The draft
First the summary, then the source of each piece of data
The bar says how many items need attention. The Source column separates what is calculated from what the AI has read.
Open in the prototype →Case study · Product design for AI · Product and concept
Nyra
An AI system that is configured around the documents, rules and roles of each company. In the case study it prepares the tender files of a construction company, and the person responsible checks them item by item before signing.
Live product · ai.nyra.itReview and configuration · conceptFictional company and data
Open the prototype (in Italian)
- Where the data comes fromRead by the AI, calculated, company records or generated.
- The field that was readHighlighted on the scan, with its confidence.
- The reason for the doubtThe clues and the most likely alternative.
- Approval blockedWhile an item still needs to be resolved.
The problem
A person who signs a file prepared by AI must be able to trust every item.
The idea
Every piece of data says where it comes from and how certain it is. The code does the calculations.
The result
A system that can be configured for any company, and a prototype with 21 states to try.
Context
Agents already work on documents
43%
of European tech professionals delegate documentation to AI. It is the task they delegate most often.
Shakers, State of European AI Builders 2026, p. 1273%
have at least one AI agent or automated workflow in real work.
Shakers, State of European AI Builders 2026, p. 12The weak point is the move into production. For 37% of people who worked on AI projects in the last year, less than half of them reached production.
All 12%More than 75% 23%About half 28%Less than 25% 28%None 9%
Shakers, State of European AI Builders 2026, pp. 9 and 15. Base: 878 respondents with AI projects in the last 12 months. “In production”: used by real people or systems and still active after 90 days.
Read moreMore data from the report, and why now
What people delegate to agents
Source: Shakers, State of European AI Builders 2026, p. 12. Up to three answers per person.
Why now
In August 2026 Shakers surveyed 1,035 tech professionals in 23 European countries. In the Italian sample, which is small and so only indicative, 87% have at least one agent in real work and 80% see European regulation as a professional advantage (p. 33, n = 70).
An expert quoted in the report notes that the barrier to adoption is cultural, and is about internal readiness more than budget (p. 18).
The problem
Delegating the file to AI without losing control over what you sign.
Hypothesis to be validated Fictional scenario
ToolFills in the DGUEfrom the company records
ToolAnalyses the bill of quantitiesagainst the regional price list
ToolGenerates the POSstarting from the PSC
Today each tool works on its own. It does not see the whole file, it cannot tell a calculated number from one read on a scan, and it cannot show how the file was prepared.
- …show the person who signs where each piece of data comes from and how much to trust it?
- …make every correction update the calculations that depend on it automatically?
- …show management how much work the system does and who checks it?
Read moreFor whom: proto-persona, hypotheses and secondary users
Without interviews, I worked with a proto-persona and a map of hypotheses, stated as such. In the prototype the company is invented: Costruzioni Lèmura, with an equally invented tender from the Municipality of Valmarina.
Proto-personaHypothesis
Giulia, tender manager at a construction company
- She prepares the files for public tenders, often with only a few days available.
- She is responsible for what goes into the file; the legal representative signs it.
- She works with a site surveyor, a labour consultant and the contracting authorities.
“When a tender is published, I want to know at once what is missing, what is expiring and what is uncertain, so I can send a complete file without checking everything again from the start.”
The hypotheses and how I would test themTo be validated
- Time is lost re-entering data and checking expiry dates.Observe a real file and time each phase.
- The person who signs does not trust numbers without a source.Test two versions of the item, with and without the data source.
- The fear of an expired document leads people to check everything again.Ask which errors led to requests for extra documents or to exclusions.
- Scans are the main cause of reading errors.Classify the corrections made in the first month.
Secondary users. The site surveyor completes the POS but does not approve it. The legal representative signs. The labour consultant and the contracting authority receive the requests. Management looks at adoption and reliability and decides whether to renew the subscription.
Where it comes from
Two projects, one system
Nyra is the brand that LL Studio and PV Design use to present a single offer. I combined its real product with a review concept.
Nyra AI ToolsIn use
Four tools for constructionRules already tested with a pilot company.
Review concept
The experience of the person who signsItem-by-item review, with the source always in view.
Nyra
A single, configurable fileIt uses the existing tools and shows where each piece of data comes from.
