Applied AI
We use AI for drafts and checks, with human review.
We have internal tools for drafting text, testing logistical replies and checking the site. Here we describe the available functions, human controls and dependencies on setup and suppliers.
Page updated on 28 September 2026.

- Someone asks for information or help with an administrative task.
- We configure the assistant with approved information and rules for handing requests to a person.
- Staff review the response or take over the request when needed.
Why this page exists
We test drafts and replies before proposing a service.
Our internal consoles prepare drafts and test the assistant with a fictional practice. A check looks for replies outside its scope, and a person reviews the result. These tests do not prove that a practice’s phone, WhatsApp or calendar is connected.
The systems
Six internal systems, each with its limit.
For each tool, we describe its function, human control and limit. Functions that depend on external services need setup and verification before use.
A model writes the draft, a person publishes it
- Problem
- Replying to every review and preparing recall messages takes time, and time runs short exactly when the workload grows.
- How it works
- From our internal console a language model prepares the draft: a reply to a review, a recall message, a front-desk text. It stays a proposal until a person approves it.
- Human control
- No draft goes out on its own. Reviewing and publishing always belong to whoever follows the practice.
- Stated limit
- If the model service is unavailable the console says so and the draft is written by hand: work does not stop, but it does not continue automatically either.
Healthcare rules live in the code, not only in the prompt
- Problem
- A text can look harmless and still contain a discount, a promise of results, or an implicit confirmation that the person who wrote is a patient of the practice.
- How it works
- Every generated text goes through a separate check that looks for discounts and promotions, promises of results and absolute superlatives, clinical terms and phrases confirming a doctor-patient relationship. The flags appear next to the text, before publication.
- Human control
- The check warns, it does not decide: the final call stays with a person, who reads the flags alongside the draft.
- Stated limit
- It is a check on words and phrases, not a legal opinion. It does not replace reading by someone who knows the case.
We run the front-desk assistant on ourselves first
- Problem
- A virtual assistant that steps outside its scope is not a cosmetic defect: in healthcare it is a risk.
- How it works
- Before we propose it, the assistant runs in our console with the configuration of a test practice. Its instructions are built server side only. After each reply a second check re-reads the text: if it drifts into clinical advice, the reply is replaced with a message inviting the visitor to contact the practice.
- Human control
- Only outcomes are logged, never the text of the messages. The practice configuration is not decided by the model.
- Stated limit
- The scope does not change with the place the assistant runs in: logistics and organisation, never clinical guidance. A public chat, when it is active, follows the same rules plus the ban on discussing prices, which are defined after the assessment.
The site provides alerts for new enquiries
- Problem
- Someone who writes to a supplier and gets no quick answer usually writes to another one. That is true for medical practices and it is true for us.
- How it works
- The site provides for storing enquiries and notifying the owner. It can send a confirmation email when the service is configured. Email and alert delivery depends on connections and suppliers.
- Human control
- A person calls back. No system answers in their place.
- Stated limit
- The code handles notification errors separately. If saving fails, the form reports the problem; it does not show a confirmation without a successful outcome.
We check the site before publishing
- Problem
- A bad change on a page is noticed late, and in the meantime that page stays online as it is.
- How it works
- Every change goes through a full build and three layers of automated tests (unit, API, browser), then a check that calls every page listed in the sitemap and a visual audit at phone, tablet and desktop widths. A dedicated test verifies that each Italian page with an English twin declares the other one.
- Human control
- No line goes live before those checks have been run and read.
- Stated limit
- Tests tell you whether a page works, not whether it is well written. Judging the text stays human.
We describe services and pages for search engines too
- Problem
- Search engines and AI assistants read pages without seeing them: if the information is not declared in a readable way, it gets reconstructed by guesswork.
- How it works
- The site publishes a summary file for AI assistants, structured data on commercial pages and articles, using the relevant types (organisation, services, frequently asked questions, navigation path), Italian-English correspondences generated from a single list shared by pages and sitemap, and update dates taken from the real files instead of from today’s date.
- Human control
- Notifications to search engines run in simulated mode: without an explicit command the system prints what it would send and sends nothing.
- Stated limit
- The instant-notification protocol is used by some engines, not by Google: for Google it is still sitemap and crawling.
The boundaries
What we do not let the AI do.
These lines are not good intentions: they are constraints written into the code or working rules already in force. They apply to us and to the systems we deliver.
- Drafts for reviews, campaigns and pages need human review. Automated chat uses instructions and checks, with referral to a person when needed.
- Recall drafts use generic campaign information. Chat or review text may contain personal data; health information and identifying details must be kept out of text sent to the model.
- It does not answer health questions. If a reply tries to, it is replaced with an invitation to contact the practice. The chat does not automatically forward the request.
- It does not set prices or terms: those are defined after the assessment, talking to the practice.
- It does not invent numbers. If a figure is not verified, it leaves the page instead of being rounded.
- It never writes reviews and never confirms in public that a person is a patient of the practice.
Where we work
Born in Italy, built to work remotely.
Studio Bonanno Consulting is an Italian business based in Mazara del Vallo (TP). Almost all the work happens remotely: analysis, design and configuration take place inside the systems of the practice, not in its waiting room. That makes distance an organisational detail rather than an obstacle.
The site lives in two languages, Italian and English, with the versions declaring each other and Italian as the reference. The English perimeter is a selection of pages, not the full translation: we prefer saying so to making the site look bigger than it is.
The constraints of Italian healthcare communication remain the yardstick in English too: no promise of results, no public discounts, no selectively solicited reviews.
Where it continues
The same principles, applied to your practice.
Frequently asked questions
The questions we get about this page.
Do you use AI to write the copy on this site?
Yes, as a drafting tool. Then every statement is checked against its source and whatever does not hold up is removed rather than softened: that is why you will not find percentages, client counts or success stories on these pages. Andrea Bonanno is responsible for what you read.
Does the AI answer patients instead of the practice?
No. The assistant handles logistics and organisation only: hours, location, how to book, reminders. Diagnosis, triage and clinical advice stay with the people of the practice. The same applies to us: an assistant on the site answers about services, method and how to get in touch, and points to the contact form when a person is needed. It does not automatically forward the conversation.
What happens when the model gets it wrong?
Checks outside the model flag words or replies beyond its scope and may replace a reply with a referral to the practice. Drafts need human review; automated chat replies are not all reviewed by a person before sending. Checks can miss problems and are not legal advice.
Does patient data pass through the AI?
Recall drafts use generic campaign data without the recipient’s name. Chat and review text, however, may pass to the model supplier and contain personal data. Before use, assess the data, supplier and applicable conditions; do not put health information or identifying details in free-text fields.
Do you work outside Italy?
We work remotely, in Italian and in English: the site exists in both languages, with the pages declaring each other, and Italian remains the reference version. The business is registered in Italy, in Mazara del Vallo (TP).
Why publish the limits as well?
Because a system whose boundaries are unknown cannot be assessed. In healthcare that matters more than a list of features: knowing where a system stops is what tells you whether it is worth activating.
Want to see how this would apply to your practice?
It starts with the analysis: what already works, what gets lost along the way, what is worth automating and what is not.
