The lab
Explore text workflows, answers with sources and a first direction for an audit. Each experiment explains how it works; search and audit include examples and rules you can use without a model.
“Available” means the experiment can be used now. Search examples are written and verified by hand; the example audit card follows predefined rules. Generative features are separate from the examples and explicitly identified. Their results should be checked.§
Available
From an order email to structured data for your systems: try an order, a quotation request or an overdue reminder.
Start with the preloaded examples. This is the same example available on the home page.
Transparency notice, Article 50 EU AI Act: Experiment with an AI system — the data you paste is not stored.
Available
Four questions to explore how we work. Each answer cites the pages that support it, so you can check the sources.
The four answers are ready; you can also write a question of your own.
Transparency notice, Article 50 EU AI Act: Example answers are written and verified by hand. Answers to free-form questions are generated by AI and should be checked.
How much does it cost to start with an AI integration?
ANSWER PRELOADED ANSWER
There is no single price before the process is understood. It starts with a one-week audit that maps data and systems and delivers a roadmap; you then decide whether to stop or move on to a pilot. Each service can also be purchased on its own, and we do not provide flat-fee quotations for processes that have not been mapped.
Sources
Where is our data processed, and how long do you retain it?
ANSWER PRELOADED ANSWER
The privacy policy describes the data processed through this site, its recipients and retention periods. In a project, data, providers, residency and retention periods are agreed and disclosed in advance in the contract and DPA; we do not use the data to train models. The preloaded questions send nothing; the free-form question sends the text to the model provider, as stated next to the field, and the site does not store it.
Sources
What happens when the model gets it wrong?
ANSWER PRELOADED ANSWER
The model can be wrong, so error handling is part of the project: human review where errors are costly, continuous quality measurement and the ability to roll back. In assistants, an answer without supporting evidence is not returned: the system says that it does not know and escalates to a person. The criteria are set before the pilot; if they are not met, the work stops.
Sources
How long does it take to see something in production?
ANSWER PRELOADED ANSWER
The first draft roadmap arrives within 24 hours of the kick-off meeting, and the audit closes in one week. The pilot on one real process takes 4–6 weeks; it moves into production only after meeting written criteria. The final date depends on the process: first we map it, then we estimate it.
Sources
The sources open the pages of this site used to verify each answer.
Available
Choose your sector, process and volume. The card suggests a first step and the conditions for assessing whether it pays off.
The example card is ready; the personalized one is generated on request.
Transparency notice, Article 50 EU AI Act: Example cards follow predefined rules. The personalized card is generated by AI and should be reviewed with a technical specialist.
Typical friction in the sector
ManufacturingOrders, bills of materials and technical documents often move between email, PDF and the management system: look for the point where someone retypes or compares the same data.
ServicesRequests, cases and hand-offs between tools accumulate when knowledge remains with individuals: look for waiting, forwarding and repeated answers.
CommerceCatalogues, orders and customer requests change channel but not content: look for duplication across email, CRM, e-commerce and the management system.
First step for the process
Orders and quotationsCollect a sample of real orders and quotations, define the fields to extract and leave final confirmation to a person.
Documents and casesChoose a single document type, define mandatory fields and exceptions, then measure time and errors before the pilot.
Customer requestsStart with recurring requests, specify the permitted sources and have a person approve answers where errors are costly.
Viability threshold
A few dozen a monthWith a few dozen cases a month, first check whether simplifying the workflow or connecting the systems would be enough: AI is not automatically the first step.
A few hundredWith a few hundred cases, a pilot can measure whether repetition and errors justify automation: decide on the evidence, not on volume alone.
More than one thousandAbove one thousand cases, prioritise exceptions, human review and monitoring: volume increases the value of testing, but does not justify shortcuts.
Use these three lines as a hypothesis to verify against your processes, data and systems: they do not replace an audit.
The three choices stay in your browser and are neither sent nor stored.
THE LAB RULES
Every experiment speaks both Italian and English and states how it works before the interaction, including when it does not use AI: this is our rule, and goes beyond the law. For experiments that use AI, transparency towards those who interact with them follows Article 50 of the AI Act.
Every generated output is marked as such, both visibly and in a machine-readable form.
No experiment pretends to be a person. If it communicates, it is a system, and says so.
The data collected by an experiment is listed in the privacy policy, together with the stated retention periods. The extraction experiment stores nothing.
An experiment goes online when it works, not when it can be announced. Modes without a model exist only when they can genuinely be used, and say so before the interaction.
The lab experiments cover our standard processes. You tell us about yours: the first roadmap arrives within 24 hours of the kick-off meeting.
Talk to us