Essay / Nate Busa

The six levels of an AI-native organization

What AI can read, what it can change, and what your company learns from the result.

Enterprise AIOperating modelsAI autonomy

An organization becomes AI-native when useful work can move through it with AI participating in the decisions, actions and feedback that make the business run. Buying assistants is an early step. Building that operating capability requires changes to information, permissions, processes and ownership.

The distinction becomes visible when a customer problem crosses a department boundary. Does a person have to collect the evidence, explain the history again and chase the next team? Or can the company carry the case forward through a shared, accountable process?

The six levels help locate that boundary. L1 improves an individual's work. L2 makes a team's method repeatable. L3 connects work across the enterprise. L4 turns experience into better operations. L5 asks whether the system can also initiate useful work. L0 describes the gap between an AI announcement and an operating capability.

The aim is better business performance: shorter resolution times, fewer errors, stronger service and a lower cost per completed outcome. A higher autonomy label is useful only when it helps deliver those results.

This essay develops the L0–L5 framework proposed by Ann Miura-Ko. The customer scenario and practical interpretation below are my own. Treat the levels as a diagnostic proposal, rather than an industry certification.1

The useful lesson from autonomous vehicles

The automotive comparison forces precision about who does the work and who takes responsibility when the system reaches its limits. SAE's driving taxonomy distinguishes supervised assistance from automated driving within defined conditions. A vehicle's capabilities have a scope; its fallback responsibilities matter.2

Enterprise AI needs the same precision. Preparing an invoice reconciliation, changing the ledger and releasing a payment involve different authority. Calling all three autonomous obscures the decisions that determine how the service should operate.

Classify an operating loop and its authority boundary before classifying the whole company. Support triage might operate at L3 while treasury remains at L1. That can be a sensible allocation of responsibility.

Four questions reveal the operating model:

Lens What I would inspect
Visibility Which records can AI retrieve, under whose identity, with what provenance and freshness?
Action Which changes can it make, and where must it stop for approval or expert judgment?
Extension Can useful workflows become maintained company assets that other people can improve?
Organization Have ownership, handoffs and capacity planning changed because the capability exists?

There is also a separate question of coverage. One excellent workflow serving five people demonstrates a capability. Report how much eligible work it handles successfully before making an enterprise-wide claim.

Six levels grouped by the capability they add: operational substance, individual assistance, a shared team method, connected execution, evaluated improvement and delegated initiative. L5 is marked as a research horizon.
Each level changes how work is carried.The progression concerns context, authority and ownership. It does not prescribe a universal target for every decision.

One customer problem, six operating models

Consider a customer reporting an export failure shortly before renewal. Support holds the complaint, engineering holds the release history, customer success holds the account context, and finance holds the invoice record. The business needs a verified resolution, a reliable account history and an appropriate customer response.

Follow that same case through the six levels. The scenario is illustrative: it makes the operating differences concrete without claiming measured results from a particular company.

L0: AI as theater

At L0, the organization has an AI story but little operational capability. A strategy announcement does not give support access to the release record. A transformation committee does not resolve the customer case. People still gather the information, interpret it and carry it between systems.

The export complaint follows the old route. Support messages engineering, someone checks the account, and finance answers a separate question about payment. The company may discuss AI extensively while the customer waits through the same handoffs.

The first intervention is concrete: choose one recurring process, identify its owner and define a successful outcome. Establish where its evidence lives and how completion will be checked. That work can create value before a sophisticated agent exists.

The test: which recurring business process now completes differently, and who can verify its result?

L1: personal productivity

At L1, a support employee pastes the complaint into an assistant and receives a useful suggested reply. Drafting is faster. The employee still has to find the release history, establish what happened and update the account.

Individual gains matter. A clearer response or faster analysis can improve service. But the context and method may remain private: another employee has to reconstruct the prompt, locate the same information and decide which checks to repeat.

In the customer case, AI helps the employee prepare the work. The employee continues to carry it through the business. The next investment is to capture the useful method in a form that colleagues can use and someone can maintain.

The test: if the strongest AI user leaves, can another employee reproduce the result? Measure time saved alongside checking and correction effort.

L2: team workflows

At L2, support has a shared playbook, common case records and a maintained workflow. AI classifies the export complaint, retrieves permitted support history and updates the triage queue. The next employee benefits from a method the team already owns.

The boundary remains functional. Engineering reconstructs the release context separately. Customer success waits for an update. Finance handles its own queue. Every team can become more efficient while the customer's resolution still stalls between them.

This is an important diagnostic distinction. A collection of useful departmental automations can leave the company dependent on people who relay messages, reconcile conflicting records and chase decisions. Improving each queue does not automatically improve the journey through all of them.

The test: follow a case into the next function. How much evidence must a person collect again, and who owns the unresolved outcome during the handoff?