Read moreWhat comes from where, item by item
From Nyra AI Tools · real
- The tools for the DGUE, bill of quantities and POS
- The AI interprets, the code does the calculations
- No missing data is invented
- Human validation cannot be skipped
- Every function is tested by breaking it on purpose
- A subscription: we sell a system, not hours
From the review concept
- Item-by-item review, with source and page
- Confidence that changes what you can do
- “Why” panel and corrections that teach
- Requests prepared by the system, sent only by a person
- Roles and permissions
- Sample checks that unlock with the data
- A statement on how the document was prepared
New in Nyra
- A single file that uses all the tools
- The source of every piece of data
- “Exact” for calculations instead of confidence
- Corrections that recalculate what depends on them
- The test bench visible to the people who use the system
- A configuration for each company
The idea that holds product and process together
The AI interprets. The code calculates. A person checks and signs.
Read by the AIMedium
The AI interprets. It reads documents and scans, and says how certain it is.
Calculated=Exact
The code calculates. Guarantees, deadlines and comparisons follow fixed rules.
Approved by the manager
The person decides. Uncertain items are opened one by one, and approval cannot be skipped.
Simplify Review · 2025
Record pair 14Similarity 92%
MergeKeep both
The number says how much, but not why.
Nyra AI Tools · todayDiagram
DGUE completed52 fields
☐ I have checked the data and I take responsibility for it
Download
The confirmation covers the whole document at once.
Nyra · concept
ISO 9001 certificateRead by the AITo check
I read “10”, but the shape is compatible with “01”. The most likely date is 31/01/2027.
It is correctCorrect it
Attention goes only where it is needed.
Read moreThe six design principles, with sources
From Nyra AI Tools · Calculation
The AI interprets, the code calculates
The guarantee, the deadlines and the bill of quantities analysis show “Exact”; confidence applies only to what the AI has read.
In the interface: the data source and the “How I calculated” box.
From Nyra AI Tools · Data
No invented data
If a document is missing, the field stays empty and the system says where it looked.
In the interface: the “Missing” state and “Where I looked”.
From Nyra AI Tools · Control
Human validation cannot be skipped
Uncertain items cannot be confirmed all at once. A person sends the requests.
Among the barriers to production, the report mentions unclear responsibilities (p. 15).
From Nyra AI Tools · Tests
You test it by breaking it on purpose
Before every update the system runs the tests with known results again, and the result is visible.
In the interface: the test bench in Adoption and reliability.
From the concept · Confidence
Confidence changes what you can do
An item with medium or low confidence must be opened, and approval stays blocked until it is resolved.
In the report, an expert says they look for people who have learnt to delegate with a method (p. 18).
From the concept · Transparency
Every item shows its source
File, page and the field that was read, highlighted. The file goes out with a statement on how it was prepared.
Transparency and explainability in design (Adigital in the report, p. 26).
Today Nyra AI Tools runs on Base44. The concept also includes processing on the company’s own server: trust and privacy come first among the European advantages named by professionals (50%, Shakers, State of European AI Builders 2026, p. 43, base 1,039).
The prototype
Six moments when the AI asks for trust
From the tender notice to the approved file: 21 states you can navigate, no empty screens. Each image opens the prototype at the right point. The prototype is in Italian.
Uncertain AI
The doubt, explained on the document
The field that was read is highlighted on the scan, with the clues behind the doubt and the most likely alternative.
Open in the prototype →
Correction
A correction that spreads
You can undo it with one click, it stays in the log, and it automatically recalculates the guarantee that depends on that date.
Open in the prototype →
Exact calculation
The calculation is shown, not estimated
Calculated numbers have no confidence level: they have the step-by-step calculation and the rule from the tender specifications.
Open in the prototype →
AI without a source
If it cannot find it, it says so
No invented documents: the system shows where it looked and prepares the request to the RUP, which a person sends.
Open in the prototype →
Approval
The statement comes from the log
Before approving, the manager reads what the system did and what she checked herself. Then the file goes for signature.
Open in the prototype →Around the flow
Scroll sideways · each screen opens the prototype
Five components for AI
CalculatedRead by the AICompany recordsGenerated
Data source
How much attention a piece of data deserves, before you read it.
