General-purpose AI tools know a little about everyone's business and nothing about yours. A private model knows your products, your specifications and your past projects — and your data never leaves your organisation.
Where the problem starts
Ask a general-purpose tool to "prepare a quotation for this specification" and the answer looks plausible but knows nothing about you. Not your product codes, not your standard unit prices, not which team does which work.
So you explain the same background every time. You were going to save time; you spend it explaining.
The second problem is heavier: explaining means exporting. Your specification, your cost table, your customer list — all of it goes to somebody else's server.
What a private model actually is
A private company AI model is a system trained on your own data and running on your own hardware, or on a server under your control.
It knows your catalogue, your technical specifications, your past quotations, your work orders and your reports. Give it a new specification and it answers by comparing it against your own history.
And crucially: neither the question nor the answer goes to the internet.
Two deployment routes
The first is document-grounded deployment. The model reads your documents like an archive; when a question arrives it finds the relevant document and answers from it. Faster to deploy, updates instantly as documents are added, and it can show which document the answer came from.
The second is training the model in your organisation's own language — your writing style, your terminology, your decision patterns. It takes longer and needs more data, but the output reads as if your own team wrote it.
In most organisations the right answer is a combination: knowledge from documents, tone and format from the trained model. Which work is solved by which route is decided during discovery.
If a vendor insists on a single route, be careful — they may simply be selling the only route they have.
What drives the cost
Not data volume, but data disorder. Ten thousand well-organised documents deploy more cheaply than one thousand scattered ones. Most of the cost sits in cleaning and structuring.
Number of integrations. A standalone system is quick; connecting to ERP, CRM and accounting means separate work for each link.
Hosting choice. Your own hardware needs capital; a cloud server under your control becomes an operating cost.
User count usually does not change deployment cost — which is why cost per user falls sharply as adoption grows. The opposite of a subscription model.
Where these projects fail
When data cleaning is skipped. A catalogue listing the same product under three names teaches the system there are three products. Once a team gets one wrong answer, it stops trusting the system.
When training is skipped. Deployment finishes, nobody uses it. This is the most common failure and it is not a technical one.
When no owner is assigned. Someone must keep the knowledge current. An unowned system is stale within six months.
When expectations are set wrong. AI does not make the decision; it gives time back to the person who does. Projects launched on "let the system decide" end in disappointment.