AI readiness: the 11 domains to prepare your development team for AI

Your team has Cursor or Claude Code licenses, but half of the AI-generated code has to be reworked? The problem rarely comes from the tool. Here is the AI readiness framework we use at Fragments Studio: 11 domains and 4 levels to know what to put in place, and in what order, so that AI really works for you.

The AI readiness (or AI maturity) of a development team measures how far its code, processes and tools allow an AI to produce reliable work without a human going back over it. It plays out across 11 domains grouped into four families: Product Design, Prompt Engineering, Tech & DX and Tooling. Assessed properly, it tells you exactly what to fix to move from autocomplete to agents working in parallel, with a controlled level of security.

Why isn't AI enough to make a development team more productive?

AI amplifies what already exists, it does not fix it. That is the central conclusion of the 2025 DORA report on AI-assisted software development: AI acts as an amplifier, magnifying an organization's strengths and weaknesses alike. According to Google Cloud, 90% of the nearly 5,000 professionals surveyed already use AI at work, but its adoption is still tied to lower delivery stability when guardrails are missing.

Developers feel it. In the 2025 Stack Overflow survey, 84% use or plan to use AI, but 46% distrust the accuracy of its answers, against 33% who trust it. And 66% name "almost right, but not quite" solutions as their main frustration.

A controlled study by METR goes further: on open-source repositories they knew well, experienced developers were 19% slower with AI, while believing they had saved 20% of their time. The time spent reviewing and correcting wiped out the gain.

At Fragments Studio, we see the same thing with our clients: the difference between a team that saves time with AI and a team that loses time does not come down to the model it picked. It comes down to the foundation: clear specs, written rules, tests that automatically check the agent's work.

What are the 4 levels of AI maturity in software development?

The AI readiness framework distinguishes four levels of AI use, from fully manual work to autonomous agents. Each domain of the framework is scored on this scale.

Our position: a team's real level is capped by its weakest domain. A team equipped with agentic tools but without automated tests stays, in practice, at level 2: every agent output has to be checked by hand. That is why we draw the framework as a radar chart: it shows at a glance the axes holding everything else back.

What are the 11 domains of the AI readiness framework?

Each domain answers a single question: does this foundation let an agent work correctly, without anyone going back over its work?

Product Design: say clearly what to build

• Specs: features described without ambiguity, with user stories, acceptance criteria and edge cases. An agent only builds well what is clearly specified, and our specification document guide details this format.
• Design system: documented, reusable UI components and tokens. With a clean design system, AI produces consistent screens instead of inventing its own styles.

Prompt Engineering: provide the context once and for all

• Rules: the framework the agent follows at all times, in a file read at every session (CLAUDE.md, .cursor/rules or the open AGENTS.md format). It holds the project's conventions and, above all, what the agent must never do.
• Skills: packaged instructions for recurring tasks, for example "create an API endpoint our way". You write them once, the agent reuses them, and Claude Code supports them natively with Agent Skills.
• Codebase intelligence: the AI's ability to understand the repository, thanks to architecture docs, a readable structure and an explained folder map.

Tech & DX: let the agent check its own work

• Tests: automated coverage that lets the agent check what it produces. It is the safety net that makes autonomy possible.
• Quality: linting, strict typing and code review automatically block faulty output before it reaches production.
• Builds & environment: a one-command reproducible setup and a reliable CI/CD let the agent run, build and deploy without human tinkering.
• Observability & security: logs, monitoring, secrets and permissions management. You see what agents do and you limit the possible damage.
• Code standards: consistent patterns and naming. The more predictable the code, the better AI imitates it.

Tooling: adopt the right tools at the right time

Tooling measures the AI tools actually adopted day to day: Claude Code, Cursor or MCP servers. An MCP (Model Context Protocol) server is a standard connector that gives an agent controlled access to your internal tools: tickets, documentation, database. Tooling is the domain where companies invest first, and the one that helps least without the other ten.

How do you secure the use of AI agents in your code?

Securing an AI agent means limiting what it can see and do, then tracing what it did. An agent that runs commands in your terminal has the same rights as you. Without guardrails, a mistake or a malicious instruction hidden in a file (prompt injection, the top risk in the OWASP Top 10 for LLM applications) can have very real effects.

The rules we apply on our projects:

1. Secrets out of reach: .env files and keys are excluded from the agent's context, and production secrets are never present on a development machine.
2. Least privilege: an allowlist of commands, no write access to production, a sandbox or container for agents running autonomously.
3. Vetted MCP servers: every connector is an entry point, so we only install servers whose publisher we know, with read-only rights by default.
4. Mandatory human review: nothing ships to production without a reviewed pull request and a green CI, whether the code comes from a human or an agent.
5. Controlled dependencies: an agent readily adds packages, hence the importance of the lockfile, audits and alerts, as we explained after the Axios npm compromise.

Where do you start when you don't know where to start?

Start by assessing yourself, not by buying tools. Here is the order that works best with our clients:

1. Run the diagnosis: score each of the 11 domains from 1 to 4 and draw your radar chart. The two or three lowest axes form your roadmap.
2. Write the rules file: one page describing the stack, the commands, the conventions and what is off-limits. It is the action with the best effort-to-result ratio.
3. Plug in the guardrails: put linting, typing and tests in CI before widening the agent's autonomy. Without them, AI simply hits the wall faster.
4. Strengthen the specs: require acceptance criteria and edge cases on every ticket handed to an agent.
5. Package recurring tasks: turn into a skill any prompt you are rewriting for the third time.
6. Go parallel only when tests and review keep up with the pace, because several agents produce more code, and therefore more reviewing.

Optimization follows naturally. A well-organized context (short rules, targeted documentation, split tasks) uses fewer tokens and gives better results than a more expensive model used blindly.

This is exactly the outline of the AI Readiness training we already run for some clients who want their developers to be supported: a diagnosis of the team across the 11 domains, then setting up the foundations directly in their own code, so that AI can work in it safely.

Frequently asked questions

What is AI readiness in software development? AI readiness is how prepared a team and its code are for using AI. It is assessed across 11 domains (specs, design system, skills, rules, codebase intelligence, tests, quality, builds, observability and security, code standards, tooling) and 4 levels, from working without AI to autonomous agents running in parallel.

What is the difference between an AI assistant and an AI agent? An AI assistant (autocomplete, chat) suggests code that the developer approves step by step. An AI agent chains several actions on its own, such as reading the code, editing files and running tests, to reach a goal. An agent therefore needs a stronger foundation: written rules, automated tests and controlled permissions.

Should you pick Claude Code or Cursor to get started? The choice of tool matters less than the foundation. Both read rules files, accept MCP servers and work in agent mode. Start with the one your team will actually adopt, then invest in rules, tests and specs, which will stay useful whatever the tool.

How long does it take to move from level 2 to level 3? It depends mostly on the state of your tests and documentation. A team that already has a reliable CI can write its rules and first skills in a few days. Without automated tests, that is the first project, and it conditions everything else.

Conclusion

The productivity gain promised by AI does not come from one more license. It comes from code the agent understands, rules it follows and tests that correct it. The AI readiness framework makes these foundations visible and measurable: 11 domains, 4 levels and a radar chart that tells you where to invest first. Teams that lay these foundations today will be able to hand entire tasks to agents tomorrow, without losing control of their product.

Raise your team's AI maturity

We support tech teams that want to unlock the full potential of AI without sacrificing quality or security.

• Discover our AI integration offer and our AI Readiness training to upskill your developers.
• If your project involves a business tool to evolve with AI, our custom development team can support you.
• Get a first estimate with our project estimator.
• Or let's talk directly.