=ExactHighMediumLow
Confidence or exact calculation
Bars and a word: you can read it without colours too.
disciplinare.pdfp. 14
Source citation
File and page, with the field highlighted.
Why I have a doubtWhere I looked
“Why” panel
Answers always in the same order.
Corrected by youUndo
A correction that teaches
It can be undone, it is logged, and it recalculates what depends on it.
The best way to understand it is to try it: correct a date and watch the guarantee change.
Try the uncertain dataOne system, every company
The construction company is the first case, not the limit
Every company prepares repetitive files that someone must sign. Nyra is configured around its documents, its rules and its roles; the way of checking stays the same.
The prototype caseConstruction companyTender file and POS
Quality officeAgri-food consortiumCompliance file for large retail chains
Client filesProfessional firmApplication for a call for funding or a loan
Operations officeLogistics companyTransport documents, vehicles and drivers
Purchasing officeManufacturing companyQualification of a new supplier
Illustrative examplesThey are not real clients. The consortium is the scenario of the first concept.
The configuration is visible: which files the system prepares, which sources it reads from, which rules the code applies and who approves.
Open the configuration →Changes from company to company
Stays the same everywhere
Read moreFive configurations compared, and how a configuration is built
| Company and office | File | Documents and sources | Fixed rules | Who approves |
|---|---|---|---|---|
| Construction company · tender officeThe prototype case | Tender file and POS | Tender notice, tender specifications, certificates, DURC, PSC, regional price list | Provisional guarantee, DURC validity, deviation from the price list | Tender manager; the legal representative signs |
| Agri-food consortium · quality office | Compliance file for large retail clients | Members’ certifications, lab tests, technical data sheets, packaging declarations | Certificate expiry dates, limits set by the client’s specifications | Quality manager |
| Professional firm · client files | Application for a call for funding or a loan | Financial statements, company register extracts, quotes, client declarations | Requirements and thresholds of the call, eligible costs, deadlines | The professional who signs the application |
| Logistics company · operations office | Documents for a transport, and fitness of vehicles and drivers | Transport documents, vehicle inspections, driving licences and qualifications | Expiry dates of inspections and qualifications, client requirements | Operations manager |
| Manufacturing company · purchasing office | Qualification of a new supplier | Certifications, safety data sheets, declarations of conformity | Certificate validity, minimum requirements of the specifications | Purchasing manager |
How a configuration is built
- A real fileHow the company prepares it today: documents, checks, steps.
- Sources and documentsArchive, email and management software, and what the system can read.
- Rules in the codeDeadlines, thresholds and calculations, each with its source.
- Roles and reviewWho prepares, who approves, who signs; review by function.
- Test benchTests with known results, run again after every change.
The method
I also designed the way of working with AI
I gave my process the same rules as the product: the AI proposes, I check and decide, and every step leaves a trace.
44pages of the report read by the AI and checked number by number
4problem options assessed on 7 criteria before choosing
21prototype states tested with real clicks in an automated browser
- Structured briefThe AI’s role, context, selection criteria and honesty rules.
- CheckpointsAt the end of each phase I decide; the AI does not continue without my choice.
- Source checksEvery number compared with its page and with the appendix.
- Guided generationPrototype and images from my specifications: flow, states, rules.
- LogWhat the AI did, what I decided, how I checked.
Structure first, then the interface
Hierarchy, states and actions before colour and typography. Drag to compare.
Read moreWhat the report check found
The report text uses 1,035 responses, the appendix uses 1,039. Some numbers change by one point, and in one case the text rounds 32% to “half”.
| Data | Text | Appendix |
|---|---|---|
| Weekly use of ChatGPT | 70% · p. 11 | 71% · p. 40 |
| 1 to 2 agents in real work | 55% · p. 12 | 56% · p. 40 |
| No projects in production | 9% · p. 15 | 10% · p. 41 |
| Penalised after declaring the use of AI | 6% · p. 25 | 5% yes, 6% not sure · p. 42 |
| Time spent mostly directing the AI | “half” (32%) · p. 28 | 32% · p. 41 |
The rule I adopted
- I quote the numbers from the text with their page.
- Data found only in the appendix with pp. 40–43 and base 1,039.
- I do not use the numbers where the text and the appendix differ.
- 32% stays 32%, never “half”.
