Find answers in company documents without chasing colleagues

Context

You need to reply to a client, and you know the answer exists. It's in the updated procedure, a supplier's email or a similar contract. You just need the right document. You ask a colleague, who points you towards whoever handled the matter. Your reply is on hold, and the search is now interrupting other people's work too.

In a service business or a professional firm, this can become the default way to find information. Documents accumulate, while the knowledge needed to choose between them stays with the people who know them. Tidy folders help, but you still need to know which version applies and under what circumstances.

This is a representative scenario for the sector. We describe how we approach the problem, without presenting it as a real client's results.

The challenge

1. Ask the question in the terms you use at work. "Which procedure applies to this case?" Answering means finding the relevant text, checking that it's current and reading the exceptions. An assistant can help connect your question to the available content. Asking which invoices remain unpaid is a different task: that requires access to current accounting data. Connecting a document archive is not enough.

2. Establish which answer you can rely on. A procedure in a shared folder may differ from the copy attached to an old email. If both appear valid, an answer with a citation can still send you in the wrong direction. Before connecting the content, we need to identify the current versions and who is responsible for keeping them up to date.

3. Make repeated explanations available to everyone who needs them. New joiners have to learn where to look and whom to ask. Experienced colleagues repeat explanations they have already given. An assistant can retrieve what has been documented. Knowledge that exists only in conversation first needs to be captured, written down and checked.

4. Address the reasons for scepticism. Privacy and difficulty controlling or understanding AI responses also feature in a research summary published on 24 March 2025. These are practical questions for the project. You need to know what data is processed, who can access it and how an answer is checked before you use it.

The numbers to watch

The before-and-after comparison on this page describes the intended changes. The ticks do not represent measured results. To judge whether an assistant would help your business, we start with questions you already receive and measure the work needed to answer them today.

• How long does it take to reach a usable answer, including verification?
• How many requests are resolved without involving another colleague?
• What share of answers has an accessible source that actually supports the response?
• How often does a search stall because the person who knows the case is unavailable?
• Among the answers reviewed, how many are wrong, incomplete or based on outdated versions?
• How much work goes into preparing a proposal using previous examples?
• Where do new joiners keep getting stuck?

We repeat the comparison with similar requests. We also count the time spent correcting documents, managing permissions and reviewing answers. An answer that arrives quickly may take a long time to verify. That time belongs in the calculation.

Our approach

1. Choose a clearly defined scope. We agree which questions to address and which content to connect, starting with a limited collection. For each set of documents, we identify who is responsible for it and how revisions are managed. Operating procedures and recurring client queries can be a useful starting point. Contracts and personnel files need a specific assessment before they are included.

2. Make the link between answer and evidence checkable. We provide references to the source document and relevant passage so you can open them and check. During testing, we assess whether the document supports the assistant's claims, including any conditions and exceptions. A genuine citation can accompany a false conclusion. We check the meaning, as well as whether the link is there.

3. Apply access permissions to every request. We review permissions in the source systems and use them to select the content available to the person asking. We also check what happens when access changes or a document is withdrawn. Restricted content must stay out of summaries too. Hiding the link while leaving the text readable does not protect it.

4. Set aside time for review and assign responsibility. Every week, we review a sample of answers with people who know the subject. We distinguish between an error in the document, retrieval of the wrong content and a mistaken interpretation by the model, then test the correction on fresh requests. The model can get a clear, current document wrong. Review is part of running the service.

We also agree how to involve a person when information is missing or contradictory. We instruct the assistant to acknowledge these limits and test how it behaves. The instruction does not guarantee that it will recognise every gap.

What changes

You can start with the question you need to resolve. Which procedure applies here? What terms were agreed with this supplier? Is there an approved reply to a similar query? Where contracts are within the authorised scope, you can also look for previous examples of a clause.

In the proposed workflow, the assistant suggests an answer and directs you to the content you need to check. If an exception comes up, you have a reference to discuss with the colleague responsible. Finding a precedent does not establish that it applies to the current case.

The aim is to reduce interruptions for routine searches and leave people to make judgements that need experience. We have to measure whether that happens. If you need to repeat the entire search before trusting the answer, the benefit has yet to be demonstrated.

What people fear, and what we tell them

"Our documents leave the firm." Yes. To generate a response, we send questions and the necessary passages to the service running the model. The models run on AWS Bedrock with EU inference profiles; data is processed in the European Union. This involves processing outside your own systems. Before connecting anything, we define which content to include, retention periods and logging. Processing data in the EU does not settle every question about confidentiality. We need to agree explicitly what data will be processed.

"It answers confidently even when it's wrong." Yes. The model can produce a plausible but false answer, miss an exception or draw a conclusion the document does not support. It can also answer when it ought to stop. Citations, instructions and review reduce the risk without removing it. Sampling leaves the possibility of undetected errors. If you use an answer to make a commitment to a client, you need to check it against the document and involve whoever is responsible for the decision. Convincing wording is not confirmation.

"Someone will read documents they shouldn't see." Applying existing permissions helps only if those permissions are correct. Before connecting folders, we review access and test requests from different roles. If access is too broad, we need to correct it at source. We also check that permission changes take effect in the assistant. Where we cannot apply and verify access rights, we leave that content out.

How we approach it today

In the scoping workshop, we start with the requests that take up your time and the documents that should answer them. We assess content quality, permissions, responsibilities and the work needed to keep everything current. We define what to test and what evidence would justify proceeding.

If there is no reliable version of a procedure, we say so. If improving your existing search would do the job, we say so. The project has to justify the time it asks of your people.

• AI integration in the systems you already use
• The AIRA method
• Odoo and integrated ERP

Let's talk. Bring us a question that currently means interrupting a colleague. We'll check together whether your documents already contain what you need to answer it.