Using ChatGPT, Claude or Copilot without exposing your company’s data
How to use ChatGPT, Claude or Copilot at work without exposing sensitive data: business plans, settings, anonymisation, usage rules and the right tools.
5 min read
AI software · Published by Step By Step
AIKeep, the AI software for your whole business
AIKeep is the AI software for your whole business: installed on your Windows workstations, it becomes the single command centre of your company. Accounting, banking, payroll, documents, business applications and servers are run through plain-language instructions, without moving your files.
- Windows
- APIs and MCP
- AES-256-GCM
- Offline OCR
- Windows, no server to manage
- Installed on your PCs
- For all of your business admin
- A single command centre
- Key kept on the workstation
- AES-256-GCM encryption
- Sensitive data removed before sending
- Filtered by AIGuard
Domains
What does the AIKeep AI software run in your business?
AIKeep is not one more piece of software sitting next to the others: it sits in front of them and becomes the single command centre. Here is what it handles, and the list stays open: as soon as a tool exposes an API or an MCP server, it falls within its scope.
-
Bulk processing
Hundreds of documents handled in one go, exhaustively.
-
Accounting
Entries, checks and reports, from your documents and accounting software.
-
Banking
Bank statements and reconciliations.
-
Payroll and HR
Staff records, payroll variables and regulatory obligations.
-
Quotes and invoicing
Issuing and tracking quotes and invoices.
-
Sales management
Customers, ongoing deals and commitments.
-
Text recognition
Scans and identity documents read offline, on the workstation.
-
Document creation
Reports, spreadsheets and presentations produced on demand.
-
Email
Sorting, extracting information and drafting replies.
-
Calendar
Checking availability and booking appointments.
-
Managed IT
Servers, SSH connections, monitoring and fleet updates.
-
Cybersecurity
Configuration checks and hardening of workstations and servers.
Features
Four ways to reach your tools, no dead ends
Most assistants stop where their publisher planned an integration. AIKeep has four ways to reach a system, from the most direct to the most universal: if the first does not apply, the next one takes over.
-
1. APIs and MCP servers
Any REST API is connected through configuration (OAuth 2.0, API key, credentials or token), with pagination handled. Any MCP server plugs in the same way, and several connectors ship preconfigured.
-
2. Browser automation
When a service exposes no interface, AIKeep drives a real browser: signing in to the portal, two-factor authentication included, navigating and filling in forms. Execution stays confined to the domains you allow, and it is logged.
-
3. Code execution
When no interface fits, AIKeep writes and runs Python code in a sandbox, with no network access, on authorised files only. This is what lets it handle needs nobody had anticipated.
-
4. System administration
Commands run locally or over SSH on your servers: updates, diagnostics, inventory and remediation. In supervised mode, every command is shown to you for approval before it runs, and all are recorded in an encrypted log.
Reading and writing files
Common and legacy office formats, native PDFs and scanned documents: text recognition is built in and runs offline. AIKeep also writes into your import files without altering their structure, and produces spreadsheets, reports and presentations.
A company memory
What is learnt along the way (legal identifiers, internal rules, contractual specifics, the quirks of a server) is recorded at the right level and reused afterwards. That is what separates an assistant from a tool that starts from scratch every morning.
Principles
Three properties that stem from AIKeep’s architecture
These guarantees do not rest on a revocable commercial promise: they follow from the way the software is built.
-
Runs on your infrastructure
Competing platforms rely on an ingestion phase: your documents are uploaded and indexed on their side. AIKeep runs on your workstations: it opens files where they are, calls your applications from your network and administers your machines from the inside. No external copy is made.
-
Local data the publisher cannot read
History, knowledge records, API credentials and server access secrets are encrypted at rest on the workstation, with a key that exists only there. The publisher has no technical means to read them: it is a cryptographic property, not an internal policy.
-
What leaves is filtered before it leaves
Inference requires sending the submitted excerpt to the model: we say so plainly rather than leaving it out. That is the very purpose of AIGuard, placed at the software’s exit point: everything heading for the model goes through it, including content never typed into a conversation.
Getting started
How do you get started with AIKeep, without an IT project?
No migration, no history import, no change to your existing tools: AIKeep is up and running in four steps, and your software stays the software you already use.
-
Declare the scope
A network folder, a business application, a server, an authorised website: AIKeep acts strictly within what you declare, and the model cannot widen that scope on its own.
-
Give the instruction in plain language
You phrase the request as you would to a colleague, whether it is a bank reconciliation or a server update. No scenarios to configure or maintain.
-
Exhaustive execution
Items are processed by code, not by the model: the whole declared scope is covered, with no sampling and no silent truncation.
-
Documented results
The output is produced, and anything that could not be established is left blank and listed. Traceability comes before the appearance of completeness.
Security
AIKeep’s protection mechanisms
Processing routinely involves identity documents, bank details and personal data. AIKeep’s security requirements are set accordingly.
-
Encryption at rest
Content is encrypted with AES-256-GCM using a random data key, itself sealed by a key derived (Scrypt) from your password, your PIN or your recovery key.
