AI guide · Updated August 2026
These systems are genuinely useful and they are also confidently wrong in four specific, predictable situations. Most vendors demo the parts that work. Here are both halves.
What it costs
The pilot row is the one to take seriously. Two to four weeks and a couple of lakh tells you whether the production build is worth commissioning at all.
| Route | Cost | What you get |
|---|---|---|
| Off-the-shelf tool | ₹2,000 – ₹15,000 / month | Upload documents, ask questions. Fine for a small, tidy document set and a team that can live with generic answers. |
| Pilot / proof of concept | ₹75,000 – ₹2,50,000 | Two to four weeks on your real documents with an agreed pass mark. The cheapest way to find out if this works for you. |
| Production build | ₹4,00,000 – ₹20,00,000 | Permissions, evaluation, citations, integration with SharePoint, Drive or your intranet. Three to five months. |
| Running cost | ₹8,000 – ₹80,000 / month | Model API usage, vector storage and hosting. Scales with query volume and how much context each answer needs. |
Running cost is the figure most proposals leave out. It scales with query volume and with how much document context each answer needs — a system answering 500 questions a day costs meaningfully more to operate than one answering 50.
Where it fails
These aren't bugs to be fixed in a later version — they're consequences of how retrieval works. Ask any vendor how they handle each one specifically.
Counting and aggregation across many documents. Retrieval fetches a handful of relevant chunks, so the model answers from those rather than the full set. Ask how many contracts expire this quarter and you'll get a confident, wrong number.
Policy v3 supersedes v1, but both are in the index and both look authoritative. Without version handling the system will happily quote the outdated one. This is the failure that causes real damage in HR and compliance use.
Scanned PDFs need OCR before they're readable at all. Spreadsheets lose their meaning when flattened into text — a table of rates becomes a wall of numbers with no structure. Both need handling as explicit work, not assumptions.
"Does our leave policy conflict with what we promised in the Mumbai contract?" requires holding two documents together and comparing them. Standard retrieval fetches passages; it doesn't reason across sources. Possible to build, but it's a different and more expensive system.
Where it works
All four share a shape: the answer already exists in one place, and the cost is a person spending time finding it.
Policy documents, SOPs, contracts and manuals where the answer exists in one place and someone has to find it. This is what retrieval is genuinely good at, and it's most of the value.
Every organisation has two or three people who get interrupted constantly because they know where things are. That interruption load is the measurable thing a document assistant removes.
New joiners ask a high volume of questions with answers already written down somewhere. High query volume, low risk if an answer is imperfect — close to an ideal first use case.
Agents finding the right article faster while a human still decides what to tell the customer. The human review step is what makes this safe to deploy early.
How to test it
Run this with two vendors at once if you can. Same questions, same documents, same scoring — the difference in what comes back is usually stark.
Collect them from the people who'd actually use it, in their own words. Include the awkward ones. Questions written after watching a demo are unconsciously shaped to suit it.
Counting questions, superseded policies, anything in a scanned document. You're not trying to make it fail — you're finding out whether the vendor tells you it failed.
The scanned ones, the near-duplicate versions, the spreadsheet everyone relies on. Clean data proves nothing about how this performs on Monday morning.
Which document, which section. Without citations nobody can verify an answer, and an unverifiable answer in a policy context is worse than no answer at all.
Correct, partially correct, or wrong — decided by someone who knows the right answers. Agree the pass mark in writing beforehand, because deciding what counts as success afterwards always ends in an argument.
Don't build if
The first one accounts for most failed projects we've seen. No retrieval system rescues a document store nobody has organised.
Questions
Not covered here? Ask us directly — we answer scoping questions without putting you into a sales sequence.
Off-the-shelf tools cost ₹2,000 to ₹15,000 a month. A proof of concept on your real documents costs ₹75,000 to ₹2,50,000 over two to four weeks. A production build with permissions, evaluation, citations and integration into SharePoint or Drive costs ₹4,00,000 to ₹20,00,000 across three to five months. Running costs add ₹8,000 to ₹80,000 monthly depending on query volume.
RAG stands for retrieval-augmented generation. Your documents are split into passages and indexed. When someone asks a question, the system retrieves the passages most likely to contain the answer and passes them to a language model, which writes a reply grounded in that text. This is why RAG can cite sources — and why it struggles with questions requiring the whole document set rather than a few passages.
Four situations reliably. Counting or aggregation questions, because retrieval fetches a few passages rather than everything. Questions where an older document was superseded but both remain indexed. Content locked in scanned images or spreadsheets, which need OCR and structured handling. And questions requiring reasoning across two documents at once. Ask any vendor how they handle each of these specifically.
It can be, with three things in place. Document-level permissions so the assistant only surfaces what that user could already open. A named model provider with data retention terms confirming your content isn't used for training. And an audit log of queries and answers. Ask for all three in writing — vague reassurance that it's 'completely secure' is not an answer.
Below roughly 50 documents, a good folder structure and search will usually serve you better. The value grows with volume and with how scattered the information is — a few hundred policies, SOPs, contracts and manuals across several systems is where a document assistant genuinely earns its cost. The other requirement is that the documents are reasonably organised to begin with.
A proof of concept on your real documents takes two to four weeks. A production system takes three to five months, and most of that time goes into permissions, evaluation and handling edge cases rather than the retrieval itself. Anyone promising a production-grade internal assistant in two weeks is showing you a demo and skipping the evaluation work that keeps it trustworthy.
Not a requirements document — the actual questions your team asks each other. We'll tell you which ones retrieval handles well, which ones it won't, and whether a pilot is worth running at all.
Scope a pilot
