Where should we start?
With one process, measurable, that today takes somebody hours of repetitive work. A first contained case costs little, is judged quickly, and tells you whether to go further.
AI that works inside your processes,
not alongside them
Assistants, quote builders and automations inside the systems you already use. With a human checkpoint wherever a mistake needs to stay visible.
measure --repetitive-time
hours per week spent copying, pasting and retyping
check --is-the-process-clear
if the rules are not written down, write them first
./pilot --one-case-only
[ok] one measurable process, not ten at once
[ok] no training on your data — written into the contract
waiting
$ ./assistant --with-sources
Every sentence says
where it came from
An assistant that just answers gets used once: with no source you cannot check it, and nobody signs a quote they cannot check. Here every part of the answer lights up the document it came from.
Does the via Marconi job need scaffolding? And what does that add?
On the north side the measured height is 7.4 metres,1 so certified scaffolding with an assembly plan is required,2 and the 18% surcharge for work above 6 metres applies.3
- specification-2026.pdf p. 14 Work at height requires certified scaffolding and an assembly plan.
- internal-price-list.xlsx row 88 An 18% surcharge applies to work above 6 metres.
- site-survey-report.docx p. 2 Height measured on site: 7.4 metres on the north face.
An example built on sample documents. The numbers are not the point: the point is that every statement carries the document and the exact passage it came from, and that when there is no source the assistant says so instead of inventing one.
// where it pays
AI pays where the work
is repetitive and checkable
It is not needed everywhere, and saying so is part of the job. These are the cases where it gives back real hours from the first month.
- 01 Answering the same questions by citing documents you already have
- 02 Drafting quotes and paperwork that today get retyped by hand
- 03 Pulling data out of invoices, contracts and PDFs instead of re-keying it
- 04 Routing incoming requests to the right person
- 05 Summarising long threads while keeping a link back to the original
// where it does not
And where it
wastes your time
Four cases where we say no. Automating a muddled process produces the same muddle, faster.
-
01
When the process is not clear
If the rules are not written down anywhere, they need writing first. AI amplifies whatever it finds, disorder included.
-
02
When mistakes cannot be caught
If nobody is able to notice that an answer is wrong, automation moves the risk without reducing it.
-
03
When a rule would do
Plenty of automations need no model at all: one condition and a webhook cost less, do not make things up, and need no supervision.
-
04
When it is only there to be seen
An assistant put in the window with no defined job gets tried once and then ignored.
// how we get there
One process at a time,
with the human checkpoint inside
Every automation has a point where a person sees it and can correct it. That is what makes a mistake recoverable rather than invisible.
Process discovery
01We look at how the work happens today, with the people who actually do it. Out of that comes the pilot case — and sometimes a "you do not need this".
A contained pilot
02One process, measurable. It costs little, it is judged quickly, and it tells you whether to carry on.
Integration into your systems
03The automation goes where the work already happens: your business system, your mail, your portal. Not another browser tab to remember to open.
Measure and correct
04Readable logs and a comparison against the time it used to take. If it does not give hours back, we change it or switch it off.
// from 2 August 2026
What your assistant says about itself
is now a legal obligation
The European AI Act has reached the phase that applies to organisations that USE artificial intelligence, not only those building it. If you run an assistant that writes to customers, or you publish generated content, these four points apply to you from now.
-
01
Whatever writes must say so
A system that converses with a person must declare that it is automated, clearly, at the first exchange at the latest — not at the bottom of the terms page. The only exception is when it is already obvious from context, and "obvious" is not decided by whoever installed it.
-
02
Generated content must be marked
Text, images, audio and video produced by a generative system must carry machine-readable marking. Not a caption under the photo: data inside the file, which has to survive the file being re-uploaded somewhere else.
-
03
Deepfakes must be declared
Anyone publishing content that depicts people, places or events in a realistic but manipulated way must say so perceptibly. For artistic or satirical work it is enough to flag that the manipulation exists, without ruining the piece.
