Private AI for business
Give your people a place to use AI that the organisation controls, rather than banning the tools they already reach for. In that place, access is granted project by project, personal details are masked before anything reaches a model, the organisation chooses the model provider and can use its own key, and every AI call is recorded.
Forge, by GradTensor, is that place. Your people ask questions of their documents, draft from them and run repeated document work as workflows, under rules an administrator sets and members cannot weaken.
When AI is useful and the approved route is slow or missing, people paste documents into whatever chatbot they have open. A ban moves that out of sight rather than stopping it.
The answer is a better place to do the work: one that is as quick to use, and where the organisation, not each individual, decides what is allowed.
In Forge, being a member of the organisation opens nothing on its own. Access is granted per project, so each client, matter or team can have its own project, and a person sees only the projects they have been invited to.
An administrator can see that a project exists, who is in it and what it costs, without being able to read its documents or conversations.
Masking is on by default in projects. Before a prompt leaves, recognised identifiers are replaced with placeholders, and the real values are put back into the answer your person reads. The model never sees them.
An administrator can make masking a rule for the whole organisation. A project can then be stricter but never looser, and the refusal happens on the server, not behind a hidden button.
The organisation chooses whether its calls go to Anthropic or to OpenAI. It can add its own API key, so model spend sits on its own account under its own commercial terms, and the administrator sees usage per person.
Every access and every AI call goes on an append-only record: which model, which surface, whether masking applied, whether the answer declined and whether its citations checked out. The record never holds content, which is why an administrator can be allowed to read it.
Answers drawn from documents cite their source, and each cited file is checked against what was actually retrieved, so an invented citation is flagged.
The business tiers of ChatGPT, Claude and Microsoft Copilot all offer administrator controls and terms that keep your data out of training. If what you need is a good assistant for each person, they are a reasonable choice.
Forge is for organisations that need the rules to sit with the organisation: access granted per project, masking before any model is called, a choice of model provider, and a record of every call. It uses Anthropic's and OpenAI's models underneath.
Is it safe to upload confidential documents to ChatGPT? →Give them an approved place that is as easy to use, and make the protections rules rather than habits. In Forge, masking can be an organisation rule that no member can switch off, and a person reaches only the projects they are invited to.
Yes. An organisation chooses Anthropic or OpenAI for its calls, and can supply its own API key for the provider it chooses.
No. It catches email addresses, phone numbers, card numbers, Aadhaar and PAN numbers, and written US social security numbers. Names and addresses are not detected. That is why access per project matters as much as masking.
Yes. A project holds a set of documents and the conversations about them, and access is granted per project. One project per client keeps each client's material apart.