Two years ago, running your own language model meant a research team and a server room. Today, open-weight models that rival the commercial chatbots for office work run on a single well-specified machine or in a private cloud tenant that only you can reach. The question for most organizations is no longer whether they can, but whether they should, and for which work.
What “private” means
There are three tiers, and the words get used loosely. A local model runs on hardware you own, in your building or your data center; nothing leaves. A private tenant runs in a cloud account that belongs to you, isolated from other customers, with the model and your documents inside it. A commercial enterprise tier is a vendor's chatbot with a contract that says your data is not used for training and is deleted on a schedule. All three are better than a free consumer account. Only the first two keep the data in your hands.
What it changes
Your documents stay inside the boundary. Staff get an assistant that already knows your manuals, your contracts, your policies, your past reports, and answers from them. Every question and every action is logged where you can read it. Costs are a fixed shape (hardware or a tenant plus maintenance) instead of a per-seat bill that grows with use. And when a law or a trustee asks where the data went, the answer is one sentence.
What it does not change
A private model can still be wrong. It still needs the same rules about human review, the same approval gates in front of anything that affects a person, and the same policy card on the desk. Private is a data decision, not a judgment decision.
The one rule
Sort your work by what the data names. Anything that names a person (a member, a patient, an employee, a client, a donor) belongs inside. Anything that names your organization's private material (curriculum, contracts, financials, strategy) belongs inside. General drafting, research and public information can live in a well-configured commercial tool. Write that split down by class of data, and the private-versus-public question answers itself for every new use.
What it takes
A plan that names the first three uses. A machine or a tenant sized to them. Connections to the systems the work lives in. Four to eight weeks, and one person on your side who will be its champion. We deliver on Parseek, our private AI platform, when it fits, and recommend something else when it doesn't. Either way, you own the setup and can leave with it.
Share this note
Forward the link, or email it to the person who should read it: send by email.