Read moreHow I chose the problem: four options on seven criteria
| Criterion | A · Pilot → adoption | B · Agent supervision | C · AI literacy | D · B + adoption + transparency |
|---|---|---|---|---|
| Strength of the evidence in the data | 4 | 3 | 3 | 4 |
| Relevance for people hiring on Shakers | 3 | 5 | 2 | 5 |
| UX depth I can show | 3 | 5 | 3 | 5 |
| Feasibility in the time available | 2 | 4 | 3 | 3 |
| Link with LL Studio | 5 | 4 | 4 | 5 |
| Value for a hybrid profile | 4 | 2 | 5 | 5 |
| What it adds to my case studies | 3 | 3 | 3 | 5 |
| Total out of 35 | 24 | 26 | 23 | 32 |
The first concept applied D to the quality office of an agri-food consortium. When I combined it with Nyra AI Tools, I moved the scenario to tenders and safety in construction, where the product is already in use.
Read moreThe log, phase by phase
| Phase | What the AI did | What I decided | How I checked |
|---|---|---|---|
| Brief | It reread the brief and flagged the data to check before using it. | The AI’s role, selection criteria, honesty rules, checkpoints. | A checkpoint at the end of each phase. |
| Report | It read 44 pages as images and compared every number with the text and the appendix. | The citation rule and the data to exclude. | Every number on this page shows its page in the report. |
| Problem | It assessed four options on seven criteria. | Combination D and a concept scope: no interviews, persona and scenario stated as hypotheses. | Scores reviewed after a full reading of the report. |
| Merge | It compared the concept with Nyra AI Tools and moved the scenario to tenders and safety in construction. | To merge the two projects under the name Nyra and to start from the principles of the real product. | No invented product usage data. |
| Prototype | It wrote HTML, CSS and JavaScript for 21 states and five components for AI, with no empty screens or fake buttons. | Critical flow, required states and rules: the system does not sign, does not send and does not invent data. | Full path in an automated browser; calculations redone by hand; contrast above 5:1; no horizontal scrolling at 1280, 768 and 390 px. |
| Page | It produced screenshots, mockups and the layout in the series of my case studies. | Texts, microcopy and data, reviewed before publication. | A “fictional” or “to be validated” stamp on every piece of data that is not real. |
Adoption and validation
A subscription is renewed if the value is visible
AdoptionTarget: 80% of tender files with Nyra at 90 days
QualityTarget: zero items approved without a source
TimeTarget: time per file cut in half
Project targets, not results. The 90 days come from the report’s definition of production (Shakers, State of European AI Builders 2026, p. 9).
To do
The test plan is ready, the results are not yet
Five tasks with 5–8 people from tender offices, then the AttrakDiff questionnaire. Target: pragmatic quality of at least 5 out of 7.
Read moreAll the targets and the five test tasks
Adoption targets
- Target: 80% of tender files with Nyra at 90 days, measured on the files sent.
- Target: zero items approved without a source.
- Target: tender office and site active every week.
- Target: halve the time per file compared with the first month.
- Target: all tests passed before every update.
- Target: fewer corrections per function, the condition for sample checks.
The five tasks
- Find the items that need your review and open the first one.
- Correct the date on the ISO certificate and check the guarantee.
- Ask the RUP for an appointment for the site visit.
- Choose the price list item for the screed and explain why.
- Approve the file and check the attached statement.
I will measure completion, errors, times and AttrakDiff, as in my other case studies. Targets: at least 4 tasks out of 5 without help, no uncertain item confirmed without opening it, everyone can tell calculated data from data read by the AI, pragmatic quality of at least 5 out of 7 (reference: 5.38 for Simplify Review).
What I learnt
Calculation and interpretation do not deserve the same level of control
A real product is worth more than a concept
Rules already tested, and a sector where an expired document costs you the tender.
The data source guides attention
Saying whether a number is calculated or read takes the eye where it is needed.
Checking sources is design
The differences in the report only came out by comparing every number with the appendix.
- Usability tests with 5–8 people from tender offices
- Interviews to validate the proto-persona and the hypotheses
- Bring the single file into Nyra AI Tools
- Try the configuration in a second sector
Now try to break it
Correct a date, change role, upload a file the system does not recognise. Nyra must hold up, and always tell you why.