On-prem data work can make or break a local LLM. Why isn’t it on the cost slide?
Companies spend serious money standing up a local model so the files stay inside the building. Then they point that model at the share. Demo day looks fine. By week two the answers are junk.
The budget slide that justified the box was full of GPUs, power, and API comparisons. It almost never had a line for making the corpus fit to quote. That missing line is not a footnote. It is often the whole failure.
Call the gap by its name
Public talk about on-prem versus cloud is loud on inference cost: tokens, cards, utilization, and break-even against commercial APIs. Published TCO frameworks itemize hardware and kilowatts. Privacy and sovereignty show up as reasons to go local.
What barely shows up — almost never as a first-class line — is the cost of data processing for that local model: turning a messy share into something a person or a model can actually trust.
That is not a soft oversight. It is the part that fails after the purchase. Your local model can be fine. The corpus is not. Until that sentence is normal in planning meetings, teams will keep buying more model and more search, and then wonder why the answers are still wrong.
We are not saying nobody ever cleans documents. We are saying the usual public cost story does not treat corpus fitness as a line item next to the GPU. Leadership hears “AI is live” when the box answers. The trust problem arrives later, unpaid for on the original slide.
The same hole shows up when a partner writes an on-prem estimate for a client. Hardware and “we’ll wire RAG” are already on the sheet. Document preparation can sit next to them as $10 a brick on this door, only if the ZIP ships.
Two halves of the deployment
A local LLM deployment has two halves. The box runs completions. The corpus decides whether those completions are about your procedures or about a fluent guess. Data work is everything between “we have a folder” and “someone will sign a citation.”
When data work is skipped or faked, the deployment breaks in predictable ways:
Ghost data with a manual’s face
Scanned TMs and shop manuals get chunked as if they were text. The model stares at empty context and invents a typical torque or warning. See scanned procedures the model cannot read.
Hallucination fuel
The extract still looks like English, but it is broken: doubled letters, mashed tables, OCR junk with a number sitting in it. Retrieval surfaces it. The model quotes it with confidence. See tables that still invent numbers.
One melted index
Ten years of SharePoint land in one index. The 2014 procedure answers the 2024 question. The model is still sure. The other choice is a bounded package you can replace when that subject changes.
No mute list, no board
Nobody can say what the model is allowed to quote. Responsible people will not put their name under it. The “deployment” becomes a demo toy.
When data work is done, the same local model suddenly looks competent — because it is answering from a package that was manufactured for that job: structured where it can be, junk off the answer path, limits listed.
The GPU did not get smarter. The corpus got honest.
Other serious approaches (fairly)
Extractors
They get words out. That is useful, and cheap per page. They do not by themselves produce a leave-behind someone will sign.
RAG
Retrieval searches whatever you already have. It is the right tool after the corpus is fit to search — not a substitute for manufacture.
Internal plants
Platform projects and “AI centers” can build plants of their own. When they work, they still end up with something like a manufactured package, a mute policy, and a trust story — not an infinite dump. Six figures of internal time to arrive at the same architecture is a real option. It is not free.
So the claim is not “only we exist.” The claim is simpler: pointing a local model at an unmilled share is a bad default, and manufacturing a residual-honest, portable unit is a sound, competitive method for the job public TCO slides forget. Argue with public bricks and HF retrieval eval if you disagree.
One more use for the same artifact: if you later bake domain knowledge into weights, an honest brick is also a clean training set — and the listed mutes tell you what not to bake in. Context-time today; weights-time later if you choose. Same leave-behind.
The missing line, bought once
The line item the slide forgot is not “more search.” It is a portable ZIP you keep, $10 a brick, only if it ships, so the model has documents it can use.
Drop the pile on the mill when you want that line filled without a staffed project. You keep the ZIP. No honest brick, no charge. Last word: your trusted people. Price detail lives on the door notes; this card is the argument for why the line belongs on the slide at all.