01

System Description

Purpose, audience and usage context

System Name Evaluator.gr
Version 1.0
AI Model Claude Sonnet 4.5 (Anthropic) via API
EU AI Act Classification Limited Risk — Article 52

Purpose and Functionality

The Evaluator.gr platform is a digital tool based on Artificial Intelligence technologies, created to strengthen Greek entrepreneurship. The system focuses on improving the quality of business documents, providing evidence-based feedback on critical files such as investor presentations (Pitch Decks) and business plans. At the same time, the platform assists in generating business content, facilitating the strategic communication of new ventures.

Audience and Fields of Application

The system is primarily aimed at new entrepreneurs and founders of early-stage startups. The platform can also prove useful to participants of business acceleration programmes (accelerators) or to investors who wish to get a quick, automated first assessment of the documents they receive.

Usage Context and Constraints

The application is accessed through the evaluator.gr website. It is important to note that the system's analyses are purely advisory in nature. The platform does not replace the decision-making process of its users or investors, and does not determine their final judgements.

ℹ️
Role Clarification: Evaluator.gr does not operate as an autonomous decision-making system. It does not determine who receives funding or who is admitted to an accelerator. The evaluation focuses exclusively on the structure and content of documents, not on the personality of the founders or the overall value of the business.
02

Capabilities and Limitations

What the system can and cannot do

The system has the capability to thoroughly analyse the structure and content of business documents, providing scored feedback based on academically documented criteria. Specifically, the evaluation of Pitch Decks is carried out through ten specific parameters, examining the material from three parallel perspectives: academic rigour, investment potential, and communication effectiveness. Regarding business plans, the analysis is grounded in the foundational research of MacMillan et al. (1985) and Gompers et al. (2020), ensuring alignment with established literature in the field.

Additionally, the platform incorporates tools for calculating startup valuations, applying six different methodologies adapted according to the stage of development, such as the Berkus, Scorecard, VC Method, as well as DCF, Comparables, and Revenue Multiple analyses.

✅ What it can do

  • Analyse the structure and content of business documents
  • Search real-time market data via web search (Startup Evaluator, Pitch Deck Evaluator, Business Plan Evaluator, Idea Validation, Competitive Landscape, Investor Readiness, Go-to-Market Plan, Legal & Compliance)
  • Provide scored feedback with academic documentation
  • Evaluate Pitch Decks through a triple perspective (academic, investment, communication)
  • Evaluate Business Plans based on MacMillan et al. and Gompers et al.
  • Calculate valuations with 6 methodologies per stage
  • Stage-adaptive evaluation with dynamic criterion weights
  • Generate business content (Pitch Deck, Business Plan, KPIs, etc.)

❌ What it cannot do

  • Predict the probability of business success
  • Access real-time market data (only for the 6 generation modules — the 8 evaluation modules have web search enabled)
  • Evaluate the execution capability of the team
  • Process information outside the submitted document
  • Make investment or funding decisions
  • Provide legal or tax advice

Content Generation Modules and Regulatory Framework

Beyond the evaluation tools, the platform incorporates 4 specialised modules for the automated generation of business content. These include tools for creating presentations (Pitch Deck Generator), performance indicators (KPI Generator), corporate identity support (Branding Assistant), and a business model canvas generator (Business Model Canvas Generator).

⚠️
Open Issue — Article 52(3) EU AI Act: The content generation modules raise an additional transparency question: whether recipients of AI-generated business documents (e.g. investors) should be informed of their origin. This issue is recognised as an area requiring further examination, both at the level of regulatory interpretation and business ethics.
03

Known Limitations, Weaknesses and Risk Management

Technical challenges and mitigation strategies

🎯

Position Bias

Position Bias

The model may assign different weights to evaluation criteria depending on the order in which they appear in the prompt.

Mitigation: Fixed evaluation structure and triple-analysis system for Pitch Decks.
🌀

Hallucination

Hallucination

The phenomenon of generating information not grounded in the user's document.

