Generative AI answers questions; Agentic Artificial Intelligence completes tasks. For a company today, value lies not in asking for a piece of text, but in delegating entire processes, with the certainty that what matters remains under human control. In this article, we explain without technicalities how Revelis builds reliable agents, when it is best to use cloud models versus internal on-premise models, and we show a concrete use case the automation of the tendering process where agents generate a measurable competitive advantage.
Generative Artificial Intelligence has now firmly entered business processes, but the real leap in value for companies is not asking for answers: it is delegating tasks. This is the promise of Agentic Artificial Intelligence, where autonomous software systems plan and execute tasks from start to finish by coordinating tools, data, and models. For an organization, however, autonomy also means responsibility: reliability, security, traceability, and cost control become indispensable requirements.
At Revelis, we support companies through this transition with a guiding principle: making AI actionable in a reliable, repeatable, and governable way. Below, we describe our vision for building agentic solutions including the use of on-premise Small Language Models the architecture that makes it possible, and a directly applicable use case example.
From Generative AI to Agentic Artificial Intelligence
With generative AI, the user formulates a request and gets an answer: text, code, a summary. With Agentic Artificial Intelligence, the user assigns a goal: the system breaks down the problem, decides which tools to use, executes the necessary steps, and returns an operational result. It is helpful to think of it as a virtual collaborator to whom you can delegate repetitive tasks that carry a time cost but do not require high-value decision-making.
This paradigm shift unlocks high-value use cases from back-office automation to software development assistance, from document management to the preparation of complex bids but it introduces a technical challenge that becomes decisive in regulated contexts: the unpredictability of language models.
The critical friction point of enterprise adoption: unpredictability
By their very nature, language models can give different answers to the same question. In a creative chat, this is an asset; in a business process, it is a risk. On a practical level, the impact is twofold:
- Compliance and control: Difficulty in faithfully reproducing an action taken in the past, logs that are not always complete, and obstacles in meeting regulatory and audit requirements (GDPR, ISO, SOC 2).
- Operational risks: Varying behavior for the exact same request, decisions that are not always traceable, and unpredictable costs.
Revelis’ vision stems from this premise: the autonomy of agents must be channeled into an architecture that separates decision-making (what to do) from execution (how to do it), making the latter reliable and repeatable.
Revelis’ Vision: Reliable Agent Orchestration
With Agentic Artificial Intelligence, Revelis’ agentic framework is structured into layers with clear-cut responsibilities. The core concept is simple: AI proposes, but it does not improvise. The sequence of actions that produces concrete results is defined by pre-established, logged, and verifiable procedures. In this way, for the same request, the system always executes the same steps in the same order a necessary prerequisite for a company to trust an agent.
Fig. 1 — The Revelis agentic framework: from security to execution, with pre-established procedures.
The architecture layers, in simple terms
- Guardrail (security): A filter that blocks irrelevant or non-compliant requests at the source, before they consume time or resources.
- Process Catalog (process memory): The list of tasks the system knows how to perform: for each one, it defines what data is needed and the correct procedure to follow.
- Semantic Router (routing): Understands the request expressed in natural language and routes it to the right task.
- Data Extraction: Extracts necessary information in an orderly fashion; if anything is missing, it asks for clarification instead of “inventing” answers.
- Execution Engine: Executes the procedure step-by-step, with the ability to resume where it left off and reconstruct every action for auditing purposes.
- Security Chain (fail-safe mechanism): When the system is not confident enough, it escalates through subsequent levels up to human intervention, never acting blindly.
- Memory/cache: Reuses decisions already made to reduce time and costs, without ever putting data at risk or repeating sensitive operations.
Fig. 2 — The three possible paths: full path, answer already in memory, and security block.
Two adoption scenarios, one architecture
The strength of this approach lies in its adaptability: the same architecture fits two very different starting points, without forcing the company to choose between reliability and pragmatism.
Fig. 3 — Two starting points for the adoption of agentic AI, governed by the same architecture.
A. Already structured contexts: leveraging existing assets
Organizations that have already adopted copilot suites and cloud models possess a valuable technological heritage, as well as established governance and security processes worth capitalizing on. In these cases, Revelis reuses the existing infrastructure and adds the orchestration layer on top: agents become reliable and traceable directly on the infrastructure the company already owns, enabling a fast rollout without requiring new hardware investments.
B. On-premise LLMs: Small Language Models (SLMs) for data sovereignty and cost control
SMEs and organizations that prefer to keep data within their own perimeter for privacy, compliance, or vendor independence can adopt Small Language Models run on-premise: compact models that, thanks to optimization techniques, operate on limited hardware. Integrated into the architecture, they power reliable and repeatable execution of tasks, with predictable costs and full data control. This is where Revelis’ model evaluation work becomes a tangible advantage.
