Evaluating technical proposals in grant calls is an intensive, repetitive process highly susceptible to human variability. Artificial intelligence offers a real opportunity to improve the consistency, speed, and traceability of this process, but there is no single, one-size-fits-all solution for every context.

In this article, we present five implementation scenarios, ordered from least to most complex, alongside a cross-sectional analysis of the common challenges that any solution must address.

Scenario 1 — Prompt-Assisted Evaluation‍

Description

The user relies on a structured set of prompts, designed and validated internally, to evaluate each proposal with the help of a generic AI tool (Claude, ChatGPT, Microsoft Copilot). There is no automation or technological integration here: the value lies in the procedural workflow, not the platform itself. The user copies the relevant content from the proposal, applies the corresponding prompt to each criterion, and manually records the result.

This scenario is not simply "using AI informally." Its utility depends entirely on the prompts being well-designed, shared across the team, and applied consistently, functioning almost like a standard internal operating procedure.

When it makes sense

As a starting point for any organization wishing to understand what AI can bring to the table before making any financial investment. It is also an essential prerequisite exercise for any of the subsequent scenarios: the process of writing effective prompts forces teams to formalize evaluation criteria with a level of precision that often does not yet exist.

Auditability and traceability

Very limited. It depends entirely on the individual discipline of the reviewer. If the evaluating officer saves the input and output of each evaluation as part of the case file, a minimal audit trail exists. Without that manual logging protocol, there is zero traceability. It is not suitable as a definitive, long-term solution.

Advantages

• Zero implementation cost.
• Implementable in hours.
• Standardizes criteria across reviewers if applied with strict discipline.
• Lays the conceptual groundwork for more advanced scenarios.

Limitations

• Wholly dependent on the individual discipline of the reviewer.
• Without a supporting technological framework, consistency degrades over time.
• Traceability is almost non-existent without a manual record-keeping protocol.
• Does not scale with high volumes.

Scenario 2 — Agent on an Existing Platform (Claude / ChatGPT / Copilot)‍

Description

The reviewer directly uploads the technical proposal PDF to a conversational AI platform (such as Claude.ai or Microsoft Copilot). The AI agent, pre-configured with the specific grant call's rubric, reads the document, applies the criteria, and returns a score sheet complete with justifications. The reviewer then validates and signs off on it.

When it makes sense

When an organization wants to validate the concept quickly, without any technological investment or reliance on an IT team.

Auditability and traceability

Limited. The conversation logs remain on the provider's platform, but there is no structured record-keeping system. Reconstructing the historical context of a specific evaluation can be difficult. It is not suitable as a definitive solution in a public administration environment where strict traceability is a mandatory requirement.

Advantages

• Implementation takes days.
• Zero development costs.
• Allows teams to learn what works before investing further.

Limitations‍

• The evaluation rubric must be manually updated for each individual grant call.
• Does not scale well when multiple reviewers are working in parallel.
• The audit trail is incomplete.
• Case file data passes through external third-party servers.
  

Scenario 3 — Automated Workflow on Existing Infrastructure‍

Description

When a reviewer uploads a PDF to a folder or document management system, an automated workflow is triggered: it reads the call's rubric, calls an AI model, and writes the results into a structured, shared format.

What differentiates this scenario from the previous one is not the specific technology used, but the fact that the process itself is automated: the reviewer simply receives the generated result rather than having to manually prompt for it.

In the context of Spanish public administration, where Microsoft 365 is widely deployed, this scenario can be naturally implemented using Power Automate for orchestration and Azure OpenAI as the AI model, with the results seamlessly exported into SharePoint or Excel. It requires no new infrastructure acquisition. For organizations preferring non-Microsoft tools, alternatives like Make or n8n offer equivalent automation capabilities and can connect to various AI models.

When it makes sense

When the evaluation model has been successfully validated (Scenarios 0 and 1) and the organization wants to take a step toward consistency and scale without undertaking custom software development.

Auditability and traceability

Medium-high. The document management system (SharePoint or equivalent) provides native version control and access logs. Results are stored with precise timestamps and authorship. It is fully possible to reconstruct which AI model evaluated which proposal and against what rubric, provided the workflow is well-designed from the outset. This is a highly valid solution for public administration if complemented by clearly defined data retention and access policies.

Advantages

• Leverages infrastructure already present in most public sector organizations.
• Reduces reliance on the discipline of individual reviewers.
• Multiple reviewers can work in parallel.
• Full history and traceability reside within the secure document system.

Limitations

• Requires an IT profile for the initial setup and configuration.
• Adaptation to different types of grant calls must be continuously validated.
• Does not natively generate official award resolution reports.
• Data still passes through a third-party provider's servers. In the case of Azure OpenAI, however, it is possible to configure the deployment in EU regions to ensure data residency and strictly prevent the use of content for model training.
  ‍

Scenario 4 — Custom Evaluation Platform

Description

A web application designed specifically for this exact process. It manages calls, case files, and reviewers. The AI automatically scores each proposal, the human reviewer validates the output on-screen, and the system directly generates the official award resolution report. Everything is logged and fully auditable.

When it makes sense

When the sheer volume of grant calls and proposals justifies the financial investment, and when the organization requires absolute audit guarantees, seamless interoperability with other internal systems, and total end-to-end control over the process.

Auditability and traceability

High. It enables a complete, easily exportable audit trail: detailing exactly which criteria were applied, the specific score proposed by the AI, any modifications made by the human reviewer, and the exact timestamps of these actions. It is tailor-made to meet the stringent transparency and accountability requirements inherent to public administration. Data can be hosted on on-premises infrastructure or within a European sovereign cloud.

