Skip to content
Hendez
← All services

Services

Private company AI model

Last updated

Short answer

A private company AI model works from the organisation's own documents, own history and own terminology, and its data never leaves the organisation's server or private cloud. The difference from a general-purpose chatbot is not only answer quality — it is where the data sits.

Gartner: by 2027 task-specific small models deployed 3× more than general LLMs
2 routesConnect to a knowledge base (RAG), or train the model on your data
0 egressIn an air-gapped install, data never leaves the organisation

What separates this from a general tool

A general-purpose tool knows the average of the world; it does not know your business. It does not know your pricing policy, your contract template, the decisions in your past projects or your client's specific requirements — and where it does not know, it invents.

A private model closes that gap: it answers from the organisation's own documents, shows its source, and says so when it does not know.

The second difference, decisive for most organisations, is data sovereignty. Contracts, personnel files or patient records never travel to a third party's servers.

Two deployment routes — which one fits

Connect to a knowledge base. The model stays as it is; the organisation's documents become a searchable knowledge base and the model answers from it. Fast to deploy, inexpensive, updates the moment a document changes. For the large majority of organisations this is the right answer.

Train the model on your data. The model is adjusted to the organisation's language, format and decision logic. Expensive, slow, and repeated whenever the data changes. It only makes sense where the organisation has a genuinely distinctive voice or decision pattern that the first route cannot reach.

Most vendors prefer to sell the second route because it is expensive. We try the first route first, and move to the second only where the first proves insufficient.

A small model is usually the right answer

Gartner projects that by 2027 enterprises will deploy task-specific small language models three times more often than general-purpose large models. The reason is simple: a small model costs less to run, answers faster, and is more accurate within a narrow domain.

Its second advantage is security: because it can run on the organisation's own server or private cloud, the data stays inside.

The assumption that "the biggest model is the best model" is wrong for most enterprise tasks, and expensive.

Air-gapped and public sector deployment

In public institutions and in regulated sectors such as defence, health and finance, a deployment with no internet connection at all is often required. This is technically achievable and is one of our standard modes of work.

In an air-gapped install the real issue is not the model but the update and maintenance regime: if the system never reaches the internet, how do updates arrive, who carries them, and through which approval.

In the Gulf, data residency is not a preference but a redline; in Saudi Arabia, Oman and Egypt sensitive data is expected to remain in country. We design the deployment accordingly.

Where these projects fail

Starting before the data is ready. Scattered, duplicated, contradictory documents do not produce good answers. The real work is usually in document order, not in the model.

Not measuring accuracy. "It gives nice answers" is not a criterion. Before starting we write down which questions must be answered correctly, and we measure the result against that.

Leaving it unowned. Documents keep changing after handover. A system without an update regime loses its reliability within six months.

The two routes side by side

Knowledge baseTraining on your data
Deployment timeShortLong
CostLowHigh
When a document changesUpdates immediatelyRequires retraining
Citing sourcesDoes it naturallyNeeds extra machinery
Organisation-specific voiceLimitedStrong
When it is the right choiceMost organisationsDistinctive decision pattern

Frequently asked

Frequently asked

Will our data be used to train a model?

No. The deployment stays on your side and your data does not go to a third party as training material. In an air-gapped install there is no external connection at all.

How much data is needed?

On the knowledge-base route even a small but well-organised document set works. What matters is not volume but that the documents are current and non-contradictory.

Can it connect to our existing systems?

Yes. Connecting to ERP, document management, CRM and file servers is ordinary practice. Legacy systems get their own integration plan.

Can we change the model later?

We build the system model-independent from the start, so when a better or cheaper model appears you can switch without discarding the deployment. Locking into a single vendor is the most expensive mistake in this field.

What drives the cost?

Document volume and tidiness, the number of systems to integrate, whether an air-gapped install is required, and how high the accuracy threshold must be.

Let us begin

Let us discuss this on your own operation. The first session is free.