Blog

Insights

Analysis, method and real cases on adopting AI in the enterprise: what we learn working on our clients’ processes.

September 4, 20263 min read

Qwen3.8-27B: an open model that runs on a single GPU

Alibaba released an open-weight model with 27.8 billion parameters, competitive with much larger models and deployable on a single GPU. What that changes for anyone evaluating an on-premise or hybrid architecture.

Architecture
Security

Read the article

September 2, 20263 min read

How we measure 92% automatic payment reconciliation

92% of transactions reconciled automatically, on a sample of roughly 15,000 manually verified payments. The method behind the number, and why raising automation can worsen precision.

Data
Finance

Read the article

August 31, 20264 min read

How we measure 86% accuracy extracting data from CAD drawings

86% accuracy isn't a number you just state: it's a scope, a sample and a metric that can be verified. The method behind the published figure for the Gruppo SAG case.

Data
Industry

Read the article

August 29, 20263 min read

In voice, knowing when to stay silent matters more than in text

90% of calls handled without human intervention. The number that matters most, though, is a different one: in 70% of uncertain cases the system chose to route the call instead of guessing.

Architecture
Security

Read the article

August 27, 20263 min read

Why we don't publish a single percentage for semantic search

A real project produced two different numbers: 86% on one phase, no single percentage on the other. It isn't reticence: it's a consequence of how evaluating a RAG system actually works.

RAG
Finance

Read the article

August 25, 20263 min read

How to evaluate a system that has no single right answer

85% of recommendations judged relevant, 65% catalog coverage. Two numbers, not one, because on a recommendation engine there is no correct answer to compare against.

Data
Method

Read the article

August 23, 20264 min read

Evaluating a RAG in three layers: retrieval, generation, faithfulness to sources

A RAG system that gets something wrong can fail at three different points, and each one calls for a different measurement. Conflating them produces a single number that doesn't say where to intervene.

RAG
Method

Read the article

August 22, 20266 min read

How to choose an AI consulting firm in Italy

Four types of providers, the signals that separate a real production system from a demo, and the AI Act and Legge 132/2025 obligations a serious provider needs to handle.

Method
Compliance

Read the article

August 6, 20265 min read

It's not the model. It's the data.

In most AI projects the bottleneck isn't the model but the state of the company's documentation: what can be cleaned up during the project and what can't.

RAG
Method
Data

Read the article

August 4, 20265 min read

What RAG actually means, explained without acronyms

What RAG (Retrieval Augmented Generation) actually means: how it works in three steps, why source citation is decisive in a regulated context, and where it falls short.

RAG
Method

Read the article

Futura AI — Chatbot or AI system? The difference that matters

August 3, 20265 min read

Chatbot or AI system? The difference that determines the return

A chatbot answers. An AI system integrates with company data and processes and produces a measurable result. The distinction sounds terminological. It isn't.

Method
ROI

Read the article

August 2, 20265 min read

Why most AI projects stall at the demo

A demo convinces in twenty minutes. A system has to work every day for years. The distance between the two is the real project.

Method
Production

Read the article

August 1, 20265 min read

What to ask an AI vendor before signing

You can tell a good vendor by the questions they ask about your process, not the technologies they list. Four areas that decide how a project turns out.

Method
Compliance

Read the article

July 29, 20265 min read

Prompt engineering isn't an individual skill, it's a design layer

Writing a good instruction helps one person. Designing a system's instructions is what makes the result repeatable, and it needs versioning, testing and governance.

Architecture
Method

Read the article

July 28, 20262 min read

Adopting AI without rewriting your processes

The fastest way to fail an AI project is to start from the technology. Our method starts from the process that already works.

Method
Processes

Read the article

July 27, 20266 min read

How to build an AI roadmap in a complex organization

A credible roadmap doesn't list technologies: it lists processes in an order that accounts for real dependencies, and lets you check whether you actually got there.

Method
Integration

Read the article

July 24, 20264 min read

Beyond text: where Speech AI and Vision AI solve real problems

Not all company information is written down. Call transcription, technical drawing reading and visual inspection in production: where these work today, and under what conditions.

Architecture
Industry

Read the article

July 22, 20265 min read

Case study: extracting structured data from documents with no fixed format

Different documents, same information every time: how a data-extraction project defines what to automate and, more importantly, what to leave out.

Data
Industry

Read the article

July 19, 20265 min read

AI doesn't replace people: it redesigns processes around them

In projects that work, the job doesn't disappear — the part nobody wanted to own does. The line between task and role decides whether a project gets adopted or quietly bypassed.

Method

Read the article

July 17, 20265 min read

Data security in AI projects: seven questions to ask your vendor

In organizations with high compliance requirements, security isn't a contract clause: it's a selection criterion to verify upfront, in writing.

Security
Compliance

Read the article

July 14, 20265 min read

Quoting and procurement: where AI moves the margin

Quoting looks like judgment work, but the part that eats most of the time is reading and transforming documents. That's where response speed and margin are decided.