Mitigation: Explicit instructions prohibiting the invention of data, requiring reference exclusively to the submitted text.

Over-reliance

Over-reliance

Risk of adopting recommendations without critical evaluation by the user.

Mitigation: Clear disclaimers about the advisory nature of results in every output.
📊

Quality Variance

Quality Variance

The quality of feedback depends on the completeness and structure of the submitted document.

Mitigation: Document submission guidelines and input quality indicators.
🌍

Language Bias

Language Bias

The system has been optimised for Greek. Documents in other languages may produce lower-quality results.

Mitigation: Dedicated Greek terminology glossary and tailored prompts for the Greek ecosystem.
04

Evaluation Methodology and Academic Sources

Scientific basis and evaluation framework

Business Documentation Evaluation Framework

The platform's methodological approach to Pitch Deck evaluation is built on ten core pillars covering the full spectrum of the business proposal: from Problem-Solution Fit and Market Opportunity, to the business model, competitive advantage, team, and Go-to-Market Strategy. Each criterion is scored on a scale from 1 to 5, accompanied by a detailed, evidence-based justification drawn exclusively from the submitted document. Correspondingly, the evaluation of business plans focuses on team quality, market attractiveness and model viability, following the academic guidelines of MacMillan et al. (1985) and Gompers et al. (2020).

Temperature (Output Consistency) 0.3 — Low, for maximum consistency and reproducibility of results
Real-time Web Search Active in 8 evaluation & strategy modules — Inactive in 4 content generation modules

Dynamic Adaptation and Valuation Methodologies

One of the system's innovative features is Stage-Adaptive Evaluation. The weight of criteria changes dynamically according to the stage of business development. For example, at the Idea stage, team quality is given increased weight (35%), while at the Growth stage, the emphasis shifts to financial metrics and traction indicators, which together influence 40% of the final score.

In parallel, the choice of valuation method is made automatically, ensuring the appropriateness of the analysis. For very early-stage companies, qualitative methods are applied, such as Berkus, Scorecard, and Risk Factor Summation. For more mature ventures, the system favours quantitative approaches such as Discounted Cash Flow (DCF) and Revenue/EBITDA Multiples.

Generation Modules and Regulatory Observations

The use of content generation tools allows the creation of documents intended for presentation to external parties, such as investment bodies and accelerators. AI-generated content raises a critical transparency issue directly related to Article 52(3) of the EU AI Act. This issue represents an open area in academic literature, making further investigation of its implications for the startup ecosystem necessary.

📚 Academic Sources

  • MacMillan, I.C., Siegel, R., & SubbaNarasimha, P.N. (1985). Criteria used by venture capitalists to evaluate new venture proposals. Journal of Business Venturing, 1(1), 119-128.
  • Gompers, P., Gornall, W., Kaplan, S.N., & Strebulaev, I.A. (2020). How do venture capitalists make decisions? Journal of Financial Economics, 135(1), 169-190.
  • Hall, J., & Hofer, C.W. (1993). Venture capitalists' decision criteria in new venture evaluation. Journal of Business Venturing, 8(1), 25-42.
  • Berkus, D. (1996). Berkus Method for Early-Stage Startup Valuation.
05

Data Management and Security Framework

Policy on data collection, use and protection

Collection Policy and Data Classification

The platform operates under a strict information management framework that ensures user privacy and compliance with the General Data Protection Regulation (GDPR). The system stores submitted documents, evaluation results, and critical metadata relating to the business's stage of development, sector of activity, and funding needs. Technical data such as timestamps and user identifiers are also recorded to ensure the proper functioning of the service.

