Reduce the Cost of AI on IBM i: Save Your Tokens!

Reducing the cost of AI on IBM i has become a concrete challenge for teams adopting artificial intelligence. The cost of tokens, which you pay for with every use, can quickly rise when the AI lacks context about your application. Anyone can open ChatGPT, Claude, or Copilot, but the real challenge is not gaining access to AI. It is ensuring that the AI truly understands your application without requiring multiple costly prompts.

AI hallucinations are widely discussed. Its cost is discussed far less. Yet this cost (the cost of tokens) is directly linked to the same issue: a lack of context. In an IBM i environment, you pay for this twice, both in reliability and in euros.

Without sufficient context about your application, the AI continually asks for more information: source code, documentation, and explanations. You rewrite the prompt once, twice, or three times, hoping to obtain a reliable answer. With every reformulation, you consume more tokens.

In other words, the less the AI knows about your environment, the more it has to guess. And the more it has to guess, the more expensive it becomes.

On a critical system such as IBM i, an AI without context can provide a highly convincing answer, sometimes even more convincing than an expert. However, convincing does not necessarily mean correct. The AI may overlook a dependency, ignore a business rule, or underestimate the impact of a change.

The result is a series of back-and-forth prompts to correct the response, followed by additional requests and reformulations. More tokens are therefore consumed, while the reliability of the final answer remains uncertain.

To be genuinely useful in a critical environment, AI needs three things:

This is precisely where the ARCAD Repository comes in by introducing determinism. It can be viewed as the application memory of your IBM i environment. Instead of asking AI to guess how your application works, you give it access to the facts: source code, objects, dependencies, procedures, relationships; in short, everything that reflects the reality of your application landscape.

This moves the AI from a logic based on assumptions to one based on interpretation and knowledge. This factual foundation helps reduce both errors and unnecessary token consumption.

The key principle is simple: AI should query, not guess. The ARCAD Repository becomes the source of truth.

To connect AI to the Repository, ARCAD relies on the open-source Model Context Protocol, or MCP. This protocol enables different types of AI (including conversational agents, coding assistants such as Claude Code or IBM Bob, and AI embedded in other tools) to communicate with external systems.

ARCAD MCP Server acts as a translator. A developer’s natural-language request, such as “perform an impact analysis on this file,” is converted into structured calls to the ARCAD Repository. The server currently provides more than 70 tools dedicated to the Repository and 200 tools across the entire ARCAD ecosystem, covering DevOps, data anonymization, and deployment with DROPS. Organizations can also create their own tools, sometimes with as little as twenty lines of code, to connect additional solutions.

In practical terms, developers can remain in their code editor (Claude Code, for example) from the beginning to the end of a development cycle. They can perform an impact analysis through the Repository, edit the code using AI, validate it through compilation with ARCAD Builder, and then deploy it with DROPS, without changing interfaces or multiplying exploratory prompts.

The second pillar is Skills. Unlike a simple predefined prompt, a Skill is a structured folder (Markdown files, descriptions, checklists, and examples) that reflect each customer’s specific development process. It is activated only when relevant, based on its description.

The direct benefit for token consumption comes from a simple mechanism. For example, a developer may ask the AI to add a field to a file without specifying its name, purpose, or size. A properly configured Skill prevents the AI from guessing this missing information. Instead of inventing an answer, the AI directly requests the necessary details. As a result, a potential hallucination is avoided, along with the tokens that would otherwise have been wasted. The Skill intervenes upstream, before the AI even begins working.

A real-world implementation of ARCAD MCP Server clearly illustrates these benefits.

The initial requirement was to allow users to ask natural-language questions (in French or English) about the application repository and quickly obtain precise answers regarding the impact of a change, rather than searching through the source code or consulting documentation that may no longer be up to date.

The complete deployment (scoping, installation, configuration, and initial training) required approximately one day of work. The customer is now considering extending MCP usage to other tools within its ecosystem and to additional teams, particularly customer support, enabling them to query the technical repository in the same way.

The key takeaway: that context is essential to implementing and using ARCAD MCP Server effectively. The quality of the answers will always depend on the quality of the questions being asked. Skills help structure those questions, even when they are initially expressed imprecisely.

Ultimately, the issue is not simply having access to AI: everyone already has access to it. The real challenge is providing AI with the right context, the right tools, and the right safeguards. This is the approach embodied by ARCAD MCP Server: fewer hallucinations, fewer wasted tokens, and greater control over your IBM i environment.