-
Inactivity lock
Locking is decided by the local service. At that moment, the data key is wiped from memory and the content becomes encrypted again.
-
Multi-entity isolation
Each company managed on the workstation has its own workspaces, settings, connectors and memory. An identifier belonging to another entity is refused, not merely hidden.
-
Backup and device transfer
Export produces a consistent, fully encrypted snapshot that can be restored on another machine with your usual password, without the content being decrypted during the operation.
The exact scope of confidentiality
Saying that “no data leaves” would be inaccurate, and we do not say it. What does not leave is storage: files, history and knowledge records stay on the workstation, encrypted. What is transmitted is the excerpt submitted for inference, meaning the portion of content strictly needed for the task requested. Large volumes are processed by code running locally and are never submitted to the model.
Comparison
AIKeep compared with index-based AI platforms
The difference comes down to where your data lives and what is actually sent to the artificial intelligence model.
| Criterion | Index-based platform | AIKeep |
|---|---|---|
| Where is your data? | Uploaded and indexed by the platform provider | On your workstations, encrypted: no external copy is made |
| What is sent to the model | The indexed content, transferred once and for all | Only the excerpt needed for the current instruction, filtered by AIGuard |
| Tools without an interface | Out of reach until an integration is built | Driven through a real browser, or code written for the task |
| Large volumes | Submitted to the model, with a risk of sampling | Processed exhaustively by code running locally |
The model
Reasoning entrusted to Anthropic’s Claude models
AIKeep’s engine is not hidden behind an in-house brand: the inference provider is named, and its commitments are quoted as they stand. Anthropic’s commercial terms state that its customers’ content is not used to train its models, and its data processing agreement, with the standard contractual clauses governing transfers outside the European Union, is automatically incorporated into those terms.
What reaches the model is limited to the excerpt needed for the current instruction. When AIGuard is enabled, that excerpt is also stripped of its sensitive data before it leaves, and the real values are given back to you in the answer.
Claude and Anthropic are trademarks of Anthropic, PBC. AIKeep, published by Step By Step, is an independent solution: it is not affiliated with, sponsored by or endorsed by Anthropic.
Support
AIKeep Portable and integration services
AIKeep also comes in a portable edition, on a USB key, for technicians working in the field. And because Step By Step also builds custom software, the studio integrates AIKeep for its clients.
-
AI audit and needs analysis
We study your tasks, tools and data to decide what AIKeep should take on first.
-
AI integration
Connecting your software, data and servers, including through custom-built connectors.
-
AI governance
Defining who can do what, tracing actions and keeping control over what AI does in your business.
First call on us
See AIKeep at work on your own tasks
Name the task that costs you the most time: bulk processing, repetitive data entry, a tool with no API or an admin operation. We will show you how AIKeep handles it, with no obligation.
Going further
Our other software and artificial intelligence services
- AIGuard Anonymises your data before it reaches ChatGPT, Claude or Copilot.
- AI agents AI agents that carry out multi-step tasks in your tools, under human control.
- AI automation Emails, documents, data entry: repetitive tasks handed over to AI.
- AI integration Large language models connected to your software, CRM or ERP.
Blog
Further reading: our articles on the topic
-
-
The Model Context Protocol (MCP): connecting AI to your business software
The Model Context Protocol (MCP) standardises how AI models connect to your software and data. How it works, what it enables, security and how to adopt it.
5 min read
FAQ
Frequently asked questions about AIKeep
What is AIKeep?
AIKeep is artificial intelligence software published by Step By Step and installed on a company’s Windows workstations. It becomes the single command centre of the business: accounting, banking, payroll, documents, business applications, IT estate and security are all driven by plain-language instructions.
Do we need to install a server or move our files?
No. AIKeep installs on a Windows workstation, opens in your usual browser and works on your folders where they already are. No migration or history import is needed, and your current tools stay exactly as they are.
Is our data sent to the publisher or to the AI model?
Storage stays on the workstation: files, history and knowledge records are encrypted there, with a key the publisher does not hold. Only the excerpt needed for the current instruction is sent to the model for inference, and AIGuard removes sensitive data before it leaves, then restores it in the answer.
Which AI model does AIKeep use?
Reasoning is handled by Anthropic’s Claude models. Under Anthropic’s commercial terms, customer content from its services is not used to train its models, and its data processing agreement, including standard contractual clauses, is incorporated into those commercial terms.
Does AIKeep work with websites or portals that have no API?
Yes. When a service exposes no interface, AIKeep drives a real browser as a person would: it signs in, fills in forms and downloads documents. When a two-factor code is sent, it asks you for it and carries on.
Can several companies be managed from the same workstation?
Yes. Each company has its own folders, settings, connectors and memory, fully isolated. You switch from one to another in a click, and nothing from one is accessible from the other.
How can we see AIKeep in action before deciding?
A guided thirty-minute demonstration is available, by video call or at your premises, with no commitment and nothing to install. You name the task that costs you the most time, and it is replayed in front of you on a demonstration dataset of equivalent complexity.