What is your company’s own AI?
It is an AI assistant that works across your company’s systems: email, ERP, documents, spreadsheets, CRM and calendar. It has the same permissions as your people, answers questions with the source of every fact, and prepares the work where it is needed: a draft reply, an order to confirm, a follow-up to approve. It asks before it acts. Stolen Orbit sets it up, connects it to your tools and looks after it over time.
Read the details
It is not software to download or a chatbot for website visitors. It is a custom service built on the tools you already use: it starts with one process, and every automation you switch on connects one more system.
Scroll the table to compare all columns.
| Request | What the assistant does | What stays with you |
|---|---|---|
| “What did we promise this customer?” | Searches the emails, the order and the sales rep’s notes; answers with dates and sources. | Deciding whether to update the customer. The draft is ready; sending it is up to you. |
| “Which quotes have waited over a week for a reply?” | Cross-checks open quotes, recent contacts and the calendar; drafts the follow-ups. | Reviewing the follow-ups and choosing which to send. |
| “How do we handle an out-of-warranty return?” | Finds the current procedure in the service manual and terms of sale, with the page. | Confirming the reply, or correcting it if the procedure has changed. |
Illustrative examples with made-up data. They do not describe a client’s work or results.
How is it different from ChatGPT or Copilot on their own?
The AI model may be similar; what surrounds it is not. On their own, ChatGPT or Copilot work with what you give them in the moment, or with the sources their vendor can connect. Your company’s own AI is connected to your ERP and your rules, prepares the result inside your tools, is tested on your cases, and someone looks after it when things change.
Read the details
Business versions of these tools can already connect to some sources. Microsoft, for example, documents that Copilot uses Microsoft 365 email, documents and calendar while showing each person only the data they can see. If you use them, we start from what they cover. What usually stays out is the ERP, your data rules (which price list applies, who approves what) and writing the result where the work actually ends.
Scroll the table to compare all columns.
| Aspect | On its own | Your company’s own AI |
|---|---|---|
| Data | Pasted text, uploaded files or sources the vendor connects. | The systems you agree on: email, ERP, documents, spreadsheets, CRM, calendar. |
| Permissions | Whoever pastes or uploads decides; connected sources follow the vendor’s rules. | Those of the person asking, checked on every search. |
| Result | Text to copy wherever it is needed. | A draft in your own tool: a reply, an order or a follow-up. |
| Checking | Up to whoever reads each answer. | Tests on your cases before launch, measurements after. |
| Over time | Nobody looks after your rules and connections. | Sources, permissions and connections looked after by Stolen Orbit. |
Which systems does it connect to?
The systems where the work happens: email, ERP, documents, spreadsheets, CRM and calendar. Not all at once: we start with the ones the first process needs and add one with each new automation. For every system we first check how the data can be reached, for example through an API, exports, shared folders or a dedicated mailbox.
Read the details
If a system offers no suitable access, we say so up front: you can start read-only, from a regular export, or with a different scope. An available connector does not prove that permissions, edits and deletions are handled properly; it has to be tested on your system.
Scroll the table to compare all columns.
| System | What it reads | What it can prepare |
|---|---|---|
| Threads with customers and suppliers, attachments. | Draft replies and follow-ups. | |
| ERP | Orders, stock, customer records and documents. | Draft orders and documents to confirm. |
| Documents | Approved manuals, price lists, contracts and PDFs. | Answers that point to the page. |
| Spreadsheets | The sheets your team uses every day. | Summaries and lists to check. |
| CRM and customers | History, quotes and commitments made. | Updates and tasks to approve. |
| Calendar | Deadlines, deliveries and callbacks. | Reminders for the right person. |
Permissions and data: what can it see?
Only what the person using it can see. If a colleague has no access to a document, the assistant does not show it or use it to answer them. Access is checked when it searches, not by asking the model to keep quiet about something it has already read.
Read the details
Before connecting a system, we put in writing which data is needed, where it is processed, which providers receive it and how long it is kept. We also check the AI model provider’s terms, such as whether data is used for training. Your designated contact authorises each connection, and access can be revoked.
- Access under the identity of the person asking, not an account that sees everything.
- Read-only connections until a write has been tested and approved.
- Emails and documents treated as data, never as instructions to follow.
- A record of the sources consulted and the work prepared, kept as agreed.
It asks before it acts.
The assistant prepares; a person decides. A reply waits as a draft in your inbox, an order waits for confirmation in the ERP, a follow-up waits for approval. Nothing is sent or recorded without the agreed check.
