An AI assistant on your own documents: what it really does and where it stops
How an assistant that reads your company archive works, the six things it does well, the five limits it must not cross (starting with legal and tax opinions) and the usage rules to give your team.
“Now with AI” appears in every software proposal. On its own the phrase says nothing: it can mean a better search, or a system promising to read your contracts for you. The difference is large, and it lies mostly in the second part, the one nobody raises during a sales call: where it stops.
This article explains how an assistant that works on your company’s documents actually operates, what it does well, and which limits it must not cross, not out of commercial timidity but because crossing them creates a problem for you.
How it works, in two minutes and without jargon
The mechanism is simpler than it sounds, and understanding it helps you trust it in the right way.
When a document enters the archive, the text is extracted: directly from a digital PDF, or through OCR for a scan or a photo. The text is then split into chunks, and each chunk is turned into a series of numbers representing its meaning. Chunks about similar things end up near each other.
When you ask a question, this happens: the question is turned into the same kind of numbers, the system retrieves the closest chunks, and hands those chunks to the language model along with your question. The answer comes from those chunks, not from whatever the model learned somewhere on the internet.
Two practical consequences follow, and they are worth remembering:
- the assistant only answers about what it has read. A document you never uploaded does not exist for it;
- the assistant can tell you which document an answer came from. If it does not, you are using the tool badly.
The six things it does well
Searching by meaning rather than exact wording. You look for “the maintenance contract with the penalty on response times” and you find it, even if none of those words appear in the file name.
Summarising. A forty-page contract, a specification, a long set of minutes. Useful when you have to decide whether reading the whole thing is worth it.
Extracting data. Number, date, amount, deadline, counterparty. This is the task that saves the most time, because today it is done by retyping.
Translating. A quote received in German, a certificate in English, a document in Turkish. Enough to understand and decide, not enough to sign without a human read.
Comparing two versions. What changed between the March draft and the June one. The model sees the differences even when nobody marked them.
Answering questions about the archive. “Which contracts expire before November?”, “what documents do we have from this supplier?” Precisely the questions that cost somebody twenty minutes today.
The five limits
This is the part that matters, and the part sales material rarely contains.
1. No legal, tax or medical opinions
This is a hard limit, not temporary caution. The assistant can find the termination clause, show you what it says and summarise it. It must not tell you whether you may terminate, whether the clause is enforceable, or how to respond to a claim.
There are two reasons. The first is substantive: judging a specific case requires knowledge that is not in the document and professional responsibility that a machine does not carry. The second is practical: if you act on a wrong opinion, the damage is yours, not the software’s. In several countries, giving legal advice is also a regulated activity reserved to licensed professionals.
The formula to remember is simple: the assistant organises facts, it does not judge the case. Where the facts end, your accountant or your lawyer begins.
2. It does not decide instead of a person
It can propose that an invoice comes from a given supplier, for a given amount, due on a given date. Approval remains an act by a person, with a name and a timestamp in the log. Automation that approves on its own is convenient right up until the wrong invoice arrives, and then nobody knows who decided.
3. It is never more reliable than the documents it reads
If the document is wrong, the answer will be wrong with exactly the same confidence. The assistant does not verify facts about the world, it reports what the archive says. An amount misread in a scan stays wrong even when it is delivered to you in a convincing tone.
4. It can be wrong, and should be used accordingly
A language model can produce a plausible and incorrect answer, especially when the question is vague or the document ambiguous. This is not a defect an update will fix, it is a property of the tool. So the correct use is always the same: ask, then open the document it cited and check the passage. It takes ten seconds and changes the risk profile completely.
5. It does not replace proper retention
An assistant that reads and summarises is not a retention system, and it does not make redundant the requirements we covered in the article on audit-proof archiving and procedural documentation. These are two different functions: one makes you faster, the other keeps you compliant.
Three rules to give the team
Deliberately only three, so everyone remembers them.
- Every answer is checked against the source. The assistant names the document: open it. If it names none, the answer is not used.
- Ask questions with checkable answers. “When does this contract expire?” is a good question. “Should I renew it?” is not.
- Do not paste in data that is not in the archive. If a piece of information matters, it belongs in the system as a document, not in a chat thread.
Where your documents go
This is the question a company should ask first and usually asks last. Three answers to insist on:
Isolation. Your documents should not sit in the same container as another company’s. In Blina Space each company gets a dedicated database, not a row tagged with a customer id in a shared table.
Purpose limitation. The content serves to answer your questions, not to train general-purpose models.
Location and rules. Hosting in Germany and processing under the GDPR are not a formality when the archive holds contracts, payroll and customer data.
When it is genuinely worth it
To be honest, not every company needs this to the same degree. The assistant earns its cost when at least two of these are true:
- a lot of documents arrive from outside and somebody retypes data from them;
- the archive is large enough that searching has become a problem, roughly beyond a few thousand documents;
- you work in several languages;
- there are contracts and deadlines somebody has to keep an eye on.
If you have two hundred documents and know them by heart, a good search is enough, and anyone telling you otherwise is selling.
How it looks in Blina Space
The assistant works through four declared actions: summarise, translate, analyse and compare documents. Alongside sits search across the content, which works on scans too thanks to OCR, and automatic data extraction on incoming invoices.
Usage is measured in actions: Starter includes 50 a month, Team 500, Business 2,000. A number you can keep an eye on, rather than an “unlimited” that turns out to be limited later.
And the boundary stays as described above: the assistant works only on your company’s documents, shows where an answer came from, and stops short on legal and tax cases. Not because it could not write something, but because that something is not what you should be acting on.