Staff have already decided that large language models are useful. The question for the organisation is whether they get to use them in a way the security team, the legal team and the board can live with. Left unmanaged, the default outcome is that confidential documents get pasted into consumer chat tools on personal accounts. Self-hosting an enterprise AI platform is the practical alternative: the same class of capability, inside the organisation’s boundary, with the controls that any business system is expected to have.
Why the public chatbot is a shadow IT problem
The trouble with consumer AI tools is not that they are bad. It is that they sit outside every control the organisation has built. There is no single sign-on, so departed staff keep their accounts. There is no record of what was asked or what came back. Retention and training policies are set by the vendor and can change. And the document pasted in to be summarised has now left the country, whether or not a policy said it could.
For a small firm this may be an acceptable risk. For anyone handling client data, health information, government work or anything under a contractual confidentiality obligation, it typically is not. The fix is not a ban, which staff route around, but a sanctioned tool that is at least as convenient as the one they are already using.
Multi-model routing behind one interface
“Self-hosted AI” is often heard as “run a small open-weight model on a server and hope it is good enough”. That undersells what is possible. A well-designed platform presents one interface to staff and routes each request to the most appropriate model behind it.
Open-weight models running on the organisation’s own hardware handle the sensitive work: anything involving client files, internal documents or personal information stays on the organisation’s infrastructure end to end. Frontier models can be made available alongside them, for tasks where the data classification allows and quality matters most. Which model answers which request is a policy decision, made centrally and enforced by the platform, rather than something users are asked to remember.
The user sees one chat window, one login and one history. The organisation gets to choose, per team or per data classification, where inference happens.
Document-native workflows
The productivity gain from these tools comes mostly from working with documents rather than typing questions. A useful platform reads Word, Excel, PowerPoint and PDF files directly in the conversation, extracts what is in them, and can write changes back into the same formats.
In practice that means a policy document can be dropped into a chat and rewritten in plain English, a spreadsheet can be interrogated about the numbers it contains, and a slide deck can be drafted from a briefing note, all without the file leaving the platform or being converted to something else on the way. The difference between a tool that does this and one that only handles pasted text is, in our experience, the difference between staff adopting it and staff ignoring it.
Governance and audit per user
An enterprise AI system needs the same controls as any other system holding business data. At minimum that means:
- authentication through the existing identity provider, with access removed when the person leaves
- a per-user audit log of prompts, files supplied and responses returned
- retention rules that the organisation sets, not the vendor
- the ability to restrict which models, features and data sources each group can reach
- a documented answer to “where did this data go” for every request
This is also where governance becomes a feature rather than a burden. When a manager can see how their team actually uses the tool, they can spot where it is saving time and where it is producing nonsense, and adjust policy accordingly. Auditability also answers the question that arrives eventually from a client, an auditor or a regulator: show me what your AI did with our information.
Cost visibility
Token-metered commercial APIs make spend hard to predict and easy to lose track of, particularly once several teams are experimenting at once. A platform that sits in front of every model can attribute usage to people, teams and projects, and report it in terms a finance team understands. Open-weight models on owned hardware have a flat cost profile, which makes the split between “run it here” and “send it to a frontier model” a decision that can be made on numbers rather than guesses.
The cost conversation is usually the second one an organisation has, after the security one, and having the usage data ready shortens both.
Where Bizix fits
Bizix builds and operates self-hosted enterprise AI platforms that put frontier and open-weight models behind one governed interface, with document-native handling of Office and PDF files, per-user audit, and usage reporting. Customer data stays inside the customer’s boundary, and the platform runs on the same sovereign private cloud we build for everything else.