04 / Services
Models inside working processes
Sorting letters, draft replies, search across your own documents — where people keep doing the same thing.
For departments drowning in similar inquiries, protocols, and internal instructions.
Model
Documents
Document
Draft
A person checks
Source
Own
Reply
Draft
Who
Staff member
What it consists of
What goes into the system
- 01
Where a model fits
First we look at a repeated step: sort a letter, draft a reply, find a clause in your own documents. We do not put a model “on the whole department.”
- Which inquiries are of one kind
- Where a mistake is costly
- What a person checks anyway
- 02
Your documents
The answer rests on your instructions and correspondence, not on the model’s general knowledge.
- Which folders and databases may be read
- What must not go outside
- How to update the documents when a rule has changed
- 03
A draft, not an auto-reply
The model prepares the text or the selection. A person sends it when the case is not a calm one.
- A draft in the same window where the staff member works
- A mark of where the fragment came from
- A button for “not this”
- 04
Where we fit it in
Not a separate chat off to the side. The step shows up in email, the workspace, or the queue, where the work already happens.
- The current system stays the main one
- A queue, if there are many requests
- A log: what was asked, and what was answered
- 05
Quality after the start
For the first weeks we watch where the model is wrong or stays silent. Then we extend only the calm branches.
- A sample of answers, read by eye
- Which topics we switch off
- Who answers if the model is unsure
What we do
- 01Which steps are worth handing to the model at all
- 02Choosing a method, without “AI for the sake of AI”
- 03Fitting into systems that are already standing
- 04A short training for the people who will check the answers
- 05Watching quality after the start
How we proceed
- 01Where time is lost
- 02A trial on your data
- 03Embedding
- 04Checked by people
- 05An extension
Stack
- Python
- Model APIs
- Vector search
- Queues
A guide for the timing
A trial takes two to three weeks. A working system takes from a month, if the data is already gathered.
Outcome
A repeated piece of the process closes faster, and disputed cases stay with a person.
What usually comes later
Voice, sending to a client automatically with no person involved, and “your own model on your own server” — only if a trial on your data already holds.