PRIVATE AI, CONNECTED TO YOUR SYSTEMS
Your company’s own AI
One assistant connected to your email, ERP, documents, spreadsheets, CRM and calendar, with the same permissions as your people. It answers with sources and drafts replies, orders and follow-ups in your own tools, and it asks before it acts. It can run in your cloud or on your servers, with no data sent to external AI services; your team’s corrections become examples and tests you keep. We set it up and look after it, one process at a time.
Our services, in detail
What each service covers, how it is tested and when it is not worth it. Services at a glance
Where projects start
Connect data and systems
Practical guides
One question per page, with examples, criteria and sources.
Decide and budget
Design
Test and launch
Tools, templates and reference
Calculators, guided decisions and worksheets. Free, with no sign-up. All tools and templates
Interactive tools
Templates to choose and design
Templates to check and operate
Industry examples
Typical processes, told by industry. They are examples: we work with businesses in any industry. All industries
Production & operations
Services & relationships
Commerce & hospitality
About us and how we work
Stolen Orbit builds and runs custom AI and automation for businesses in Italy and Europe. We work in Italian and English.
The company
Working together
Work with us
Start with the decision you need to make.
Short paths: an answer, a test or worksheet, then the relevant service. You do not need to read the whole library.
We want to start using AI
Define a process, compare alternatives and prepare a brief.
We use ChatGPT, but manual work has not decreased. Why?
A standalone assistant can draft text while leaving every later handover intact. If somebody still copies, checks and assigns each request, examine the whole journey. A required form field or an existing business-software rule may solve the problem; consider AI where interpreting variable content is actually necessary.
First check Follow one request from arrival to completion. Record each manual step, who performs it and which missing information prevents progress.
We copy data between systems
Map data and access, then test failures, duplicates and reconciliation.
Why do order totals differ between our CRM and ERP?
Connecting two systems does not guarantee that they represent the same order. A later amendment, different customer identifier or repeated attempt can create differences. Before adding AI, define a shared identifier, the authoritative system for each field and a check for missing, duplicated or changed records.
First check Compare one order across both systems using its shared identifier. Check lines, status and last update; assign responsibility for resolving differences.
Answers are scattered
Understand how to retrieve sources and which actions people should retain.
How can we answer customers without always asking a colleague?
A document assistant is useful only when it retrieves an authorised, current source. An old price list in a shared folder should not override the approved version. Establish who can see each document, who updates it and how to flag questions for which no adequate source exists.
First check Choose one recurring question. Identify the valid document, its owner and permitted roles; include an outdated source when checking the answer.
We need to know whether it is reliable
Build a sample, inspect failures and decide when the workflow should stop.
The demo reads PDFs well. Can we trust it in production?
A successful demonstration does not measure the errors your team will encounter. A wrong quantity may matter more than an imperfect description. Prepare expected answers for ordinary, incomplete and ambiguous documents; check critical fields separately. Decide beforehand which errors stop the workflow and which need human review.
First check Choose one critical field and write its acceptance rule. Try a document with that field missing: the system should flag the absence.
We need to justify the investment
Measure remaining work, running costs and benefits actually realised.
If we save working hours, does the project automatically pay for itself?
Recovered hours create available capacity; they become cash savings only when an expense actually changes. A team may handle more requests with that time, but measure the result. Include implementation, software, ongoing care, human checks and corrections in the comparison. Consider months with lower volumes too.
First check Measure the complete handling time for one task, including corrections. Separate removable spending from additional work the team could take on.
We need to use it in daily work
Define rollout, training, ownership and operational responsibilities.
The automation stopped. Who should deal with it?
A connection may fail after a password, permission or format changes. The project needs an owner, an understandable alert and a procedure to recover unfinished work without duplicating it. Agree who maintains the workflow and which interventions belong to ongoing management, including what happens when that owner is unavailable.
First check Simulate expired access in a test environment. Check who receives the alert, how the workflow stops and how processing resumes safely.
We want an AI that knows our company
Connect systems one at a time, with permissions and approvals.
Can we have a ChatGPT that knows our customers, orders and procedures?
Yes, if it is connected to the right systems and respects permissions. A general assistant answers from what you give it; to know your company it has to read email, ERP and documents with the permissions of the person asking, cite its sources and prepare work without sending it on its own. Start with one process, measure it and add one system at a time.
First check Pick a question that today means opening three tools. Note which systems it needs, who may see them and who has to confirm the answer.