A custom AI document assistant that answers questions from your own content has two price tags. The build usually starts around $15,000 to $40,000 for a first customer-facing version, according to a 2026 guide from one custom-AI vendor. Running it is a separate monthly bill, from roughly $200 in AWS’s simplest published example to several thousand dollars at high query volume. Larger, permissioned production systems are quoted by another vendor at $80,000 to $500,000 or more. Those figures are vendor estimates for specific scopes, not market averages, so the most useful question is what each number actually includes.
Separate the build budget from the running budget
Most proposals mix two kinds of cost. One-time implementation covers discovery, document ingestion, retrieval design, the interface, testing and launch. Recurring operation covers model usage, vector storage or search indexes, hosting, networking, monitoring and maintenance. A quote that shows only one of these lines is incomplete.
As an Amazon Associate I earn from qualifying purchases.
One-time build costs by scope
Vendor pricing guides cluster around a few levels of scope. The table below lists the ranges as each vendor published them. Neither guide defines a market-wide average, and the scopes do not map perfectly onto each other.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches| Scope | Published estimate | Timeline or conditions | Publisher |
|---|---|---|---|
| MVP chatbot grounded in your own content | $15,000–$40,000 | 3–6 weeks delivery range | 4xxi 2026 custom-AI guide |
| Controlled pilot (enterprise RAG) | $35,000–$75,000 | Not stated | NextPage enterprise RAG cost guide |
| Production knowledge assistant | $80,000–$180,000 | Not stated | NextPage enterprise RAG cost guide |
| Regulated or operationally managed deployment | $180,000–$500,000+ | Assumes regulated data, document-level permissions, source synchronization, evaluation datasets, audit logs and managed operations | NextPage enterprise RAG cost guide |
The gap between the MVP and the production tiers is wide because “chat over documents” describes very different systems. A single document collection behind a simple web interface is a small project. A system that pulls from several live systems, respects each user’s access rights, logs every answer and is tested against a fixed evaluation set is a different engineering effort.
#1 Best Overall
- 🎙️ Hands-Free Voice Typing for Windows & Mac – Powered by iOS & Android dictation technology, AI VoiceWriter allows fast, accurate speech-to-text directly on your desktop. Simply speak, and your words appear in real time. Compatible with Windows 10 & above, macOS 13 & above.
- ✍️ AI Writing Assistant for Effortless Editing – Boost productivity with AI proofreading, rephrasing, and formatting. Perfect for emails, reports, creative writing, and professional content.
- 💻 Works Seamlessly in Any Desktop App – Type with your voice in Microsoft Word, Google Docs, PowerPoint, Teams, emails, and more. Just place your cursor in any text field and start speaking!
- 📱 Mobile App for Enhanced Voice Input – The AI VoiceWriter mobile app enhances voice recognition by using your phone’s microphone as an input device for clearer, more accurate dictation—while typing on your desktop. Supports iOS 15 & above, Android 9.0 & above.
- 🌎 Multilingual Voice Typing & AI Assistance – Supports 33 languages for dictation, plus AI-powered features in Chinese, English, Japanese, Korean, French, German, Spanish, Italian and, Swedish.
Add-ons that the 4xxi guide prices separately
- Scanned-document processing: $10,000–$30,000, to handle OCR and poor-quality scans.
- Multilingual processing: $5,000–$15,000 per language.
These add-on figures come from one vendor and should be treated as prompts for scoping a proposal, not as standard rates.
Recurring cloud and model costs
AWS publishes worked examples in its implementation guide and cost pages. Its own guide states that “the cost of the use case will vary depending on the configuration, such as Text use cases with different model providers, with or without Retrieval Augmented Generation (RAG) enabled, and so on.” The examples below show how far that configuration choice moves the bill. They are AWS scenario calculations, not prices for custom development, and cloud rates change over time, so check current pricing before budgeting.
Four AWS scenarios
| Scenario (AWS) | Estimate | Assumptions |
|---|---|---|
| Simple production-ready chatbot on Amazon Bedrock, no document access | About $200/month | US East (N. Virginia) |
| Sample agent proof of concept with Bedrock Knowledge Bases and Guardrails | About $840/month | About 100 daily interactions |
| VPC-enabled use case over tens of thousands of documents | About $1,500/month | About 8,000 queries per day; includes a Kendra index and other solution components |
| QnABot example with embeddings and model inference | $775.33–$2,755.33 (example total) | 8,000 daily questions, 2,000 input tokens per request |
The QnABot example rises to $1,508.33–$5,468.33 when its Bedrock knowledge-base RAG option is modeled with the same workload. The period for that total is not restated on the page, so confirm it in AWS’s calculator before comparing it with the monthly figures above.
