Start with the questions that really come up
In a small business, certain questions come up every week: where is the return procedure, which sheet corresponds to this product, who validates a particular order? Sometimes the answer exists, but it's split between a shared folder, an old attachment, and a colleague's memory. Before choosing a tool, write down the ten questions the team asks most often.
For each question, indicate the document intended to answer it and the person who can confirm the answer. If no one finds the source or if two colleagues give different versions, the first problem is documentary. Adding AI to this mess may just make the error more convincing. An organized file and a classic search can already solve a large part of the difficulty.
Then choose a limited scope, for example public notices for a range of products. Avoid starting with all of the company’s files. A narrow topic makes it easier to recognize wrong answers and missing documents. It also makes the cost of the test more understandable: you know how many files have been prepared and what questions need to be covered.
Understand how it works in three steps
The principle sought is the following: the system receives a question, finds passages in the authorized documents, then prepares an answer from these passages. This method is often called RAG. This acronym designates a response enriched by documentary research. The public service Albert API presents this possibility in its offer for administrations. See this official explanation.
The presence of a search does not guarantee accuracy. The system may select the wrong passage or rephrase it incorrectly. This is why the expected output must contain the file name, its date or version, and the passage that justifies the response. A vague reference to “your documents” does not verify the work. The user must be able to open the source.
You also have to decide what the assistant does when it finds nothing. A correct answer may be: "I cannot find this information in the available documents." » This sentence has more value than a plausible invented answer. To learn to distinguish these cases, the guide check an AI response offers point-by-point verification.
Prepare a clean file before the first test
Create a small folder with public, fictional, or authorized documents for this test. Give readable names to files and keep a baseline version. A name like “procedure-return-version-2026-09” helps more than “new-final-2”. Remove duplicates from the test folder without destroying the original archives. The goal is to define what is authentic, not to reorganize the entire company at once.
Then check that the content is actually readable. A PDF composed of photographs may require text recognition. A table can be misinterpreted if its columns get mixed up during extraction. Open the documents and test a few passages. If the tool does not correctly read a unit, reference, or exception, correct this step before evaluating the quality of its responses.
For internal or personal data, the choice of service and access rights must be examined before import. The CNIL offers resources for small businesses considering generative AI. They help to ask questions about information protection and organization. Consult CNIL resources. A test on fictitious data allows you to move forward without immediately exposing real files.
A fictitious example to reproduce with your team
Let's imagine a company that has twenty public notices of equipment. This scenario is educational; it does not describe an ISS client. The objective is to help a colleague find the characteristics of a product without going through each PDF. We choose five simple questions, three questions which ask to compare two documents and two questions which the notices do not answer.
The first question could be: “Which document indicates the dimensions of model A?” » The answer must give the reference and the exact passage. A comparison question might ask what dimensional differences appear between A and B. The assistant should maintain the units and report missing information. He should not complete a picture by guessing what the manufacturer would probably have intended.
An unanswered question could relate to a commercial availability date, absent from the notices. The expected behavior is a statement of absence, with a suggestion to check with the person responsible. We are not waiting for an invented date. These unanswered questions are essential: they show whether the system knows the limits of the file or whether it tries to satisfy any request at any cost.
A starting instruction to copy and adapt
You can start with this instruction: “Use only documents in this folder. For each answer, indicate the name of the file and the passage that justifies it. Retains numbers and units. If two documents contradict each other, present both versions. If information is missing, say so clearly. Don't turn a hypothesis into fact. » This instruction sets a behavioral objective; it does not replace technical controls.
Then add a specific request: “Compare the dimensions indicated for products A and B. Separate the information present from the elements to be requested from the supplier. » A limited request is easier to control than a general order like "scan all my documents." It also allows a colleague to take the test again and check whether the answers remain consistent.
Do not present this instruction as sufficient protection against any problems. The accessible files, permissions, and how the tool uses sources matter as much as the requested text. Documents can also contain extraneous instructions: a document assistant must treat them as content to be analyzed, not as orders that change its operation.
Measure usefulness before adding connections
Prepare a grid with the question, the correct source, the expected answer, the result obtained and the necessary corrections. Also note the typical research time and the total assisted time, including proofreading. For ten questions, keep all ten answers. Don't write off failures just because they involve a poorly presented file: this type of document may be part of the actual work.
Define errors that prevent you from continuing. Bad punctuation does not carry the same weight as an invented product reference or a confidential document shown to the wrong person. If access rights are unreliable, the test should not be extended to more data. If the answers cite the wrong versions, document management must be reviewed.
Connection to messaging or a client tool can come later, if the need is established. To begin with, a response displayed to an employee is easier to control than an automatic action sent externally. Our article on the possibilities of AI for people and businesses places this type of project in a broader perspective.
Frequently asked questions about document assistants
Should we train a new AI?
Not necessarily. Many projects involve organizing a search of documents and framing the response. The choice depends on the need and the tools available. Before talking about training, ask how sources will be maintained, how they will be cited, and how access will be limited to the right people.
Can we import all the files at once?
This is not our recommendation for a first try. A small perimeter makes it easier to check and spot reading problems. Start with documents that you can use in your chosen setting. The expansion comes after an assessment, not simply because the interface accepts a large number of files.
Who should take care of updates?
An identified person must know which version is valid and when to remove outdated information from research. Without this accountability, an answer may be technically well-written and commercially false. Documentary monitoring is therefore part of the cost of the project, just like the tool and the control time.
Prepare a first exchange with ISS-AGENCY
To discuss a document assistant, prepare three frequently asked questions and two examples of public or fictitious documents. Explain who seeks the information, how often, and what happens when it is missing. These items are more useful than a list of desired features. They allow you to determine whether the need is for better organization, a search engine or an assistant.
ISS-AGENCY is based in Quintin and supports digital projects for businesses. You can consult our services and present your situation to us via the contact form. The objective of the first framework is a limited test, with understandable quality criteria. A quick response is only valuable if it helps the right person find reliable information.
Transform reading into action.
Present your needs to us: we will help you define a useful and verifiable first step.
Automation: organize the follow-up of your requests →
Request the auditPut it into practice
The related guide Advanced SEO and Automation is sold by DIY Marketing Guide. Check its contents and price on the product page before ordering.
