In valuation and financial-data work, knowing the difference between a local wrapper and local processing is what protects you.
As interest in agentic systems has grown, so has the number of tools marketed as "Local Agents." But despite the word Local in the name, many of them don't actually process anything locally: with every request, they ship your documents off to an LLM provider — OpenAI, Anthropic, Google, and others. For teams handling confidential financial information, that distinction is the whole game.
To see where "local" ends and "cloud" begins, it helps to understand how an agent is actually built.
What an agent is made of
An agent is typically a stack of three components: an interface, the agent itself (the agent harness), and an LLM.
The interface is how you interact with the agent. It might be a desktop app, a web page, or a command-line tool. Its only job is to let you talk to the agent.
The agent harness is what truly sets one agent apart from another. It's the set of rules the agent follows and the tools it can call — and it's where your data lives: how it's stored, edited, created, and deleted. When a vendor says "Local Agent," this is the component they put front and center: your data sits on your machine, in your database. But running the harness locally is necessary, not sufficient — because there's one more piece.
The LLM is what makes any of this impressive in the first place. It's the language model that serves as the agent's engine: it receives its constraints and available tools as text, your content is appended to that, and the model returns a sequence of actions plus a message for the user.
The part that matters
If the model runs inside your own environment, your data never leaves the perimeter. But the moment an external model is in the loop, every document you process leaves the building. When a "local" agent reads your file, it isn't really the one reading it — the contents are sent to an LLM provider and processed by a third party.
"But providers have data protections"
The usual objection: the leading providers offer zero-data-retention modes, opt-outs from training on your data, and enterprise deployments (Azure OpenAI, Amazon Bedrock). All of that genuinely lowers the risk — but it doesn't change the nature of the risk.
First, you're still trusting a third party — their configuration, their contracts, their audits. Second, the data still physically leaves your perimeter and travels across the network. Local processing removes the question of trust entirely: nothing leaves, there's no outbound network traffic, the system runs in a closed environment (even offline), and regulatory compliance becomes far easier to demonstrate.
What's at stake for finance teams
The cost of a leak here is anything but abstract.
In the EU, processing personal and client data falls under GDPR, with fines of up to 4% of global turnover — and cross-border transfers add another layer of requirements (the fallout from Schrems II, Standard Contractual Clauses).
In the US, you're dealing with GLBA, SEC and FINRA expectations around safeguarding client information, state privacy laws (CCPA / CPRA), and counterparties that increasingly demand SOC 2 certification.
And the assets in play are the most sensitive ones you hold: financial models, M&A deal data, valuation materials, material non-public information, and client NDAs that explicitly forbid sharing data with third parties. A single confidential document run through an external model can breach a contract, an internal policy, and a law all at once.
The flip side of local models
Let's be candid: keeping everything local carries a price of its own. Deploying your own model is a serious capital expense on hardware — GPUs above all — that you have to buy, house, and maintain. These systems need looking after: updates, monitoring, engineers. And as a rule, local models trail the flagship cloud models on raw quality.
That's exactly why many teams end up back in the cloud, risks and all. The choice looks like a trade-off between security and cost.
How Nektus removes the trade-off
Nektus takes that trade-off off the table. Instead of chasing a do-everything model, we've built specialized pipelines purpose-built for valuation. That narrow focus delivers high accuracy and output quality at a fraction of the hardware requirements — which means a genuinely low total cost of ownership.
We solve the real bottleneck in today's models — cost — and let you realize a tangible ROI from deploying agents with no security compromise. Your data stays inside your environment, and the results hold up against cloud solutions at markedly lower cost.
If you're evaluating AI agents for financial-data work, start with one simple question: does my data leave the perimeter? Nektus answers it with an unambiguous no.