✅ Data we collect

  • Submitted business documents
  • Evaluation results and scores
  • Metadata (stage, sector, funding)
  • Technical data (timestamps, user ID)
  • Content generated by generation modules

❌ Data we do NOT collect

  • Biometric data
  • Health data
  • Ethnic or racial data
  • Special category data (Article 9 GDPR)
  • Data for AI model training

Use, Sharing and Information Security

Data is used to provide the service and to conduct academic research through anonymised samples. It is clarified that user information is not used for AI model training. A notable exception is the optional ability to share the business profile with investment bodies via a dedicated Investor Dashboard. This action is carried out exclusively upon explicit and active consent from the user, which can be withdrawn at any time.

At a technical level, security is ensured through the use of MySQL databases with encrypted connections, while communication with external APIs (such as Anthropic's) takes place via secure channels without data being stored by third-party providers. Data is retained for as long as the account remains active, with the user retaining the right to permanently delete their data at any time.

06

Logging Mechanisms and Traceability

Audit Trail and logging protocols

To ensure academic integrity and the technical reliability of the process, the system incorporates a comprehensive Audit Trail mechanism. For each individual evaluation, a detailed log is maintained that includes the user identifier, the exact date and time of the process, and the version of the AI model used. Additionally, the declared sector and development stage information, the individual scores per criterion, and any errors that occurred during processing are all recorded.

1

Transparency

Provides the user with the ability to fully monitor the history of their evaluations.

2

Methodological Verification

Enables academic confirmation of the applied methodology and ensures consistency of results.

3

Diagnostic Control

Facilitates the identification and correction of potential technical failures or logical errors in the system.

🔐
Log Access: Logs are retained for as long as the user maintains an active account. Access is provided exclusively to the user themselves via the /projects.php page, ensuring data remains confidential unless the user chooses otherwise through the consent process.
07

Human Oversight and Responsibility Boundaries

The role of the human factor and the limits of automation

The Role of the Human Factor

The operation of the Evaluator.gr platform is governed by the principle of human oversight, making the user the sole and final judge of the generated outputs. Every system output — whether an evaluation or generated content — requires critical assessment before any further use. The user remains the sole bearer of responsibility for the strategic decisions they make based on the system's information, which functions strictly as a support tool and not as an autonomous entity.

Automation Limits and Right of Review

In the context of avoiding uncontrolled automation, the system does not generate decisions that can directly affect third parties. No information or document is shared with external parties without the explicit and deliberate action of the user. If a user considers an evaluation inaccurate, they have the option to submit additional context or repeat the process with updated documentation.

External Oversight and High-Stakes Criteria

An additional external oversight mechanism is provided through registered investors in the Investor Dashboard, who are able to verify and critically evaluate the results presented by entrepreneurs. However, for high-stakes decisions — such as acceptance into an accelerator or investment decisions exceeding €100,000 — assessment by an experienced human expert is explicitly recommended before using the system's results. This approach ensures that AI remains a reliable assistant without replacing the necessary expertise in critical business scenarios.

⚠️
Known Limitation: The current human oversight mechanism relies on the judgement of the user and investors. Random sampling evaluation by independent experts is a priority for future development.
08

Version Management and Continuous Updates

Versioning, change documentation and reproducibility

Current Version 1.0
Release Date February 2026
AI Model Claude Sonnet 4.5 (Anthropic)
Status Production

Technical Infrastructure and Change Management

In its current operational phase, the system utilises Anthropic's advanced language model Claude Sonnet 4.5. To ensure scientific consistency, every material change to the underlying model or the structure of evaluation prompts is systematically recorded as a new version, enabling the traceability of changes in system performance.

Update management follows strict documentation protocols. Any modification to the evaluation criteria, the weighting per development stage, or the valuation methodologies is accompanied by the corresponding implementation date and relevant justification. This approach is considered essential for the reproducibility of results, ensuring that academic research can verify the system's methodological trajectory over time. Provisions also exist for informing users in cases of significant upgrades that may affect the interpretation of evaluations.

📋 Version History

v1.0
February 2026 — Initial Release

Full implementation of 10 modules (4 evaluation + 6 generation), stage-adaptive evaluation with dynamic weights, investor dashboard, and EU AI Act compliance framework.