On-premise Small Language Models (SLMs): models evaluated by Revelis
Not all compact models are created equal. Revelis expanded its testing to eleven coding-oriented Small Language Models, measuring the correctness of the generated code against a sample set of test problems. The minimum benchmark set by Revelis was to solve at least 30% of the sample problems: four out of eleven models passed this threshold.
The highest score was achieved by Qwen3-Coder-Next (60% of problems solved), which is currently being tested within a coding agent using a context window of up to 200K tokens. This is followed by Gemma4 26B-A4B (48%), top of the list for upcoming agent testing. Qwen3-Coder-30B-A3B-Instruct (40%) has already been integrated into a coding agent, but in practice proved to be slow, heavy, and penalized by a necessarily restricted context window (32K tokens).
Fig. 4 — Positioning of the tested Small-LLMs: requested resources vs. suitability for use with coding agents (★ also tested with coding agents).
| Model | Features | Ideal Application Contexts |
| Qwen2.5-Coder 7B (base + instruct fine-tuned) | Compact 7B parameter model, available as both base and instruct versions optimized for coding. Reduced footprint, executable on midrange workstations or servers. In Revelis tests: 8% of problems solved. | Assistance in code writing, local lightweight autocompletion and refactoring. |
| Qwen3.5-Coder 9B (Q4_K_M) | A 9B parameter evolution with improved reasoning capabilities on code, maintaining a good balance between quality and required resources. In Revelis tests: 28% of problems solved. | Development assistant for the team, generation of documentation and medium-complexity codebases. |
| DeepSeek-Coder-V2-Lite-Instruct (16B, ~2.4B active) | “Lite” version specialized for code, Mixture-of-Experts architecture with few active parameters per token. In Revelis tests, it did not solve any problems from the sample (0%): an underwhelming result, to be verified in further experiments. | Local and edge inference with minimal resources — coding suitability to be re-confirmed. |
| Qwen3-Coder-30B-A3B-Instruct ★ | Mixture-of-Experts architecture (~3B active parameters on 30B): high quality with contained inference costs. In Revelis tests: 40% of problems solved, above the set minimum threshold (30%). When integrated into a coding agent (OpenCode, Cline) it proved to be slow and heavy, penalized by a necessarily reduced context to 32K tokens. | A good benchmark for non-agentic flows; direct use in agentic scenarios to be carefully evaluated for latency and context. |
| Gemma4 26B-A4B | A model from the Google Gemma family with a Mixture-of-Experts architecture (~4B active parameters on 26B). In Revelis tests: 48% of problems solved — the second best result after Qwen3-Coder-Next. | Entirely local deployment on contained hardware; first Revelis candidate for the next tests in a coding agent. |
| Codestral 22B | A model dedicated to code with broad multilingual coverage and support for contextual completion, designed for heterogeneous stacks. In Revelis tests: 6% of problems solved. | Complex and multilingual repositories, code completion across multiple technologies — results to be confirmed. |
| Devstral 24B | Agentic-oriented software development model, designed for use in coding agents on real repositories. In Revelis tests: 10% of problems solved. | Structured development environments; agentic suitability to be re-evaluated in light of results. |
| North-Mini-Code (Cohere) | Enterprise-class compact model from Cohere’s North platform, designed with a focus on data security and privacy. In Revelis tests: 6% of problems solved. | On-premise enterprise deployment with compliance requirements, reduced footprint — coding performance to be monitored. |
| Qwen3.6 27B | A 27B parameter evolution of the Qwen family dedicated to coding. In Revelis tests: 18% of problems solved. | Mid-complexity development assistant for team and medium-sized projects. |
| Qwen3.6 35B-A3B | Mixture-of-Experts architecture with 35B parameters (~3B active). In Revelis tests: 40% of problems solved, above the set minimum threshold (30%). | Alternative to Qwen3-Coder-Next for agentic flows, to be adopted if latency and context result as sustainable. |
| Qwen3-Coder-Next ★ | 80B parameter Mixture-of-Experts architecture (~3B active): the model with the best results in Revelis tests so far. In Revelis tests: 60% of problems solved. | More demanding agentic scenarios, multi-step autonomous activities, and tool orchestration; currently in testing phase in a coding agent (OpenCode) with context up to 200K tokens. |
A concrete use case: automating the tender process with Agentic Artificial Intelligence
To understand where agents generate a real competitive advantage, it is worth looking at a process that many companies know well: participating in public tenders. It is the perfect example of a complex, high-value process involving multiple stakeholders, where most of the work today is performed manually.
Why it is a difficult (and expensive) process without Agentic Artificial Intelligence
A structured organization may participate in hundreds or thousands of tenders every year. Each tender brings an impressive volume of documentation: technical specifications, tender rules, and annexes that often exceed 200 pages, sometimes featuring thousands of individual “lots” to evaluate one by one. All of this must be read, interpreted, and turned into a bid within tight deadlines and under rules that vary from entity to entity.
