AI document management for accounting firms is rarely a priority — until it becomes a necessity.
Most accounting firms survive by cutting costs. The margins are thin, and the instinct is to defend them by spending less – so equipment ages, software stays frozen, and the way the work gets done belongs to an earlier decade. Plenty of firms still run the accounting package they installed years ago – the same every firm in the industry uses – and have not touched the workflow, the processes nor the customer service ever since. When a firm of this kind does invest, the money tends to go into new premises or a “brochure” website. The work itself – how documents move through the office, how clients are served, how the firm finds what it already knows – stays as it was.
Against that background, Debitus is worth a closer look. It is a Portuguese accounting firm that decided to do the thing its industry almost never does: change how it operates, on purpose, before anything forced it to.
“I either do this or I die”
The decision came from the top, and it was not hedged. When Rui Costa, Debitus’ CEO, told his team “ou faço isto ou morro” – “I either do this or I die” – he was not reaching for a metaphor. He meant that the firm’s trajectory, and the industry’s, does not hold unless the way the work is done is rebuilt. In a profession where this kind of investment is postponed almost as a reflex, that clarity is rare.
The point was never digitisation as such. What Rui wanted was an operational strategy that touched every process in the firm – filing, client communication, reporting, reconciliation. ArquivoDigital, built by Swell, is the first piece of it.
The proof: Arquivo Digital
Arquivo Digital works as an operational layer over the firm’s core processes. Documents arrive through several channels and are captured, classified, and filed without anyone handling them by hand. The system extracts the content, gives each document a type, and indexes it so it can be found by meaning rather than by filename. The whole team works from one searchable, auditable repository.
The early numbers were concrete enough to be persuasive. In the first few days we loaded more than 160 clients, and the model was already classifying hundreds of documents, their content indexed and searchable. Client sharing, notifications, and collaboration went in early, so the platform served the people inside the firm and the clients on the other side from the start.
And not everything worked first time. About 5% of documents were not being read at all, and the pattern was clear: they had arrived on paper and been scanned. A degraded scan defeats text extraction, and with no text there is nothing to classify. The fix was to swap the Azure Document Intelligence model for one that holds up better against poor originals. It is the kind of problem that never appears in a demo and only shows up with real documents, at real volume.
Here is what the interface looks like:
The hard part is choosing
The team made the project easier than it might have been. The vision came from the CEO, but the whole team took it on. They wanted it to work, and that changes everything about a project like this.
That enthusiasm brings its own difficulty. Technology can now do almost anything in a firm like this: classify documents, answer questions, generate reports, reconcile accounts, forecast cash flow, talk to clients. Building any one of them is seldom what makes a project hard. The hard part is choosing which to build, and in what order.
Every conversation with Rui and his team began in possibilities and had to end in focus – not because the other ideas were weak (many were excellent), but because a transformation that tries to do everything at once does none of it well. The discipline this demands is editorial: deciding what comes first, what comes next, and what waits.
Starting with the document flow was that decision. Documents are the least glamorous part of accounting and the most foundational. Get that base right and everything built on top is worth more; get it wrong and even the most sophisticated analytics cannot be trusted.
Why it won’t stop
This is the part that interests me most about Debitus, and it has little to do with any single feature. The first move did more than make the firm more efficient. It put Debitus somewhere its competitors cannot reach by cutting costs – and that position only widens as the next capabilities arrive.
A firm that starts investing in AI tends not to stop, for a structural reason. Every few months, new capabilities make cheap what used to be expensive, and each one opens a door to the next. The investment compounds: each step lowers the cost of the one after it, the target keeps moving forward, and the firm keeps following it. The roadmap already points that way – reporting and analytics in Power BI, reconciliation automation, a client portal, a conversational assistant that lets clients ask questions about their own data in plain language.
There is a human side to this that the technology tends to hide. As processes are automated, the work does not vanish; it moves. The Debitus team will spend less time on repetition and more on judgment: analysis, advice, the client conversations that need a person. That shift is demanding in a way no software is, because it asks people to give up tasks they have done for years and take on roles that are less defined, more exposed, and worth more for exactly that reason. Pushed far enough, the firm Debitus becomes is one where almost nobody does the accounting itself and almost everybody does the work around it.
An exception, not a wave
I would not claim Debitus is the start of a trend. The honest position is the opposite: most accounting firms will go on cutting costs, running the old software, and putting up a new website when they invest at all. That is precisely what makes this firm worth writing about. It is an exception – and once an exception like this starts to move, it rarely moves back.
The architecture, for those who want it
ArquivoDigital runs on a Microsoft stack. The frontend is a Power Apps Code App in React and TypeScript; the data sits in Dataverse, with SharePoint Online for storage, separating inbound documents, triage, and final archive. On the backend, Azure Functions orchestrate upload and OCR, Azure Document Intelligence extracts the text, and OpenAI classifies each document by type, journal, and likely client. Power Automate handles email ingestion and routing. The point of the stack is governance: a custom application with a full audit trail, built inside the Power Platform.
If you’re looking to automate your firm’s document workflows with AI and Power Platform, see how we can help.
