"We're interested in custom AI development or RAG (internal data AI), but we have no idea what it costs — so we haven't taken the first step." Many SME owners share this concern. Because AI development costs vary widely depending on requirements, relying on a single figure can make budget planning difficult. This article organizes the cost breakdown and drivers, how to think about pricing, our own plan rates, and how to progress from PoC to production. Use it to understand what cost items to expect and find an approach that fits your business.
What is custom generative AI development and RAG?
Off-the-shelf SaaS AI products offer fixed functionality billed monthly. Custom AI development, by contrast, designs and builds an AI-integrated system tailored to your own business workflows, data, and IT environment.
One approach attracting particular attention is RAG (Retrieval-Augmented Generation). RAG prepares your company's internal documents, manuals, past project records, and product databases as "reference material," which the AI consults when generating answers. The key advantage is accurate responses to questions that generic generative AI cannot handle — such as company-specific policies or product details. Common use cases include automating internal FAQs, supporting sales teams, and streamlining inquiry responses.
What drives the cost? (Breakdown)
Costs for custom AI / RAG development are generally built up from the following phases. The volume of each phase varies with scale and requirements, which is why overall costs vary.
| Phase | Key activities | Cost drivers |
|---|---|---|
| Requirements definition | Clarifying the problems, scope of work, and required features. Deciding which AI model to use. | Complexity of business processes, number of stakeholders |
| Data preparation & preprocessing | Converting and cleaning existing internal documents/data into a format the AI can reference effectively. | Data volume, format variety, data quality |
| Development & implementation | AI model configuration, RAG pipeline construction, API integration with existing systems, UI development. | Number of systems to integrate, UI scope, customization extent |
| Testing & tuning | Accuracy evaluation, hallucination (misinformation) checks, validation by business users. | Target accuracy level, number of revision rounds |
| Operations & maintenance | Maintaining the production environment, data updates, model version management, incident response. | Update frequency, support scope, SLA requirements |
The four main cost drivers are: ① data volume and quality, ② scope of integration with existing systems, ③ target accuracy level and number of business processes, and ④ operations and maintenance scope. Starting with a PoC (proof of concept) focused on a single specific process makes it easier to manage uncertain cost risk.
How to think about pricing
The market for AI development and RAG broadly offers two approaches: fixed-price individual development and monthly subscription.
Monthly subscription: Development, operations, improvements, and support are bundled into a monthly fee. This lowers the initial investment barrier and makes it easier to start small and iterate continuously.
Our plans at MRI Inc.
To help SMEs start small with custom AI and RAG, we offer monthly subscription plans. Starter and Standard plans come with no initial setup fee (Enterprise is custom-quoted).
| Plan | Monthly fee (excl. tax) | Typical use case |
|---|---|---|
| Starter | ¥98,000/month | Automating 1–2 specific processes, piloting RAG — starting small |
| Standard | ¥198,000/month | Expanding to multiple processes, production deployment, ongoing accuracy improvement |
| Enterprise | Custom quote | Large-scale data, multi-system integration, advanced security requirements, etc. |
Typical timeline and the PoC-to-production path
Development timelines vary by requirements, but the general approach of starting with a PoC and validating effectiveness before moving to full production is considered easier to manage in terms of both risk and cost.
| Phase | Indicative timeline | Description |
|---|---|---|
| PoC (proof of concept) | A few weeks or more (varies by scope & data readiness) | Build a prototype focused on one process; verify accuracy, effectiveness, and fit for your workflow |
| Production development | Several months or more (varies by complexity) | Informed by PoC findings: design, build, test, and release for production |
| Operations & improvement | Ongoing | Continuous data updates, accuracy improvements, and process expansion |
Starting small, confirming ROI, and then scaling to production is also effective for budget management. At the PoC stage you can decide whether to proceed, adjust the spec, or pivot to a different process — reducing the risk of a costly mismatch at full scale.
Subsidies can reduce your out-of-pocket costs
Depending on the conditions, custom AI and RAG development costs may be eligible for Japanese national or local government subsidies and support programs. Leveraging subsidies can allow you to move forward while keeping self-funded costs down. For details on subsidies available in 2026, see the related article below.
→ 2026 Subsidies for SMEs in Chiba Adopting AI | Up to ¥100M & how to apply
Tips to avoid common mistakes
① Don't start with a large-scale build
Trying to build a company-wide, multi-process system from day one tends to let requirements expand, leading to budget and timeline overruns. Instead, run a PoC on 1–2 processes where impact is easy to measure, get the numbers, and scale from there.
② Don't underestimate data preparation
RAG accuracy depends heavily on the quality of the documents you feed it. If internal docs are in mixed formats, contain outdated information, or aren't organized, data prep alone can take more effort than expected. A data inventory before the PoC significantly improves estimate accuracy.
③ Define your goals and KPIs clearly
Without concrete objectives, you can't measure effectiveness after deployment. Setting targets upfront — such as "reduce time spent on this task by X%" — gives you a clear decision framework throughout development.
④ Budget for ongoing operations
Production launch is not the finish line for AI/RAG: data updates, accuracy tuning, and model version management are ongoing. If choosing a monthly subscription, confirm in advance whether operations costs are included.
Find out in 3 minutes
Related articles:
- 2026 Subsidies for SMEs in Chiba Adopting AI | Up to ¥100M
- How SMEs Can Use ChatGPT: First Steps to Better Efficiency
- SME AI Adoption — Where to Start? 5 Steps
* The costs and timelines in this article are general indicative figures only; actual costs and timelines vary significantly based on your requirements, business context, and data situation. The plan rates listed (Starter ¥98,000/month, Standard ¥198,000/month, Enterprise by quote; no initial setup fee for Starter/Standard) are as of June 2026 (amounts shown excl. tax). We have not stated specific market prices from unverified sources. Subsidy eligibility and approval depend on requirements and review. MRI Inc. does not guarantee subsidy approval.