The work touches multiple departments: the tenders office, sales, technical experts, finance, management control, all the way to executive sign-off. Today, coordination happens largely via email and spreadsheets, triggering a domino effect: the same specification document is read independently by multiple people, multiplying the hours spent. Above all, the margin for error is minimal: a single misread requirement or a non-compliant bid means immediate disqualification and a lost business opportunity.
How Agentic Artificial Intelligence intervenes
Revelis addresses this process using Agentic Artificial Intelligence. Thanks to a multi-agent system, an orchestrator coordinates multiple specialized agents, each responsible for a specific phase. The output of one agent feeds into the next, culminating in the draft proposal. No agent “decides” on its own: each one prepares, verifies, and proposes, leaving the decisions that matter to human experts.
Fig. 5 — The tender process governed by agents: an orchestrator coordinates six specialized agents, with human validation and signature/submission handled by people.
- Scouting & Triage: Monitors portals, identifies relevant tenders, and assigns a feasibility/suitability score, allowing attention to focus where participation is truly worthwhile.
- Call Analysis (Tender Document Analysis): Reads the technical specifications and tender rules, extracting lots, requirements, criteria, deadlines, and grounds for exclusion—always citing the source document and page number so every piece of information is verifiable.
- Checklist & Envelopes: Generates the list of required documents for the administrative, technical, and financial envelopes, flags what is already prepared, and assigns tasks to the right team members.
- Pricing Support: Combines the tender formula, cost structures, and competitor estimations to propose a reasoned price alongside various scenario simulations. It proposes and explains, but the final decision remains human.
- Quality Simulation: Steps into the shoes of the evaluation committee: it applies the scoring rubric, estimates the proposal’s score, and highlights how to bridge any gaps to maximize points.
- Drafts & Memory: Prepares initial proposal drafts using company templates—eliminating the need to start from scratch—and builds a historical market memory (prices, outcomes, similar tenders) valuable for future bids.
The competitive advantage for companies using Agentic AI
Employing Agentic Artificial Intelligence means that the combined effect of these agents isn’t just about “doing things faster” it’s about competing better. By drastically reducing preparation time, companies can participate in more tenders including those currently discarded due to lack of time and focus on the ones with the highest chance of winning. Reducing compliance errors protects against disqualifications. Structured preparation helps avoid leaving money on the table: a more accurate bid raises the average winning price. And historical memory transforms even lost tenders into reusable market intelligence.
Governance and controlled evolution with Agentic Artificial Intelligence
Deploying Agentic Artificial Intelligence means utilizing an enterprise agentic system that is not static: it must grow without losing control. In the Revelis framework, every new capability starts as an activity in the process catalog and follows a governed path: drafting, overlap check, “shadow” testing (running in parallel without execution, to measure reliability on real-world cases), approval, and production deployment. Every human intervention thus becomes a learning opportunity, enriching the system in a fully traceable manner.
The result is an AI that observes, logs, and improves: every step is documented for the benefit of control, security, and continuous optimization.
The Revelis vision, in summary
Introducing Generative AI and Agentic Artificial Intelligence into a company does not mean giving up control. It means designing an architecture where reliability stems from the separation between decision and execution, security precedes all processing, and execution follows defined procedures rather than improvised choices.
Whether starting from an already structured copilot ecosystem or an on-premise setup based on Small Language Models, and whether the goal is automating back-office processes or winning more tenders, the objective remains the same: reliable, secure, and governable agents capable of generating measurable value.
It is our interpretation of the principle that has always guided Revelis: making AI actionable.
Mini-glossary for non-experts
- Generative AI — Systems that produce content (text, images, code) in response to a prompt.
- Agentic AI — Systems that go beyond simply answering questions to execute delegated tasks by coordinating multiple steps.
- Agent — A specialized software “virtual collaborator” assigned to a specific task; multiple agents can cooperate with each other.
- Orchestrator — The coordinator that assigns tasks to individual agents and synthesizes their results.
- LLM / Small-LLM — The language model that “reasons” over text; Small-LLMs are compact versions capable of running on limited hardware.
- On-premise — Technology installed and managed directly within the company’s premises rather than in the cloud, keeping data in-house.
- Human-in-the-loop — An approach where humans validate and approve what the AI proposes, ensuring full control is maintained.
Want to evaluate Agentic AI in your company? Revelis supports enterprises and SMEs in designing agentic solutions from model analysis to orchestration architecture, in the cloud or on-premise. Write to us at info@revelis.eu
Authors
Achille Abritta – Enterprise Solution Engineer at Revelis
Francesco Cupello – Enterprise Solution Engineer at Revelis
Luigi Granata – Enterprise Solution Engineer at Revelis