Read the details
Over time you can decide that some low-risk actions go ahead on their own, one at a time and only after measuring them. Prices, discounts, refunds and commitments to customers stay with people, unless a specific project puts proper controls in place.
It receives a request
From someone on your team, or from an email within the agreed scope.
It searches the permitted sources
Only the connected systems, with the permissions of the person asking.
It prepares the work
An answer with its sources, or a draft in the right tool.
It waits for confirmation
A person approves, corrects or discards it. Corrections become test cases.
How does it start?
With one process that weighs on you today: orders arriving by email, quotes waiting for a reply, or customers asking the same questions. We trial it on real cases for an agreed period, with a result measured before and after. Then you decide whether to extend it, adjust it or stop.
Read the details
Every automation you switch on connects one more system: the assistant that started with orders can later read the CRM or the calendar too. That is how, one piece at a time, it becomes your company’s own AI.
A map of how you work
Processes, documents and systems: where information lives, who uses it and where time goes.
One process and its starting point
Volumes, time per case, errors and waiting times, measured before anything changes.
A trial on real work
For an agreed period the assistant prepares and people confirm. Exceptions stay visible.
Compare and decide
The same measurements before and after. If the result does not justify the cost, it stops there.
What gets measured?
The same things before and after the trial, on the same kind of work. Released time is available capacity: it only becomes a saving when a cost actually changes.
Read the details
- Time per case, from request to finished work.
- Drafts accepted as they are, corrected or discarded.
- Errors, exceptions and cases handed to people.
- Response times to customers and colleagues.
- Running costs: software, model usage and ongoing care.
Who looks after it over time?
We do, through an ongoing service with a written scope. An assistant connected to a company’s systems needs care: documents change, people change roles, vendors update their APIs and AI models improve.
Read the details
The scope sets out what is included, how to ask for support and what counts as a change. We do not promise unlimited work or response times that have not been agreed.
- We check answers and drafts on a sample, against their sources.
- We update sources and permissions when documents, roles or procedures change.
- We notice when a connection breaks and recover the work left behind, without duplicating it.
- Every approved or corrected answer becomes a test case: before changing a model or a rule, we rerun the tests.
- We add connections and automations when the work calls for them, using the same method.
When is it not the right choice?
When something simpler will do. If a rule in your ERP or a required field in a form solves the problem, you do not need an AI assistant. If volumes are low, the time released may not cover setup and ongoing care.
Read the details
In those cases we tell you, and point to the simplest alternative: a rule, a connection between two systems, or organisational work before any technology.
- Nobody can review drafts or answer for the decisions.
- Procedures are not written down, or nobody keeps documents up to date.
- Your systems offer no usable access, not even read-only or through exports.
- The process changes every week and there is no stable version to test.
- You want AI to decide on prices, payments or customer commitments by itself.
- You only need help writing text: a general assistant may be enough.
A template to work from.
First AI pilot brief
An editable Markdown template for your process, data, owners and go/no-go criteria.
Download the Markdown templateAI permission matrix
A permission matrix with approver, tests and a revocation procedure.
Download the Markdown templateKnowledge source register
Sources with valid versions, access rules, updates and retrieval tests.
Download the Markdown templatePractical questions
Is it a website chatbot?
No. It works for your people, on your systems, and prepares work for them to confirm. Answering customers is one possible starting point, with its own rules and checks.
Do we need to replace our ERP or CRM?
No. We connect to the tools you use. If a system does not allow suitable access, we flag it up front and agree a different scope.
Can it send emails or confirm orders by itself?
Not at first: it drafts and asks for confirmation. Later you can decide that some actions go ahead on their own, one at a time and after measuring them.
Which AI model does it use?
We choose it for each project on quality, cost and data terms. Approved and corrected answers become test cases, so before switching models we check that the results stay correct.
What does it cost?
There is no standard price: it depends on the process, the systems and the volumes. The proposal separates setup, ongoing care, and software and model usage. The first process is trialled with a result measured before and after.
What do we need to get started?
A description of the process that costs you time and the tools you use. Anonymised or made-up examples are enough: please do not send credentials or customer files through the contact form.
References and method
Microsoft documents that Copilot only shows the Microsoft 365 data each user can access. We cite it to describe fairly what a connected general assistant already does.
Recommends minimal permissions and autonomy, actions executed in the user’s own context, and human approval for high-impact actions.
Describes instructions hidden in emails and documents, which is why consulted content is treated as data, not commands.