How to Prepare Company Documents for an AI Assistant in 2026: A Client Checklist

📊 Section 3. Format Comparison — What AI Handles Well and What It Doesn't

One of the biggest pitfalls in RAG projects is document format. If you think "a PDF is a PDF," here's the unpleasant truth: there are two fundamentally different types of PDFs, and one of them is like a blank page to AI.

A digital PDF is a document created on a computer (e.g., saved from Word). The text in it can be highlighted, copied, and AI recognizes it perfectly. A scanned PDF is essentially a photograph of a paper document. To the human eye, it looks the same, but to AI, it's just an image from which text cannot be extracted without special processing (OCR).

As researchers note, parsers without OCR capabilities return a completely empty result for scanned documents — and critically, they don't report it (The PDF Problem: Why AI Struggles to Read Business Documents). AI won't say "I can't read this" — it will simply ignore the document and respond without it.

How to Check if Your PDF is Digital or Scanned?

Open the file and try to highlight the text with your cursor. If the text highlights, it's a digital PDF, and AI will process it. If only an area highlights (like in a picture) or nothing highlights, it's a scan. It needs to be processed through OCR before uploading to a RAG system.

Summary: before handing over documents to a contractor, check: are there scanned PDFs among them? Are your price lists up-to-date? Is the text in tables readable? 30 minutes of checking now will save days of work later.

💰 Section 4. How Much Time and Money Does Data Preparation Take

The timeline and cost depend on one simple factor: how organized your documents already are. Here are three typical scenarios we see in 90% of projects:

Scenario 1: "We're Organized" (2–3 days, minimal cost)

Documents are gathered in one place (Google Drive, Notion, SharePoint). The price list is current. FAQs exist. Instructions are written down, not just "in Olena's head." Files are in digital formats (Word, Google Docs, digital PDF). In this case, preparation involves checking for relevance, removing duplicates, and forming the final list. You can do this yourself using our checklist in 2–3 days.

Scenario 2: "Moderate Chaos" (1–2 weeks, $200 – 500)

Documents exist but are scattered across different systems. Some are in email, some in Telegram groups, some on local computers. Some files are outdated, and there are duplicates. There are scanned PDFs that need to be processed through OCR. In this scenario, you'll need help: either allocate 1–2 team members for a week or assign it to a contractor.

Scenario 3: "Complete Chaos" (3–4 weeks, $100 – 3,000)

There is almost no documentation. Knowledge "lives" in the heads of specific people. The price list hasn't been updated in a long time. There are no instructions — only "Ask Serhiy, he knows." In this case, the first step is not AI implementation, but creating a knowledge base from scratch: interviews with key employees, documenting processes, writing documents. This is an investment that pays off even without AI — simply because you will finally have documented processes.

Prices in Ukraine vs. Europe vs. USA

Data audit and preparation in Ukraine — $200–1,000 for a medium-sized project. In Western Europe, similar work costs €1,000–2,000, in the USA — $3,000–5,000. At the same time, the result is the same: a structured, cleaned, ready-to-upload knowledge base for RAG. For foreign clients, this is another argument in favor of working with Ukrainian teams.

Summary: data preparation is 20–30% of the AI project budget. This is normal. These are not "extra costs" — they are the foundation without which nothing else works.

⚠️ Section 5. 5 Data Preparation Mistakes That Kill AI Projects

Mistake 1: Outdated Documents in the Knowledge Base

You upload the 2024 price list — and the AI assistant quotes old prices to clients. Or the knowledge base contains an old version of an instruction that contradicts the new one. AI doesn't know which version is current — and might choose any. Solution: before uploading, check each document for relevance. Mark the date of the last update. Delete old versions.

Mistake 2: Scanned PDFs Without Text Recognition (OCR)

This is the "silent killer" of RAG projects. The document looks normal when you open it. But for AI, it's just an image — it can't read a single word. And it won't say "I can't see this document" — it will simply respond without it. You might not know for months that part of your knowledge base is "invisible" to the assistant. Solution: check all PDFs (highlight text with your cursor). Process scanned ones through OCR.

Mistake 3: Duplicates and Contradictory Documents

Three versions of the same FAQ. Two price lists — one for the website, another for managers. An instruction that was updated, but the old one wasn't deleted. AI will find both documents and won't be able to determine which is correct — the result will be random. Solution: one document — one version. Perform "deduplication" before uploading.

Mistake 4: Confidential Data Without Filtering

You upload documents with clients' personal data, internal financial reports, or confidential partner agreements to the knowledge base — and the AI assistant provides them in responses to external clients. This is not just a loss of trust — it's a legal risk. Solution: separate documents into "public" and "internal." RAG specialists recommend maintaining separate databases for public and sensitive corporate documents (Kapa.ai — RAG Best Practices).

Mistake 5: Lack of Structure in Documents

A document without headings, subheadings, and logical breaks is a continuous stream of text, making it difficult for AI to find specific fragments. Imagine a book without a table of contents or chapters — you wouldn't find the information you need either. Experts in building knowledge bases for RAG recommend breaking down documents into semantic blocks, not by character count, and maintaining a consistent format for headings, lists, and indents (Astera — Building a Knowledge Base for RAG). Solution: add headings to each section. Break large documents into logical blocks. Use a consistent format.

Summary: each of these five mistakes is not just an "imperfection." They are specific risks: incorrect prices, data leaks, loss of clients. 2–3 days for checking — and you will avoid them all.