-
04
In Italy there is also Law 132/2025
The Italian AI law requires businesses and professionals to declare their use of AI tools in the services they provide, and employers to inform staff when AI enters personnel management. AgID and ACN supervise.
// how you get compliant
Four steps,
and none of the first ones needs a lawyer
This is not a thousand-page exercise: it is knowing which systems you touch, saying what they are, marking what you generate, and writing it down somewhere. We do it together, and what comes out is a document you can show.
An inventory of what is switched on
01List every system that produces text or talks to someone: assistants on the site, automatic replies over email or WhatsApp, description generators, filters on job applications. The usual surprise is not what you bought: it is what somebody enabled inside a tool you already had.
The notice, at first contact
02We write the sentence the assistant says about itself, and decide where it appears. At the first message, not the third. Phrased well it costs a line and puts nobody off — what puts people off far more is finding out later.
Marking what you generate
03Content produced by a generative system gets the required metadata, and the content management system keeps a record of what was generated and from what. It is the most technical part, and the one that never happens unless someone sets it up once.
The register and the two dates
04A document saying who is answerable for what, which systems exist and what was decided. With the deadlines in it: the obligations apply from 2 August 2026, and systems already running before that date have until 2 December 2026 to comply.
$ ./automations --where
Not one more program.
Inside the ones you already run
The automation that works is the one nobody has to remember to open. It lives inside the email, the chat and the system where the work happens anyway — and you only notice it because something you used to do by hand has stopped asking you.
Gmail Reads the enquiry as it arrives, works out which job it is about, and drafts the reply with the right quote already attached.
$ cat stack.json
What we build it on
- OpenAI and Anthropic Models chosen per task, not per trend: swappable without rewriting the system
- RAG Answers cite the document they came from, so they can be checked
- Embeddings and search Finds by meaning, not exact wording: it copes with your in-house jargon
- Queues and webhooks Long operations do not leave anyone waiting in front of a screen
- Human checkpoint A person sees the draft and confirms: mistakes stay visible and fixable
- Logging and tracing What was asked, what was answered, and from which source
$ ./domande --frequenti
The questions that
always come up
Does AI replace people?
In the projects we do, no: it removes the mechanical part of the work and leaves people the decisions and the oversight. If the stated goal is to reduce headcount, it is fair to say so up front, because it changes how the tool should be designed.
What happens when the model gets it wrong?
We plan for it per project: every automation has a human verification point where the mistake is visible and correctable, and the logs let you work out what happened.
Is our data used to train models?
We choose configurations and providers that exclude it, and we put that in the contract. It is one of the first things to settle, not a technical detail.
Do we get locked into one model provider?
No, and that is an architectural choice: the model sits behind an interface of ours, so it can be swapped without rewriting the system. That also helps when prices change.
What does an AI project cost?
A contained pilot is the cheapest part and exists precisely to help you decide. The number for the rest comes afterwards, once the real process is visible and we know how many integrations are involved.
Does our assistant have to declare that it is an AI?
Yes, if it converses with a person: from 2 August 2026 Article 50 of the European regulation requires it, at the first exchange at the latest and in clear terms. The exception only applies when it is already obvious from context — and it is not enough that it is obvious to whoever installed it. If yours has been running since before that date, the deadline to bring it into line is 2 December 2026.
What is the risk if we do not?
For breaches of the transparency obligations the regulation goes up to 15 million euro or 3% of worldwide annual turnover, whichever is higher, with reduced caps for small and medium enterprises. In Italy, AgID and ACN supervise. That said: the fine is the small risk. The big one is a customer working out on their own that they were talking to a machine.
Does it apply to generated photos and text too?
Yes. Content produced by a generative system has to be marked in a machine-readable format, and content depicting people or events in a realistic but manipulated way has to be declared perceptibly. It does not cover edits that do not substantially alter the source material.
fymera init --project "yours"
Bring us a process
that wastes your time
Thirty minutes on video about one concrete case. If a rule would do instead of a model, we will tell you.
