Subcapability 04 of 05 · Artificial Intelligence

AI Architecture

RAG, agents and MCP: the architecture that moves AI from experiment to production with verifiable data and traceable action.

What is at stake

AI in production with invented data is operational risk with a natural language interface. Most AI architectures in use today respond based on what the model memorized during training, with no mechanism to guarantee that the response is grounded in the company actual data. The error reaches business processes with the appearance of verified fact.

What it is, in practice

Most AI architectures in use today respond based on what the model memorized during training. When a business process depends on current, specific and internal company data, that model fails confidently. The response looks correct. The format is right. The data is invented. That is the central architecture problem for AI in production.

How we work

RAG over internal data

We implement RAG over company data sources so the model responds based on real and retrievable documents. Every response can be traced to its source, private data stays inside the company and content updates do not require model retraining.

Agents with defined scope

We design agents for processes with multiple sequential steps, with explicit boundaries between what the agent executes autonomously and what requires human approval before advancing. Every action is logged for audit.

Integration via MCP

We connect agents to company systems and tools via Model Context Protocol, the open standard that canonicalized how agents receive context and access. With MCP in use, adding a new system to an agent scope drops from weeks to days.

Spec-driven engineering

We implement spec-driven development for AI-assisted engineering: the team describes intent and expected behavior, AI implements, the engineer reviews and approves before anything runs in production.

Build or buy decision

We define the build versus buy criterion for each AI architecture component, using commodity where the market solves the problem and building where there is real business differentiation, without accumulating vendor dependency where the company needs control.

Measurable gains

What changes in the result when this subcapability matures.

Rate of responses traceable to a verifiable data source

RAG over internal data raises the proportion of responses that can be audited and traced to the source document, reducing the risk of hallucination that reaches business processes with the appearance of fact.

Time to integrate a new system into the agent ecosystem

MCP standardizes agent-to-system connections. With the protocol in use, integrating a new system into an agent scope drops from weeks of custom development to days of configuration.

Percentage of agent actions with human approval at high-impact points

Architecture with defined approval points before execution guarantees control without eliminating automation. The human approves where impact is high. The agent executes where the criterion is already clear.

Delivery cycle speed with AI-assisted engineering

Spec-driven development shortens the implementation cycle by separating code generation from review accountability. The engineer retains ownership of what goes to production.

Frequently asked questions

What is the practical difference between RAG and fine-tuning the model on company data?

Fine-tuning incorporates patterns from company data into the model during training. RAG retrieves the data at query time and grounds the response in the recovered document. The operational difference is that RAG updates with current data without retraining the model, the response can be traced to its source, and private data stays under company control rather than being sent to the model vendor during training.

What is MCP and why did it become the integration standard for agents?

MCP is the Model Context Protocol, an open protocol that defines how AI agents connect to tools, systems and data sources in a standardized way. Adopted by major model vendors from 2025 onward, MCP reduced integration cost because it eliminates the need to build a custom connector for every agent-to-system pair.

When does it make sense to use AI agents instead of traditional automation?

When the process has multiple steps that depend on variable judgment, involves access to different systems and can be described with clear criteria for success and failure. Traditional automation executes a fixed rule. An agent executes a process that requires reading context at each step. What determines whether the agent is safe is the clarity about what it can do without human approval.

How do you ensure the agent does not execute actions outside the expected scope?

With a defined set of possible actions, human approval points before high-impact actions and a complete log of every action executed. The agent operates within defined capability boundaries. Actions that exceed a defined impact threshold require confirmation before executing. The audit log is what allows responding to any incident without relying on manual reconstruction.

Does data need to be ready before implementing RAG?

Data quality is what determines response quality. RAG over fragmented, unstructured or outdated data produces responses as imprecise as a model without retrieval. Data preparation work, structuring, cleaning and indexing, must precede the implementation of the retrieval system. Immature data is the main reason RAG projects do not deliver in production what they demonstrated in the pilot.

Want clarity on where to invest first?

A complete technology capability assessment with an evolution roadmap connected to financial result.

Get your assessment Talk to an expert