of expertise or internal resources as the main obstacle, according to a survey of more than 2,200 executives in 28 countries and 39 industries (Simon-Kucher, Global Pricing Study 2025).
This is exactly the gap that a specialized partner fills: the pricing expertise that two IT teams—no matter how skilled they may be in language models—generally do not have in-house.
The four building blocks brought together from the very beginning
BOOPER MPS integrates your pricing data, business rules, and impact simulation into a single engine. No “black-box” recommendations.
Build or buy? It only takes 30 minutes to find out.
Schedule a meetingThe Comparison
When presented in a table, the four missing building blocks and their operational consequences make the difference in category more tangible.
| Dimension | Generalist LLM (in-house model) | Specialized Pricing Solution |
|---|---|---|
| Actual price data | To be connected and maintained by the user, outside the template. | Native, connected from the moment of deployment. |
| Business rules | Reformulated with each request; compliance not guaranteed. | Integrated, versioned, and controlled. |
| Impact Simulation | Not included by default; to be developed separately. | Integrated prior to any deployment. |
| Explainability & Auditing | Plausible reasoning, rarely traceable. | A well-founded and historically grounded recommendation. |
| Required Team | Generalist data scientists, on an ongoing basis. | Pricing team, quick onboarding. |
| Time to value | Several months, uncertain (30% abandoned during the POC). | Weeks, as part of a pilot rollout. |
A 4-Question Framework for Making Decisions
The decision between building in-house and relying on a specialized solution cannot be made based solely on the quality of a model. It hinges on four specific questions, rather than a technical comparison.
- Is your data already integrated and reliable? Sales history, price elasticity, up-to-date competitor prices: without this foundation, no model—whether general-purpose or specialized—will produce reliable recommendations.
- Do you have a dedicated pricing team? A general-purpose IT team knows how to develop a language model. However, it rarely defines on its own the business constraints that make a pricing recommendation actionable.
- Does the decision need to be explainable and auditable? For a single introductory price, a plausible explanation is sometimes sufficient. For a national pricing policy, however, every recommendation must be justifiable in hindsight.
- Does the time to market matter? An internal bake-off takes several months before any pricing decisions are made. A specialized solution reduces this timeframe to a pilot deployment in a test category.
The Pitfalls of an Internal Bake-Off
- Compare only the quality of the generated text. Never its impact on a pricing decision tested on the actual catalog.
- Allowing two teams to duplicate the effort, without a quantified business objective common to both implementations.
- Do not include governance and explainability in the selection criteria, only to discover them upon deployment.
- Ignoring maintenance costs after the proof of concept. Yet this is the phase in which the greatest number of projects are abandoned.
- Never compare the time and cost of a bake-off to those of a specialized solution that has already been proven effective at other retail chains.
The choice isn't between two implementations of the same general-purpose model. It's between spending several months building the four building blocks that a specialized pricing solution already combines, or leveraging them directly so you can focus on the decision itself. Talk to our teams about your specific situation via our MPS page , Booper's modular pricing solution.
FAQ
Because both implementations fall into the same category of tool: a model trained to predict plausible text, not to reason about a product catalog, price elasticity, or margin constraints. The comparison distinguishes between two variations in prompts or settings, without verifying whether this category of tool, taken in isolation, is suitable for pricing decisions that require connected data, explicit business rules, impact simulation, and traceability.
A general-purpose LLM generates plausible text based on its training data, with no guaranteed connection to up-to-date data or explicit business rules. A specialized pricing solution combines demand modeling, versioned business rules (minimum margin, product line consistency), pre-deployment impact simulation, and explainable governance, all continuously connected to the retailer’s real-time data.
This depends on four criteria: the reliability of the data already integrated, the presence of a dedicated pricing team (not just general IT staff), the need for explainability and auditing of recommendations, and the expected time to value. Projects that combine in-house expertise with a specialized partner have a significantly higher success rate than those developed solely in-house.
According to a 2025 MIT NANDA study, 95% of generative AI projects in businesses have no measurable impact on the bottom line. The cause is generally not the quality of the model, but rather the approach: poorly defined business objectives prior to development, insufficient integration into existing workflows, and a lack of domain expertise to guide the project.
Technically, yes—through a custom-built integration—but this connection is not part of the model itself: it must be built, maintained, and secured separately, just like the business rules and the simulation layer. A specialized pricing solution natively integrates these components without requiring any additional development on the retailer’s part.
By setting a quantified business objective before development begins, ensuring the reliability of sales, elasticity, and competition data early on, incorporating explicit business rules rather than reformulating them with each query, and bringing specialized pricing expertise into the internal technical team from the outset—a combination that yields a significantly higher success rate than development carried out in isolation.
Sources
MIT NANDA, *The GenAI Divide: State of AI in Business 2025* · Gartner, July 2024 · Simon-Kucher, *Global Pricing Study 2025*. Last updated: September 22, 2026.
Comparing two implementations of the same general-purpose language model before deciding which one to use to guide pricing is an increasingly common practice among large retail groups: both variants fall into the same category of tool.
A general-purpose LLM lacks four key components required for pricing decisions: access to real-world data, explicit business rules, impact simulation, and explainable governance.
Generative AI projects that combine in-house expertise with that of a specialized partner have a success rate far higher than those developed solely in-house: 67% versus 22%.
Price optimization involves determining, for each product, the price that maximizes a defined business objective (margin, volume, market share, price image) within operational constraints. It is not about finding the highest price; rather, it is about finding the best balance among often conflicting objectives.
A retail optimization engine combines three layers: demand modeling, constrained optimization, and human oversight of sensitive trade-offs. None of these three is sufficient on its own.
Assuming a constant product mix and strategy, rigorous optimization generally yields an additional 0.5 to 1.5 percentage points in gross margin, with a return on investment achieved in less than 12 months.
Setting a selling price is based on three key factors (costs, demand, and competition), but in retail, the constraints lie elsewhere: a legal price floor (SRP+10 for food products through April 15, 2028), up to 40,000 SKUs in a hypermarket, and an operating profit margin of about 8% for every 1% change in price. Price setting becomes a managed process, with rules and safeguards, rather than an isolated calculation.