An AI assistant is only as good as the documents and permissions it is given. Three things to put in order before any project: clean data, access and an owner.
When a company decides to use AI, the discussion usually starts with choosing the tool: which model, which platform, what it costs. Experience shows that problems more often come from elsewhere: documents are old or contradictory, everybody has access to everything, and when something goes wrong it is unclear who answers for it. An AI assistant does not fix these things; it only makes them more visible and faster.
This article proposes three things to put in order before you connect an assistant to company data: clean-up, access rights and responsibility.
Why data matters more than the model
An assistant that answers from company documents (the mechanism explained in our article on RAG) does not “know” more than what it finds. If it finds three versions of the same contract, two price lists and a procedure replaced two years ago, it will answer confidently based on one of them. It will not know which is the right one unless you tell it.
The rule is old in computing and applies in full: garbage in, garbage out, except that now the garbage is nicely phrased and convincingly cited.
Step 1: the inventory
Before the clean-up, you find out what you have. A useful inventory is not a 50-page document but a short table with a few columns:
- what data: contracts, quotes, emails, procedures, customer records, invoices, support material;
- where it sits: shared folders, email, CRM, accounting software, employees’ laptops;
- who produces it and who uses it;
- what kind of data it is: containing personal data, trade secrets, public information;
- how current it is.
Usually half of the surprises appear here: important documents on a single laptop, old folders with open access, copies of the same file in four places.
Step 2: the clean-up
You do not have to clean everything. Start with the data you actually want the assistant to use, that is, with the first use case. The most frequent problems:
- Duplicates and multiple versions. Choose the valid version and archive the rest in a separate place the assistant cannot see.
- Expired information. Old prices, replaced procedures, contacts who have left the company. Mark the validity date or remove the document.
- Contradictions. If two documents say different things, decide which takes priority and write that down.
- Difficult formats. Scans without text, tables pasted as images, protected PDFs. A person can read them; an automated system may not.
- Chaotic names. “final final2” tells neither a person nor a system anything. A clear name and a date help both.
- Data you do not need. The GDPR principle of data minimisation says you keep only the data necessary for the purpose, no more and no longer than needed. Cleaning up for AI is also an opportunity to apply this principle.
A suggestion: for every important document, note on the first page who wrote it, the date of the last review and whom it applies to. That is three lines, but they make the difference between a document you can count on and one you only hope about.
Step 3: access rights
This is where the most serious risks are. The practical rule: the assistant should see exactly what the person using it is allowed to see, no more.
A hypothetical example: in a company, the “Contracts” folder is accessible only to management. An assistant is connected to the whole shared drive, using a technical account with access to every folder. An employee asks the assistant about the terms of a contract and gets the answer, although they could not have opened the file themselves. The assistant did not “break in”; it used the access you gave it.
What you do, concretely:
- Check the existing permissions. Many people find that “private” folders are in fact open to the whole company. Fix that first, regardless of AI.
- Connect the assistant with the user’s rights, not with an account that has full access, where the tool allows it. If it does not, create separate assistants for sets of documents: one for internal procedures, another for quotes, with no HR or accounting content.
- Separate sensitive data. Health data, employee data, information that identifies customers and trade secrets do not go into the first pilot unless you have a clear reason and legal agreement.
- Keep logs of accesses, so you can reconstruct what someone saw.
- Review periodically who has access and remove rights for whoever has left or changed role.
The GDPR requires appropriate security of processing and data protection by design (Articles 5, 25 and 32). How exactly that applies in your company is a question for a lawyer or the data protection officer.
Step 4: who is responsible
The third piece is the most neglected, because it is not about technology. For every set of data used by an assistant there must be a named person, not “the department”.
A simple model with three roles:
- The data owner: answers for the correctness and currency of the data (“the price list is valid and up to date”). Usually the head of the department that produces it.
- The assistant’s owner: answers for how it is configured, which documents it sees, testing, and stopping its use if problems appear.
- Who decides in case of disagreement: usually management. Without this role, discussions about “who got it wrong” never end.
Besides roles, set a few simple rules:
- how a document is updated, so the assistant does not keep the old version;
- what to do when someone notices a wrong answer: whom they report it to and how quickly it is corrected;
- how long conversations and logs are kept, and who can read them;
- what is not entered into the assistant (for example passwords, card details, health data), written in two lines that the whole team knows.
A simple test before launch
Prepare 15 to 20 real questions, with the correct answer known beforehand, including a few to which it should answer “I do not know” or “I am not allowed to see that”. Run them with accounts of different levels. If an ordinary employee receives information they should not, or if the assistant answers confidently from an expired document, you have found a problem while it still costs little to fix.
What not to wait for
The clean-up never ends, and it does not have to end before you start. A small pilot, on a single well-kept set of documents, teaches more than a year of preparation. What matters is that something remains from each step: an inventory, an access rule, a named owner.
At ME Digital we start AI projects with these three checks and configure the integrations so that data is not used for training; see AI implementation or get in touch, while the legal side stays with your own lawyer.