How do you use generative AI when your client files can never leave the building? CerFrance Maine-et-Loire, an accounting firm, found its answer in Understand Tech's AI-in-a-Box, an on-premise AI appliance powered by NVIDIA GB300. Today its legal team drafts company statutes in half the time, an AI agent processes 500,000 invoices a year, and the AI budget stays fixed. Not a single client file goes to the cloud.

"For us, it completely transforms the way we work."
Michel Besançon
CEO, CerFrance Maine-et-Loire

Key results: the ROI of on-premise AI in an accounting firm

On-premise AI results at CerFrance, an accounting firm: 50% less time per statutes file, up to 2 FTE freed, 3-minute drafts

50% less time per company statutes file, measured by the legal team on real files.
Up to 2 FTE freed in the legal team, with more files handled by the same people.
3 minutes to generate a statutes draft, instead of 1 to 3 hours by hand.
500,000 invoices a year read, classified and posted by an AI agent.
Fixed cost, no token limits, whatever the volume.

The challenge: building a private AI factory for an accounting firm

CerFrance Maine-et-Loire is an accounting and advisory firm in the national CerFrance network, supporting 5,000 business leaders. Its teams handle tax returns, payroll, bank statements, company statutes and client files every day, all covered by the professional secrecy of chartered accountants.

Generative AI could take on much of this document-heavy work. But public AI ran into two walls: client files cannot leave the firm, and at 500,000 invoice images a year, a per-token cloud bill made no sense. What CerFrance needed was its own private AI factory: an on-site platform where its teams run models, build apps and process data at scale, under their own roof.

"Every business leader today pays close attention to their data. The idea of that data going to the cloud is not acceptable, not for our clients and not for us. So there is a security issue, and above all an economic one."
Michel Besançon
CEO, CerFrance Maine-et-Loire

Watch Michel explain it in his own words

Why AI-in-a-Box: on-premise AI powered by NVIDIA GB300

CerFrance wanted a complete AI solution it could own and run itself. AI-in-a-Box brings the hardware, the models and the applications together in one appliance, installed inside the firm's network.

Why CerFrance chose Understand Tech's AI-in-a-Box: on-premise AI with data sovereignty, ready-to-use apps, no-code AI apps and a fixed cost with no token limits

"Our customers should not have to choose between AI capability and data control. GB300 lets us stop asking them to."
Naama Bak
CEO and co-founder of Understand Tech

Generative AI use cases in accounting: what CerFrance runs on site

Since September 2026, a Dell Pro Max GB300, built on NVIDIA Blackwell Ultra, runs the full Understand Tech stack on site. It keeps working even when disconnected from the internet. Four business teams use it every day:

Legal: company statutes. The app turns the client's source data into a complete draft of the statutes, which the legal team checks and finalizes. Built with Understand Tech, it was the first app CerFrance put into production.
Accounting: invoice processing. An AI agent reads invoice images, classifies them and posts them to the accounts. Accountants review the entries.
HR: assistants and documents, built by the HR team itself in Understand Studio, Understand Tech's no-code App Builder.
Training: learning apps for the firm's staff and its clients, also built in-house.
Every team uses the bundled apps: Assistants for internal procedures, Chat with CSV for accounting and payroll exports, Multi-LLM Portal for first drafts of client replies, and coding assistants.

At every step, the AI prepares and drafts, and the professional checks, decides and signs.

Measured results: AI document automation for company statutes

The legal team timed the statutes app on real files, for two types of company. In practice, the same team can now handle up to three times as many statutes files in the same time.

Type of company	Before AI (range)	With AI	Time saved
Single-partner company	2h to 2.5h	45 min	62 to 70%
Company with four partners	4h to 5.5h	2h	50 to 63%

What changes: predictable AI costs and no-code AI apps

Predictable costs. The appliance comes at a fixed cost, with no token limits and no usage-based bill. CerFrance knows its AI budget in advance, however much its teams use it.
Business teams build their own tools. New use cases go from idea to app in Understand Studio, with no new procurement cycle.
Faster and more reliable client service. The applications increase production capacity and help the firm answer client needs as quickly, and above all as reliably, as possible.

From one app to an AI factory: what's next

CerFrance started with one app, statutes drafting. Its own teams then built apps for accounting, HR and training, all on the same appliance, now the foundation of the firm's own AI factory. Next, CerFrance will measure the accounting agent and the other Studio apps against the same 50% target, with Understand Tech as a partner it is growing with.

Talk to our team about on-premise AI

AI-in-a-Box comes in two sizes, built with our partner Dell: NVIDIA GB10 at $25,000 a year for a single team, and NVIDIA GB300 from $150,000 a year for a whole organization like CerFrance. Rack configurations are available on request.

Want to see what AI-in-a-Box could do for your firm, with your data staying on site? Book a call with our team.

Prefer to explore on your own first? Discover AI-in-a-Box in detail.

On-premise AI for accounting firms: 50% less drafting time with AI-in-a-Box