Advantages

• Designed specifically around the grant evaluation process.
• Complete, exportable auditability that integrates seamlessly with other systems.
• Scales effortlessly across multiple grant calls and evaluating teams.

Limitations

• Requires significant financial investment and development time (months).
• Necessitates ongoing maintenance.
• Creates greater dependency on the technology vendor if the platform is not built using open standards.
  

Scenario 5 — Proprietary Model Trained and Deployed On-Premises

Description

Instead of relying on generic AI models from external providers, the organization develops and deploys its own proprietary evaluation model. This involves taking an open-source language model (such as LLaMA or Mistral) and either providing the rubric as context for each evaluation or fine-tuning it with actual historical evaluation data to sharpen its criteria. All processing occurs on local on-premises infrastructure or within a sovereign cloud controlled directly by the public administration; no data ever leaves the secure environment.

When it makes sense

When the confidentiality of case files is a strictly non-negotiable requirement, when the organization demands total independence from external third-party providers, or when the sheer volume and maturity of the process justify investing heavily in in-house capabilities. This scenario is also highly relevant within the context of European digital sovereignty and trustworthy AI initiatives.

Auditability and traceability

Very high in potential, but highly demanding in execution. By retaining total control over both the model and the infrastructure, it is possible to engineer a comprehensive audit system tracking model versions, applied criteria, and the exact traceability of every algorithmic decision. However, building that rigorous audit framework is entirely an internal responsibility; it does not come pre-packaged by a vendor. It requires an explicit, dedicated effort in AI governance.

Advantages

• Data never leaves the organization's secure environment.
• Complete independence from external tech providers.
• The model can be continuously fine-tuned over time using the organization's own real-world data.
• Strong alignment with European frameworks for digital sovereignty and trustworthy AI.
• Much greater control over algorithmic bias and model behavior.

Limitations

• Requires a highly specialized technical team (data engineering, MLOps).
• Demands local on-premises infrastructure with significant computing power.
• Model maintenance and continuous updating become entirely internal responsibilities.
• Validating outputs and controlling biases without external vendor support is inherently more complex.
• The time-to-deployment to achieve an operational system is the longest among all five scenarios.

Cross-Cutting Challenges

These are the structural challenges that any implementation must address, regardless of the chosen scenario. It is highly advisable to tackle these foundational issues before making any technological decisions.

1. Defining Evaluation Criteria‍

AI evaluates only as well as the criteria it is given. If the evaluation rubric is ambiguous, subjective, or incomplete, the AI will amplify that ambiguity rather than resolve it. It is absolutely necessary to invest time in formalizing criteria with rigorous precision, including explicit examples of high, medium, and low scores for every single dimension. This groundwork precedes any technological development and is likely the most important step in the entire process.

2. Validating Evaluation Quality

Before placing trust in AI, you must explicitly demonstrate that it evaluates effectively—and this is neither self-evident nor a one-time fix. Validation is not merely a pre-deployment formality, but a continuous control process. AI models, evaluation rubrics, and the nature of grant calls change over time; furthermore, third-party provider models are often updated without notice. Consequently, a system that works perfectly today can degrade silently tomorrow without anyone noticing.

The strongest recommendation is to build a baseline reference set of proposals that have already been evaluated by human experts, periodically compare AI-generated scores against human scores, and meticulously measure the degree of agreement over time—not just during the initial launch phase.

3. Bias Control‍

An AI system can easily replicate, or even heavily reinforce, biases present in historical evaluations. Essential mitigation measures include strictly separating objective criteria from subjective ones, establishing mandatory human review protocols, and conducting periodic audits to verify whether certain types of entities or sectors are systematically receiving disproportionately different scores.

4. The Human in the Loop‍

In none of the five scenarios does the AI make the ultimate final decision. The human case officer (técnico gestor) always retains the final say. This is not merely a legal safeguard; it is the fundamental mechanism through which the system improves over time. It is vital to design workflows so that this human validation process is both functionally convenient and meticulously documented.

5. Adaptation to Different Types of Grant Calls‍

Grant calls vary enormously in their specific criteria, required proposal formats, and the exact weighting of each dimension. The best practice is to begin with a single specific line of funding as a controlled pilot, validate those initial results, and then scale the solution progressively.

6. Transparency for the Citizen‍

In public administration, official resolutions must be fully justifiable and subject to legal appeal. If an AI system generates a score, the administrative case file must clearly and legibly record exactly which criteria were applied and the explicit reasoning behind why they were applied.

7. Regulatory Framework: Risk Classification and Security

Before any deployment, two major legal frameworks must be analyzed. The European AI Act requires organizations to determine whether the system is classified as "high risk"—which is highly probable when a grant call directly affects individuals—entailing strict, reinforced obligations. While human oversight is legally necessary, it does not, by itself, exempt the system from full compliance. Meanwhile, the National Security Framework (ENS) applies to any IT system processing public records, dictating exactly where and how that data is hosted and processed. In both cases, this critical assessment falls to internal legal and security departments and must be completed well before selecting any underlying technology.

The five scenarios outlined above are not mutually exclusive alternatives, but rather evolutionary stages in a broader progression. The most reasonable approach is to start with Scenarios 1 and 2 to formalize criteria and validate the core concept at a very low cost; consolidate operations in Scenario 3 once the evaluation model is thoroughly proven; and reserve Scenarios 4 and 5 for when sheer volume, strict audit requirements, or absolute confidentiality fully justify the leap. Ultimately, the most profitable investment is never the technology itself: it is the precise definition of the evaluation criteria and the continuous validation that the AI is scoring reliably

‍