← Insights

From idea to MVP of a B2B product with a language model

5 March 2026 · 3 min read

The first version of a B2B product with a language model requires decisions on data, quality evaluation and cost control. The article puts those decisions in the order of stages from idea to MVP and gives the Artificial Intelligence Act dates after the July 2026 amendments.

From idea to MVP of a B2B product with a language model

In a B2B product a language model most often organises documents, drafts replies or classifies tickets. The value of such a feature depends on the client’s data, on whether the result can be checked and on who approves the decision taken on the basis of it.

The article describes the stages from idea to MVP in the order in which decisions are made (author’s analysis), together with the dates of Regulation (EU) 2024/1689, the Artificial Intelligence Act (AI Act), as amended by Regulation (EU) 2026/1744.

In this article

Problem and success metric

The first step is to describe a single task the model is to perform and how that task is done today. Without this baseline it is impossible to tell whether the MVP improves anything. The success metric should be measurable on pilot data: time to handle a case, the share of drafts accepted without changes, or the number of tickets routed to the right queue (author’s analysis).

Four stages from idea to MVP: problem and metric, data and classes, quality evaluation, MVP, with a timeline of AI Act dates from 2 February 2025 to 2 August 2028.
MVP stages (author’s analysis) and AI Act dates after Regulation (EU) 2026/1744.

Data and where it is processed

The second stage decides which data reaches the model and where it is processed. Client documents, personal data and trade secrets have different requirements, so it is worth classifying the data before choosing a model. The data class determines whether a local model is enough or whether a cloud provider is acceptable, and on what contractual terms (author’s analysis).

Quality and risk evaluation

Before the MVP reaches clients it needs a set of test cases with expected results and a review of the output by a person who knows the process. The OWASP Top 10 for LLM Applications 2025 lists risks worth checking already in an MVP: LLM01 Prompt Injection, LLM05 Improper Output Handling, LLM06 Excessive Agency, LLM09 Misinformation and LLM10 Unbounded Consumption.

The last entry calls for limits on the number of requests per user and on cost per billing period, with a log of model calls (author’s analysis).

AI Act dates after the 2026 amendments

Regulation (EU) 2024/1689 of 13 June 2024 applies in stages (Article 113). Chapters I and II, the general provisions and prohibited practices, have applied since 2 February 2025. Regulation (EU) 2026/1744 of 8 July 2026 moved the dates for high-risk systems. Chapter III, Sections 1 to 3, applies from 2 December 2027 to systems classified as high-risk under Article 6(2) and Annex III, and from 2 August 2028 to systems under Article 6(1). Systems generating synthetic content placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026.

The product’s classification under the AI Act is a decision worth recording already at the MVP stage, together with its rationale (author’s analysis).

Sources

  • Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Article 113 – eur-lex.europa.eu/eli/reg/2024/1689/oj
  • Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI) – eur-lex.europa.eu (CELEX 32026R1744)
  • OWASP, Top 10 for LLM Applications 2025 – genai.owasp.org/llm-top-10/

Key takeaways

  • An MVP with a language model needs a baseline and a success metric that can be measured on pilot data.
  • The data class determines the choice between a local model and a cloud provider.
  • Cost limits and a log of model calls reduce the LLM10 Unbounded Consumption risk from the OWASP list.
  • After the 2026 amendments, obligations for high-risk systems listed in Annex III of the AI Act apply from 2 December 2027.

Next steps

  1. Describing one task for the model and measuring how that task is done today.
  2. Classifying the data and deciding where each class is processed.
  3. Preparing a set of test cases and having the results reviewed by a person who knows the process.
  4. Recording the product’s classification under the AI Act with its rationale.