One dossier number in ChatGPT, and the AI reads the file in Accelerate, the client's emails and the books in Nmbrs. Read-only, EU-hosted, nothing goes out without a human.
Landsman & Teune is a mortgage advice practice and a bookkeeping office in Kampen. They already used ChatGPT, but uploaded every document by hand. I built a connector that lets their own ChatGPT Business read dossiers in Accelerate, the shared inboxes in Outlook and client books in Nmbrs, plus a document service where a person approves every advice letter.
Financial advice / Bookkeeping
ChatGPT Business, MCP, Accelerate API, Microsoft Graph (Outlook), Nmbrs Accounting, Python, Docker, Hetzner (EU)
3 systems connected to ChatGPT, read-only, with human approval on every advice letter
Rein Teune already worked with ChatGPT before we met. To write an advice letter, Rein would download the rate offer, the purchase contract, the financing overview and the application, upload them one by one, and prompt. It worked. It was also slow, and every upload was a document full of personal data going in by hand.
The wish was one sentence: "I just want to give it the dossier number, and the rest gets fetched."
The firm is really two businesses. The advice side works in Accelerate, the platform many Dutch mortgage and financial advisers use. The bookkeeping side keeps around 246 client administrations in Nmbrs Accounting (formerly Yuki). And everything comes in through a shared Outlook inbox. Three systems, none of them talking to ChatGPT.
An honest note: when we started I told Rein that Accelerate had no API to read dossiers. The public documentation suggested so. I asked Accelerate directly, and it turned out there is a full API behind the licence. I corrected myself in writing. You only find that out by asking the vendor, not by asking ChatGPT.
The connector is a small server that ChatGPT Business can talk to (via MCP, the standard way to give an AI tools). Rein logs in once, and from then on ChatGPT can:
Most client documents are scans, so text recognition runs on the same server: a payslip in about 10 seconds, an 8-page contract in about 18. Nothing goes to an extra party for that.
This was the part I spent the most time on, because it's a firm full of financial data.
Rein and ChatGPT had already written their own method and Word template for the advice letter. I built the service around it: ChatGPT prepares the input, the service renders the letter in their template and runs their own checks, and then a person approves it on a separate page.
That approval is tied to a fingerprint of the exact input and the exact document. Change one number, and it's a new version that needs a new approval. An approval filled in by the AI is refused. A letter without amounts is blocked. The human releases the letter, never the AI.
The first time we filled the work queue it found 2,466 signals over 31 days, of which 187 still open, split per adviser. ChatGPT now starts the morning with that list.
The connector is half of it. The other half is the flow on top: every morning the assistant prepares the work queue, drafts the advice letter from the dossier and runs the firm's own checks. Nothing leaves the office until an adviser has read it and released it. Dossier numbers and names below are fictitious.
The pattern here isn't specific to mortgages. Any AI assistant becomes ten times more useful the moment it can read your own systems, with clear limits.
Who it fits: accounting and bookkeeping firms, mortgage and financial advisers, notaries, recruitment agencies, any office where the answer sits in three systems and a pile of PDFs.
Yes. Accelerate has an API behind the licence. With a connector, ChatGPT Business can search and read dossiers, mortgage details, documents and the work queue. Writing back is possible too, but here it was deliberately left out.
Yes, through the Nmbrs API: administrations, balances, transactions, open items and the document archive. Nmbrs allows a limited number of calls per day, so a good connector caches what it can.
It can be, if you design for it: a business account where data isn't used for training, a connector that only reads what's needed, EU hosting, no content in logs, and a data processing agreement. The point is that the rules are enforced by the software, not by good intentions.
Not in this setup. It can prepare a draft, but sending and writing are technically impossible. Every advice letter needs a person's approval.
Yes. The connector uses MCP, an open standard that ChatGPT, Claude and others support.