AI Assistant for Company Documents: Build or Buy (2026)

An AI assistant for company documents is a chat tool that answers your team's questions from your own files, such as contracts, handbooks and price lists, and shows which file each answer came from. If your files live in one office suite, try its built-in assistant first. Build one when the answers must be checkable figure by figure.

We build these, so read this knowing we have a side. Right now we are building a research assistant for Niveshaay, a SEBI-registered research and fund-management firm, and it is not live yet. What follows is how we decide whether a company needs a build at all, and what we write down before we start one.

How does an AI assistant answer from your own documents?

A language model on its own knows what it read in training. It has never seen your returns policy or last quarter's supplier contract. So the assistant does a search first. It finds the passages in your files that match the question, hands those passages to the model with the question, and the model writes an answer from them.

Researchers named this retrieval-augmented generation in a 2020 paper, and most products in this space still work this way. The part that matters to you is the search step. If it pulls the wrong passage, the model writes a confident answer from the wrong page. If it pulls nothing useful, a well built assistant says it found nothing. A badly built one fills the gap from general knowledge, and that answer reads exactly like a correct one.

That is why we treat the citation as the core feature. An answer that names its source document can be checked in seconds. An answer without one has to be trusted or redone by hand, and once a team catches one wrong figure, most of them stop asking.

Can't we just give the team ChatGPT or Copilot?

Often, yes, and it is the cheapest first step. Pasting a document into a general chatbot works for one-off questions about one file. It fails once the answer is spread across forty files, or once the files change every week and nobody remembers to paste the new version.

The assistant inside your office suite is the next step up. Microsoft 365 Copilot reads the documents, emails and chats a person already has permission to open. Microsoft's documentation says it "only surfaces organizational data to which individual users have at least view permissions", and that prompts and responses are not used to train the underlying models. If your company runs on SharePoint and Teams, that covers a lot of everyday lookups with no build at all.

The same documentation also carries a warning: generated responses "aren't guaranteed to be 100% factual", and users should review them before sending. That is fair for drafting an email. It is a harder sell when the answer is a revenue figure going into an investment note, or a warranty term going to a customer.

There is a second catch. Suite assistants follow your existing permissions, so a messy shared drive becomes a messy assistant. If anyone can open the salary spreadsheet today, the assistant will quote from it to anyone who asks.

What are the options, side by side?

Plenty of companies belong in the middle column. The build only makes sense when one of the rows on the right is a requirement and the suite cannot meet it.

When is a custom document assistant worth building?

These are the four signs we look for on a first call. One of them alone may still be solvable with a product. Two together usually mean a build.

• Every figure has to be traceable. In research, finance or compliance work, "the AI said so" is not an answer anyone can sign off on.
• The answer needs more than documents. Niveshaay's assistant answers from the firm's own files and a market data feed, which no office suite assistant joins for you.
• Some things must never happen. Rules like "the AI never acts alone" or "nothing is deleted" have to live in the system, not in a training session.
• The files are not in one place. PDFs on a shared drive, Word files in email threads and a spreadsheet on someone's laptop do not index themselves.

A note on the third point. Telling a model in its instructions to never delete anything is a request. Building a system where the delete action does not exist is a rule. Niveshaay's assistant is being built the second way. The same thinking runs through kriyo, our outbound platform: by default the AI drafts each reply and a person approves it, and the operator console keeps an append-only audit log.

What goes into the scope before anything is built?

We write a scope and the client agrees to it before any code. Whoever builds yours, make sure it answers these five questions. Most of the hard thinking happens here.

• Which questions must it answer? Collect twenty real ones the team asked last month, with the answers a person gave.
• Which files and feeds does it read, and which does it never touch?
• Who can sign in, and does everyone see everything?
• What must it never do? Send, delete, change a record, answer outside the files.
• What does it say when it does not know?

Those twenty questions become the test set. Run them against the assistant before handover and again after every change, and check each answer against what the person said. A question with no answer in the files belongs in the set too. The correct response to it is "I could not find this", and that deserves as much testing as the rest.

The last scope question matters more than it looks. In kriyo's reply drafting, a value the system does not have shows as unknown, never as a made-up zero. A document assistant needs the same habit. A blank is honest. A plausible number with no source is how a wrong figure ends up in a board deck.

How do you keep a document assistant reliable after launch?

Most of the work after launch is boring, and that is the point. The assistant has to run every day like any other system the team depends on.

For Niveshaay that means separate environments for development, demo, test and production, so a change is tried on test before anyone at the firm sees it. It means a nightly backup. It means the assistant reads the PDF and Word files the team already has, so nobody has to convert anything before uploading it.

Then someone has to own the files. An assistant reading last year's price list will quote last year's prices, with a citation, perfectly. Decide on day one who removes a superseded document and who adds the new one. In a small company that is usually the same person who updates the shared drive today.

Finally, keep asking the test questions. Model providers update their models, and the same question can come back worded differently or, rarely, wrong. Re-running the test set after a model change tells you whether anything moved.

How much does a custom AI document assistant cost?

It depends on how many file sources it reads, whether it joins a live data feed, how many people sign in with different access, and how strict the citation rules are. An assistant over one folder of policies for one team is a small job. One that cites every figure across years of research and a market feed is a bigger one.

We do not publish a figure, because a number without a scope is a guess. We agree the price after a free 30 minute call, once we know what it has to answer. If the assistant in your office suite will do the job, we will say so on that call.

Try it in an afternoon first

Before talking to anyone, write down twenty questions your team asked last month and the files that held the answers. Ask them of the assistant you already pay for, or of a general chatbot with the files attached. Mark each answer right, wrong, or right but uncheckable.

If nearly all are right and checkable, you are done, and it cost you an afternoon. If the wrong and uncheckable columns fill up, you now hold the first draft of a scope. There is a working demo on our AI consultancy page, built on three invented documents, if you want to see how a cited answer looks before you start.

Sources

• Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (permissions, training, accuracy statements), as of November 2026: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020): https://arxiv.org/abs/2005.11401
• Niveshaay research assistant, in build: https://www.mkdirhq.com/work/niveshaay-research-assistant
• AI consultancy service page and demo: https://www.mkdirhq.com/ai-consultancy