AI-agent infrastructure for accounting & bookkeeping firms
Invoice processing, financial statement support and file research with a fully self-hosted AI agent: nothing in your client administrations ever leaves the building.
An accounting or bookkeeping firm by definition works with information that must never end up anywhere else: financial statements, tax returns, payslips, correspondence covered by fiscal confidentiality obligations. The very thing that makes AI so valuable for this line of work is processing and searching large volumes of financial documents at speed. That's exactly why "just paste a financial statement into ChatGPT" isn't an option in this line of work. Every prompt to an external AI provider is client administration leaving the building.
That's why we set it up the other way around: not the data going to the model, but the model coming to the data. On the BedrijfsLab, the language model itself runs entirely on hardware that stays yours. For sectors where confidentiality genuinely matters, such as healthcare, law or accounting, that's often the deciding reason to choose self-hosted: no tax return, no draft financial statement and no client file that has ever touched a server belonging to an American AI provider.
For the technically curious
The language model runs via Ollama, llama.cpp or vLLM, set up on hardware at your own premises.
The concrete use case
On that self-hosted foundation, we deploy an agent with its own, clearly bounded task, the same approach as our broader AI-agent infrastructure. For an accounting or bookkeeping firm, that means for example:
Invoice processing & matching
Incoming invoices get read in, matched against purchase orders or existing bookings, and booked or queued for approval, so nothing needs to be manually retyped.
Financial statement support
Source documents and ledger entries get structured and summarized into a draft financial statement, so an accountant mainly has to check it rather than build it up from scratch.
File research
Prior tax returns, correspondence and sources get searched and summarized on the specific point an employee is asking about, without a complete file having to be reviewed by hand.
In every case, the same principle applies as with our broader unmanned office: if something falls outside the agreed boundaries, the agent escalates to a human instead of making a fiscal or financial judgment call itself. The agent takes the processing work off your hands, not the assessment.
Why this is safe with client administrations
The approach we set up every agent with is particularly well suited to work covered by confidentiality obligations:
A branch and pull request per change
Nothing gets applied directly; there's always a readable trail of what changed and a clean point to roll back to.
Isolated working copies
An agent working on a file does so in its own sandboxed environment, never in a copy someone else might be working in at the same time.
An independent check
A second, independent step verifies the actual result, not just the agent's own summary of what it thinks it did.
A fixed list of things asked first
A human stays ultimately responsible for every fiscal or financial judgment call; the agent takes the groundwork off your hands, not the decision.
For an accounting or bookkeeping firm, that's not an incidental detail. It's the core of it: the same diligence you'd expect of yourself with a client file applies to the agent looking over it too.