Where the money goes in a RAG system
A separate AWS RAG cost breakdown, modeling 8,000 interactions per day, shows that retrieval infrastructure can cost more than the model itself:
| Component | Monthly estimate | Notes |
|---|---|---|
| Use-case application components | $577.76 | Before knowledge-base costs |
| Embedding calls | $9 | Same workload |
| Basic serverless OpenSearch vector store | $691.20 | AWS labels this estimate rough; heavier workloads may need more capacity, and existing provisioned resources can lower the cost |
| Kendra search configuration | $1,008 | Under the query and document assumptions stated in the breakdown |
The lesson is that the vector store or search index is often the largest line in a document assistant’s bill, not the per-token model charge. Choosing between a managed index and a provisioned one is a budgeting decision, not just an architecture one.
Azure OpenAI pricing models
Microsoft describes three ways to pay for Azure OpenAI models. On-demand pricing charges per input and output token. Provisioned throughput uses monthly or annual reservations. Batch processing is advertised at a 50% discount on Global Standard pricing for eligible batch workloads, which suits jobs such as bulk document processing rather than live chat. Microsoft states that displayed prices are estimates that vary by agreement, purchase date and currency, and that deployment choice (global, data zone or regional) also affects cost. Calculate with your own region and contract.
Why a low monthly figure can mislead
The $200 AWS example has no document access, so it describes a general chatbot, not a document assistant. Treat it as a floor for hosting a bot, not as the running cost of a system that answers from your files. Compare recurring figures only when they share the same workload, retrieval setup and region.
Free tools Windows power users keep installed
One-click scans. No signup required.
What pushes a quote up
Proposals that look similar on the surface often differ on these seven dimensions:
- Documents and ingestion: number, formats, size, scan quality, update frequency, and parsing or OCR needs.
- Retrieval workload: document count, expected questions per day, context size, embedding and vector-store design, and search quality targets.
- Integrations: number of source systems and whether content must sync continuously rather than on a schedule.
- Access control and risk: identity integration, document-level permissions, data boundaries, audit logs, retention and security controls.
- Quality assurance: evaluation datasets, citation and grounding checks, human review, error handling and acceptance criteria.
- Operations: uptime, latency, peak traffic, monitoring, support hours, model changes, and who maintains the system after launch.
- Geography and purchasing: cloud region, data residency, provider, model, pricing agreement, and reserved versus on-demand capacity.
The NextPage guide names permissions, synchronization, evaluation datasets, audit logs and managed operations as the main drivers of its enterprise tiers. AWS’s scenarios show that traffic, token counts, retrieval-store configuration and network architecture drive the recurring side.
How to get quotes you can compare
- Ask for a defined first phase with a fixed scope, such as one document collection, one interface and named acceptance tests.
- Require a workload model with daily questions, average input, context and output tokens, corpus size and refresh frequency, the selected model and region, vector-store minimums, and network and security configuration.
- Ask which items are in the implementation price and which are usage-based.
- Ask who owns maintenance, support hours and model updates after launch.
- Send every vendor the same requirements document, then compare permissions, integrations, acceptance tests and support scope line by line.
A managed or off-the-shelf platform can cut build effort, while a custom system is easier to justify when permissions, workflows or integrations are strict. Neither source publishes a break-even point between the two, so that decision has to be made against your own requirements.
Reader questions behind the search
AWS’s own blog, published August 11, 2025 by Brian Clark, Srividhya Pallay and Prerna Mishra, frames a common customer question as: “How much will it cost to run our chatbot on Amazon Bedrock?” The question behind most searches is broader: what does a custom assistant that answers from my documents cost to build and to keep running? The answer above is that the build and the run are separate budgets, and both depend on scope rather than on the word “chatbot”.
Neither vendor source offers a market-wide average for this kind of project. Use the ranges above as bounds for vendor conversations, and attribute each figure to its publisher and scope when you share it internally.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
