Where to run the finance model so the data stays home

The controller has a new laptop and a plan. The model that drafts the variance commentary will run on it, locally, so no ledger line ever goes to a hosted service. It works on Monday. On Tuesday she starts an overnight run that reconciles budget, forecast and actuals across forty cost centres, closes the lid at seven and takes the laptop home. At seven fifteen the run stops, because the machine it lived on is now in a backpack.

Nothing leaked. The plan simply confused two questions: where the data is allowed to go, and where the work is allowed to live.

Why the question is new this autumn

The hardware for local models in finance changed in October 2026. NVIDIA put RTX Spark laptops on sale on 16 October: Windows on Arm, one chip combining a Blackwell RTX GPU and a Grace CPU, up to 128 GB of unified memory and, by the manufacturer’s figures, up to one petaflop at FP4, from USD 2,599. NVIDIA ships them with local agents and with Nemotron 3.5 Bolt, a 24 GB model it describes as tuned for tool calls. The pitch is that the PC becomes “an everyday teammate”.

A week later NVIDIA published its Secure Agent Workspace reference design for agents that run for hours on a person’s behalf. Its central idea fits our controller exactly: the user’s device is only the presentation layer, and the agent works in a managed environment that persists independently of the session. In NVIDIA’s words, the workspace is not the session.

The real cause: sovereignty is about the data, availability is about the place

Keeping finance data inside the company, the argument made in why financial data should never leave the building, says nothing about which machine inside the company should do the work. A finance agent that runs at night has three requirements a laptop cannot meet: it must be on when nobody is, it must not carry one person’s credentials, and its writes must wait for a person who is awake.

Where it goes wrong

The laptop becomes the server. A model on an AI PC is excellent for the controller’s own work during the day: drafting, asking questions of a local extract, checking a reconciliation. It is the wrong home for a process that runs overnight, because its availability is the owner’s commute.

The agent inherits a person’s access. If the night run uses the controller’s ERP login, then everything it does is, on paper, something she did. The reference design keeps raw credentials out of the agent’s process altogether.

Writes go straight into the ledger. A night agent that may post accruals or reclassify entries on its own turns a drafting tool into an unreviewed bookkeeper. NVIDIA’s design calls for human review of sensitive writes through governed connectors; in finance, almost every write to the ERP is sensitive.

“Local” quietly becomes “somewhere”. Moving the work off the laptop is right. Moving it to whichever cloud is closest to hand, without deciding which data may go there, undoes the reason for running the model locally in the first place.

What to do instead: three places, three jobs

The AI PC for the person. An RTX Spark-class laptop, or any machine with enough memory for a mid-sized open-weight model, is where the controller works with the model interactively. Data on it stays with the person using it, and nothing runs there that has to outlive the session.

A desk-side station for the always-on agent. A small machine in the company, such as NVIDIA’s DGX Spark station, is a better home for the night run: it stays on, it can run in its own environment with its own service identity, and it keeps the model and the data in the building. The laptop becomes a window onto it.

A European AI factory for heavy, scheduled work. Some jobs, such as re-estimating a model across years of history, need more compute than a station has. Europe is building capacity for this, including the EuroHPC AI factory in Poznań, and NVIDIA has announced modular AI factories with deliveries from early 2027. It is still outside the company, so it should get anonymised or aggregated data under a written decision, never the raw ledger by default.

Across all three, the same rule for writes: the agent proposes, a person approves, and only then does anything reach the ERP. That is the shape of a continuous close: the loop runs all night, and the decisions wait for the morning.

ObserBI is pre-launch. It is designed around this split, with local models on your own hardware and read-only connectors to the ERP, and it does not claim certification or integration with any of the products named here. More on the ObserBI homepage.

A test you can run today

List every task you want an AI model to do in finance this quarter. Next to each, write two things: does it have to run when nobody is logged in, and does it write anything to the ERP or a bank system?

Anything that runs unattended does not belong on a laptop. Anything that writes needs a named person who approves it before it lands. Anything you were planning to send outside the company needs a sentence explaining which data it carries and who decided that. The tasks where you could not fill in those columns are the ones to sort out before the hardware arrives.