Industry
Method

Read the article

July 12, 20265 min read

Vector databases and knowledge graphs: two ways to represent knowledge

One finds what's similar. The other finds what's connected. How to choose between semantic search and a relationship graph, and when a hybrid architecture is worth it.

Architecture
Data

Read the article

July 10, 20265 min read

How to evaluate an AI system before it goes into production

The question isn't whether the system works, but on which cases, with what margin of error, and who verified it. How to build a test set, which metrics to use, and why you need a human baseline.

Method
Production

Read the article

July 7, 20265 min read

The cost nobody budgets for: maintaining an AI system

The first release is the cheapest part. Models, documents and processes keep changing: what to budget and put in the contract so you aren't caught off guard by the silent decay of an AI system.

Production
Method

Read the article

July 5, 20265 min read

Five public-sector tasks you can automate today

You don't need multi-year projects for a measurable effect: repetitive, documented, high-volume processes are enough. Five concrete tasks for public agencies, with prerequisites and indicators for each.

Public Sector
Method

Read the article

July 2, 20265 min read

AI governance: who is accountable when the system gets it wrong

Before asking whether AI can do something, an organization has to establish who answers for it when it does it badly. The minimum roles, documentation and system registry needed, even with limited resources.

Compliance
Method

Read the article

June 30, 20265 min read

The AI that talks to your business systems: integrating with ERP, CRM and records management

An AI system that doesn't access operational data stays a demo. The value comes from integrating with ERP, CRM and records management — and it's the line item most often underestimated in initial estimates.

Integration
Architecture

Read the article

June 27, 20265 min read

Hallucinations aren't an unpredictable flaw: they can be governed

A language model produces the most probable continuation of a text: without a source, it invents one plausibly. Hallucinations aren't a flaw to wait out — they're a structural property that's governed through the right architecture.

Security
Compliance

Read the article

June 25, 20265 min read

PNRR and EU funds: AI-assisted document monitoring

Reporting on PNRR funds isn't a question of willingness but of document volume and deadlines. AI-assisted control moves verification from the end of the process to the moment of upload, when fixing an issue still costs little.

Public Sector
Compliance

Read the article

June 22, 20265 min read

Enterprise search: why employees can't find what the company already knows

Traditional enterprise search looks for words; people look for answers. The gap between the two is measured in minutes lost every day, and in an archive people stop consulting.

Data
Security

Read the article

June 20, 20265 min read

Fine-tuning or RAG: an architecture choice, not a trend

Fine-tuning and RAG solve two different problems, not two degrees of the same problem. Confusing them, choosing based on budget rather than need, is the most frequent cause of expensive, low-value projects.

Architecture
Method

Read the article

June 18, 20265 min read

Why a few weeks of assessment are worth more than a year of experiments

Experimenting without a criterion produces enthusiasm and no decision. A few weeks of assessment produce a list of processes ranked by value and feasibility, and the discipline to exclude what isn't ready.

Method

Read the article

June 15, 20265 min read

Manuals, bills of materials and technical support: the knowledge the company already has

In many manufacturing companies the technical knowledge already exists, but it's scattered across PDFs, paper revisions and people close to retirement. The problem isn't producing it: it's making it queryable before the people who hold it are gone.

Industry
Data

Read the article

June 13, 20265 min read

KYC, AML and due diligence: where generative AI actually saves time

In verification processes, time is spent reading heterogeneous documents, not on the final decision. Where generative AI actually saves time in due diligence, and why traceability is a requirement, not a nice-to-have.

Finance
Compliance

Read the article

June 10, 20265 min read

On-premise or cloud: how to decide when data is sensitive

The right question isn't which infrastructure is more secure, but which perimeter you're able to govern and demonstrate. The real criteria for choosing between on-premise, dedicated cloud and public cloud when data is sensitive.

Architecture
Security

Read the article

June 8, 20265 min read

How to measure the return on a generative AI project

An AI project with no indicator defined before it starts will never have a demonstrable result at the end. How to build an honest baseline and tell adoption indicators apart from return indicators.

Method
Data

Read the article

June 5, 20265 min read

Document intelligence in public administration: from the protocol system to administrative determinations

Preparing a case isn't deciding it: a system can read, classify and prepare a file, but the signature and the responsibility stay with the official. Where document intelligence actually saves time in public administration.

Public Sector
Production

Read the article

June 3, 20266 min read

Agentic AI: when a system stops answering and starts acting

An assistant suggests, an agent acts: it reads, writes, decides the next steps. The question that matters isn't what an agent can do, but where it has to stop.

Architecture
Method

Read the article

June 1, 20265 min read

AI Act: what changes in practice for companies introducing AI

The Regulation doesn't ban artificial intelligence: it asks you to know what a system does, on what data, and who answers for it. What actually changes after the delay set by the Digital Omnibus, and why it's no reason to delay the work.

Compliance
Method

Read the article