💼 Section 6. What a Contractor Should Do During the Data Audit Stage

The data audit stage is a "litmus test" of a contractor's professionalism. If the contractor says, "Just dump everything into Google Drive, we'll figure it out" — that's a red flag. Here's a checklist of what a proper contractor should do before starting development:

1. Conduct a document inventory. Create a complete list: what exists, in what format, where it's stored, when it was last updated. This sounds tedious, but it's where 90% of problems are revealed.

2. Check formats and readability. Are all PDFs digital? Are there scanned ones? Are Excel tables readable? Is the formatting in Word "broken"? The contractor should check this and provide a report: "these 15 files are ready, these 8 need conversion, these 3 are unreadable."

3. Determine priority documents for MVP. Not everything needs to be uploaded at once. The contractor should help select 10–30 key documents for the first launch, and plan the rest for subsequent iterations.

4. Check for relevance and remove duplicates. If there are three versions of a price list — the contractor should ask: "Which one is current?" If an instruction contradicts an FAQ — help to reconcile them.

5. Filter out confidential information. If the assistant will be public (for clients) — the knowledge base should not contain internal financial reports, personal data, or confidential partner agreements. The contractor should assist with this separation.

6. Provide you with a report and preparation plan. The result of the audit is a concrete document: what is ready, what needs to be refined, how long it will take, who is responsible for each item.

Summary: if a contractor starts development without a data audit — it's like a builder laying walls without a foundation. The result will be commensurate.

🏆 Section 7. How We Prepare Data in WebCraft

Our approach is built on a simple principle: it's better to spend 3–5 days on an audit than 3–5 weeks later on redoing an assistant that answers incorrectly due to poor data.

What our audit includes (free of charge)

• ✔️ Inventory: we compile a list of all documents you have — and those you don't, but need
• ✔️ Format Check: we find scanned PDFs, "broken" tables, unreadable files
• ✔️ Relevance Assessment: we mark outdated documents, duplicates, contradictions
• ✔️ Confidential Data Filtering: we help separate documents into "for clients" and "internal use only"
• ✔️ Preparation Plan: specific document — what to do, who does it, how much time

Case Study

A home appliance distribution company — 400+ documents: manufacturer catalogs, price lists, warranty terms, user manuals. During the initial audit, we discovered: 40% of documents were outdated (2022–2023 catalogs for models no longer sold), 15% were scanned PDFs without OCR (partner warranty cards), and 10% were duplicates. We removed outdated information, processed scans through OCR, and structured the rest by categories (home appliances → brand → type → model). After cleaning, 120 documents remained — and based on this data, the AI assistant answers 92% of customer inquiries accurately. The entire preparation process took 8 business days.

Summary: we do not start development until the data is ready. This is our principled stance — because we are responsible for the quality of the result.

❓ Frequently Asked Questions

What if all my information is "in my head" and not in documents?

This is the most common situation in small businesses. The solution is to start with interviews with key employees. We ask 20–30 questions, record the answers, and structure them into documents — and these documents then become the knowledge base for the AI. Bonus: after this process, you will have a documented knowledge base that is useful even without AI — for training new employees, for example.

Can I use a scanned price list?

Yes, but it needs to be processed through OCR first — a program that recognizes text in images. After OCR, the text quality won't be perfect (there might be errors in numbers or symbols), so manual verification is required. The best option is to find the original digital file or create a new price list in Word/Google Docs.

What is the minimum number of documents required to launch?

For an MVP, 10–30 key documents are sufficient: an up-to-date price list, service descriptions, FAQs, terms of service. This will cover 70–80% of typical customer inquiries. After launch, you'll see which questions the assistant can't answer — and you can add the necessary documents on a case-by-case basis.

Who is responsible for preparation — us or the contractor?

Usually, it's a joint effort. The contractor conducts the audit, provides a checklist, and recommendations. You provide access to documents, confirm their relevance, and answer questions about business processes. Some contractors (including WebCraft) can take full responsibility for preparation — for an additional fee.

What if some of the data is confidential?

There are two approaches. The first is to divide the knowledge base into "public" (for the client-facing AI assistant) and "internal" (for employees). The second is to use a private deployment where data is stored on your server and not shared with third parties. More details on security can be found in our article on data security in AI implementation.

Do I need to rewrite anything?

Not necessarily. If a document is understandable to a human — it's understandable to AI. The main thing is relevance, absence of duplicates, and a readable format. Rewriting is only necessary if a document contradicts others or contains critically outdated information.

How often do I need to update the knowledge base?

It depends on the business. If price lists change monthly — the knowledge base should be updated monthly. If the documentation is stable — once a quarter is sufficient. A good RAG system allows updating individual documents without rebuilding everything from scratch: upload a new price list — and the assistant immediately knows it.

✅ Conclusions

• 💰 Cost: data preparation — from free (DIY with a checklist) to $200–1,000 (turnkey contractor). This is 20–30% of the AI project budget
• 🎯 Main Recommendation: start with 10–30 key documents from Category 1. Don't try to upload everything at once
• ⚠️ Main Warning: scanned PDFs, duplicates, and outdated documents are the three "silent killers" of AI projects. Check them before starting development

Key takeaway: An AI assistant is only as good as your data. Invest in document preparation — and the technology will work for you, not against you.

🚀 Don't know where to start? We can help

Submit a request for a free audit of your data — we will analyze your documents, show you what's ready and what needs improvement, and create a preparation plan for implementing an AI assistant.

Order a free audit → WebCraft

Or write to us on Telegram — we'll respond within 3 hours.

📖 Read also

• 🤖 What is RAG and why your business needs it: an explanation without technical jargon
• 💼 AI assistant for small business: is it worth it with a budget up to $1000
• ⚖️ AI chatbot vs RAG assistant: what's the difference and what to choose
• 🔒 Data security when implementing AI in 2026