The Earned Autonomy Model: A New Way to Think About AI-Native Development

Every consultancy, every vendor, every conference keynote now claims to be "AI-native." Almost none of them can tell you, in concrete terms, where an organisation actually stands. Most conversations skip straight to tooling, which assistant, which agent, which model, without asking a more useful question first: how much can this organisation actually trust an AI system to act on its own, and has it done the work to earn that trust?

That question is the basis of a framework we have started using internally and with clients, which we are calling the Earned Autonomy Model.

Autonomy is earned, not switched on

The core idea is simple. Giving an AI system more independence is not a tooling decision, it is a trust decision. And trust, in any system, has to be earned through two things running alongside the autonomy itself: alignment, meaning the system is working from a genuine understanding of business and product intent rather than generic best practice, and feedback, meaning real signal from production and from real users actually makes its way back into what gets built next.

Skip either of those, and adding more autonomy simply means making mistakes faster and further from anyone noticing.

6 levels, 3 lenses

We assess every level against the same three questions: who is actually deciding what gets built (autonomy), is that decision genuinely informed by current strategy or just convention (alignment), and does real-world outcome data change what happens next, and how quickly (feedback).

Level 0: Ad Hoc Development

Work is manual, individually driven, and undocumented. There is no shared tooling and no LLM use at all; quality depends entirely on the person doing it. Alignment to strategy is incidental, and feedback, where it exists, only surfaces after something has already gone wrong.

Level 1: Tool-Assisted Experimentation

People start using LLMs, but outside the actual workflow, in a separate chat window disconnected from the codebase or backlog. The gains are personal, not organisational. Feedback is still manual and retrospective.

Level 2: Embedded Augmentation

AI assistance moves into the development environment itself, becoming default rather than optional. It is still reactive though, suggesting rather than initiating, and while team conventions are starting to get encoded into the tooling, business and user signal still has no route back in.

Level 3: Workflow Standardisation

Reusable, governed patterns appear: prompt libraries, structured tasks for tests and documentation, agreed ways of working. This is also the point where product management properly enters the picture, with specs and backlog items starting to be drafted with AI pulling directly from documented strategy rather than from what is in someone's head. Feedback becomes a deliberate, scheduled habit rather than an accident.

Level 4: Agentic Delegation

Agents take ownership of discrete pieces of work end to end: implementation, testing, opening the pull request. Human oversight shifts from directing every step to reviewing outcomes. Agents work from current priorities directly, and production and user signal arrives quickly enough to shift what gets worked on next, sometimes within the same day.

Level 5: Autonomous Value Orchestration

Business intent and live user signal flow continuously into what gets shipped, largely without point-in-time human direction. Human involvement narrows to setting strategic intent, governance, and the occasional exception review. Build, alignment, and feedback have become one closed loop rather than three separate activities.

Where most organisations actually sit

In our experience, most leadership teams believe they sit at level three or four because they have rolled out a coding assistant somewhere in engineering. Look closer, and the reality is usually level two: the tooling is embedded and people are using it daily, but alignment and feedback have not caught up, so the autonomy on offer has not actually been earned yet. The tooling got ahead of the organisation.

It is not only a development question

The same three lenses apply just as well to product management as to engineering. A backlog written without reference to current strategy is just as misaligned as a line of code written without it, and a roadmap that never hears from real users is just as broken a feedback loop as a deployment nobody monitors. The Earned Autonomy Model is, deliberately, a model for how an organisation works, not a scorecard for a development team.

Where do you sit

We built this model because it is the conversation we keep having with clients in banking, insurance, defence, and the public sector, all of whom are under pressure to move faster with AI and all of whom are, quite sensibly, nervous about what they are actually handing over when they do. If you want to work out honestly where your organisation sits, and what it would take to earn the next level, we would enjoy that conversation.

Authors: Abraham Schoots, Abhijeet Selukar (partners and co-founders at Lithe transformation)