AI in the bid process: where it already works and where the limits lie

Introduction

Few topics in industrial procurement are currently debated more controversially than the use of AI in bid preparation. The promises are considerable: automatic requirement extraction, intelligent classification, and standards detection at the push of a button. As is so often the case, reality is more nuanced.

This article provides a sober overview of the state of the art, shows where AI already delivers measurable value today, honestly identifies its limitations, and places the development in the context of the rail industry. No hype-driven prose, but a factual assessment for bid managers who want to know what works—and what does not (yet).

An overview of the four fields of application

AI in the bid process can be divided into four clearly distinct fields of application, each with a different level of maturity:

Requirement extraction: the most mature use case

The automatic extraction of requirements from unstructured documents is currently the most advanced AI use case in the bid process. The problem is clearly defined: a specification is available as a PDF or Word document, and the requirements must be transferred as structured records (ID, text, standards reference, bindingness) into a requirements management system.

Siemens integrated AI capabilities for bid preparation in Polarion in 2025. The AI layer analyses incoming tender documents (RFQs, specifications) and extracts requirements automatically, without manual input. According to Siemens, this can reduce analysis time “from weeks to hours.”

In a separate blog post Siemens quantifies the efficiency gain as “up to 25%.” Caution is warranted, however: the figure comes from the energy sector and refers to the overall process, not specifically to the rail industry. Whether comparable gains can be achieved in the far more standards-intensive rail industry remains to be seen.

semantha: AI for requirements assessment in Polarion

Alongside the native AI capabilities, semantha offers a specialised extension for Polarion that focuses on AI-assisted requirements assessment. semantha analyses requirements semantically and can search existing requirements databases to find similar requirements from previous projects. This approach is especially relevant for recurring specifications in the rail industry, where many requirements are worded similarly across tenders.

Classification support: promising, but not autopilot

Automatic classification of requirements (OK, OKB, NOK, OKM, R) is an obvious AI use case. The logic is simple: if a company has thousands of requirements already classified from previous projects, a machine-learning model can learn to classify similar requirements in a similar way.

In practice, the most effective approach is an assisted system: AI proposes an initial classification (including a confidence score), and the subject-matter expert reviews and corrects it. For unambiguous cases (high confidence), this saves significant time. For borderline cases (low confidence), the system explicitly alerts the expert that a review is needed.

Standards detection: a high degree of automation is possible

Automatic detection of standards references in free text is a use case where AI already works reliably today. The task is clearly defined: in text such as “The fire-protection requirements in accordance with DIN EN 45545-2:2020 in hazard category HL3 must be met,” the references DIN EN 45545-2:2020 and HL3must be recognised and classified.

Estimated recognition rate of AI systems by reference type (qualitative assessment, not based on a specific study).

The challenge lies less in recognition itself than in post-processing: is the detected standard current? Is there a newer edition? Does the standard reference other standards that apply implicitly? This cross-reference analysis requires a current standards database. A topic explored in greater depth in our article Standards compliance in tenders.

Expert routing: AI as a recommendation system

The automatic assignment of requirements to specialist departments is a use case that depends heavily on the quality of historical data. Companies that have documented their assignments from previous projects properly can train a model that generates assignment suggestions for new requirements.

As described in our article Expert routing in tenders, this works well for clearly assignable requirements (fire protection → fire protection team), but reaches its limits with cross-functional requirements. The most effective approach remains: AI suggests, people review and correct.

The honest limits: what AI cannot (yet) do

In addition to the use-case-specific limitations, there are fundamental constraints bid managers should know about:

1. Project context is missing

An AI system does not know the specific vehicle concept. It does not know that the customer requires a low-floor BEMU with a top speed of 160 km/h and that the company itself has so far only built 140 km/h platforms. The classification “OKM: with adaptation of the drivetrain” requires engineering knowledge that no language model possesses.

2. Liability and traceability

Every classification in a bid is a binding commitment. If an AI system proposes that a requirement is “OK” and the subject-matter expert adopts the proposal without checking it, the company—not the algorithm—is liable in the event of a dispute. ISO 22163:2023 requires complete traceability: who made which decision, and on what basis? “The AI suggested it” is not acceptable evidence.

3. Confidentiality and data protection

Specifications are confidential documents. They contain details about planned transport infrastructure, technical specifications, and commercial terms. Processing them using cloud-based AI systems (in particular LLMs such as GPT-4 or Claude) raises legitimate questions:

• Will the data be used to train the model?
• Where is the data processed and stored (GDPR compliance)?
• Can the contracting authority prohibit the sharing of its documents with third-party providers?

4. Hallucination risk

Generative AI models can produce statements that sound plausible but are factually incorrect, a phenomenon known as “hallucination.” In a bid context, this would be fatal: an AI system that references a non-existent standard or suggests an incorrect compliance assessment can lead to erroneous bids. Human review of AI suggestions is therefore not optional, but mandatory.

Practical architecture: embedding AI in the bid process

The most effective integration of AI into the bid process follows the pattern “AI suggests, people decide.” Platforms such as Tendric are built around precisely this architecture:

Market overview: what is available

The tooling landscape for AI in bid preparation can be broadly divided into three categories:

1. Integrated AI in ALM systems

Siemens Polarion has integrated native AI capabilities since 2025. The functionality includes requirement extraction from tender documents, semantic analysis, and links to existing databases. Hitachi Rail has already integrated Polarion into its project workflows.

2. AI extensions for existing systems

semantha is an example of a specialised AI extension that plugs into existing Polarion installations. The advantage is that companies can add AI functionality incrementally without replacing their existing ALM system.

3. Specialised platforms

Alongside the major ALM providers, specialised platforms are emerging that focus on AI-assisted bid preparation, with a lower barrier to entry than Polarion or DOORS but targeted support for extraction, classification, and specification export. Tendric is an example of this approach: the platform combines AI-assisted requirement extraction with structured classification and expert routing, without the overhead of a full ALM system. Such tools particularly address mid-sized suppliers for whom Polarion or DOORS is oversized.

Implementation strategy: how companies get started

Companies that want to introduce AI into their bid process benefit from a phased approach:

Recommended implementation sequence by maturity level and risk (our own assessment).

Phase 1 (extraction) offers the fastest ROI with the lowest risk: the result is a structured requirements list that people review before any assessment takes place. Specialised platforms such as Tendric enable this gradual entry point without requiring a complete ALM migration. Phases 3–5 require company-owned training data (historical classifications, assignments) and are therefore more demanding to implement.

Conclusion

AI in the bid process is neither a cure-all nor hype; it is a tool with clear strengths and equally clear limitations. Requirement extraction and standards detection are ready for productive use today. Classification support is promising, but it is not autopilot. Final assessment, commercial decisions, and liability remain with people.

For bid managers in the rail industry, the key question is not “Should I use AI?” but “Where should I use AI, and where explicitly should I not?”