L3: organizational infrastructure

At L3, the company provides a common way to retrieve relevant evidence and carry permitted actions across systems. The export complaint can be connected to the release history, the approaching renewal and the paid invoice. Source records and stable identifiers make those connections inspectable.

The change is organizational as well as technical. One service can assemble the case, open the engineering investigation, update the account and route the customer response while retaining a named owner. People spend less effort reconstructing context and coordinating routine transitions. For the economic implications, see How L3 AI autonomy changes the economics of a company.

Access remains scoped to the job. A support service needs the records and actions required to resolve its cases. Connectors and application controls enforce those permissions. A model's proposed action cannot grant itself additional authority.

Approval should follow the consequence of the action. Recording a case note might be delegated; a commercial exception might require an accountable reviewer. A blocked action must leave the case with a clear next owner. Reading across systems is useful, but completing the permitted follow-through is what changes the operating model.

Reusable skills can now move across functions. A customer-success specialist can define an investigation method, engineering can supply tested connections, and the business can assign responsibility for the resulting service.

The test: trace one case from its source evidence through the actual record changes to a verified outcome. Where did a person still have to relay context, and why?

The same support, engineering, customer-success and finance records appear in both panels. At L2, people relay context between teams. At L3, scoped connections assemble one shared case, with an authority check before action and an accountable owner through completion.
L2 → L3: the handoff becomes part of the service.The records stay in their source systems. Shared case context and explicit action limits let the company carry the work across them.

L4: compounding operations

At L4, experience changes how future cases are handled. Suppose the export defect was initially sent to the wrong queue. The correction becomes evidence for a proposed change to the investigation or routing skill. The candidate is compared with the current method, checked for regressions and released only when the evaluation supports it.

The useful asset is a maintained improvement process. Saving conversations supplies a record. Demonstrating improvement requires a baseline, cases protected from tuning, and outcomes from later use. A better prompt, retrieval method, tool description or workflow can improve operations without changing model weights.

In our case, subsequent export complaints should reach the appropriate owner with less rework, while unrelated billing cases continue to be handled correctly. That is the claim to test. A growing number of agents or stored memories does not establish it.

People outside engineering can contribute domain knowledge and build useful tools through supported release practices. Product and business owners define the intended outcome and delegation limits. Engineering provides the mechanisms that enforce them. Production tools need ownership, maintenance and retirement decisions.

The test: show an improvement proposed from prior outcomes, its independent evaluation, the release decision and its effect on subsequent work. Can another team reuse it safely?

Verified outcomes and corrections lead to a candidate change and an independent evaluation. A passing candidate enters the shared skill registry; a failing candidate leaves the current version in place. Later outcomes feed the next evaluation cycle, with rollback available after release.
L4: evidence governs what compounds.Improvement reaches future work through evaluation and a release decision. The current version remains available when a candidate fails or later needs rollback.

L5: delegated initiative

L5 is a research horizon in this essay. The defining question is whether the system can formulate useful new work from evidence, pursue it within delegated authority and improve future behavior. A company-wide claim would require sustained evidence across its operating loops.

In our scenario, several customer reports and a release record suggest a pattern before anyone asks about a regression. The system identifies the question, gathers permitted evidence and opens an internal investigation. It records the relationship as an unresolved hypothesis until further verification establishes the cause.

Initiative changes where the work begins. A scheduled run can provide the occasion to look; the capability under examination is whether the system identifies a useful question that a person has not already framed. A fixed threshold alert demonstrates a different mechanism.

The authority boundary still holds. Opening an internal investigation may be delegated. Promising compensation, changing commercial terms or making a consequential production change may require expert judgment. Humans remain responsible for purpose, risk and exceptions.

The test: what useful inquiry did the system initiate, what did it verify, and what did it learn? Inspect its behavior on unfamiliar cases and misleading evidence before accepting a claim of broader autonomy.

A system selects an inquiry from emerging evidence, checks the hypothesis and its authority, then either opens a permitted investigation or escalates to an accountable expert. Both paths retain an owner and an explicit outcome status. An open hypothesis is not a confirmed root cause.
L5: a new source of work, with explicit limits.The system selects an inquiry. Evidence and delegated authority determine what happens next; unresolved questions stay visible.

Diagnose the constraint, then measure the result

The four lenses rarely advance together. Broad visibility with little action authority can produce excellent analysis that waits in a queue. Powerful write tools with weak evidence can accelerate mistakes. Employee-built tools without a maintenance path can create a growing support burden.

Use the mismatch to choose the next intervention:

Observed constraint Practical intervention Evidence of progress
AI cannot retrieve relevant records Establish identifiers, scoped connectors and source provenance More cases investigated without manual evidence gathering
AI can investigate but routine work still waits Delegate a narrow action with policy checks and recovery More verified completions with controlled exceptions
Every team rebuilds the same method Publish maintained skills and a supported extension path Reuse across teams and fewer duplicate implementations
Workflows repeat the same mistakes Capture corrections and outcomes; evaluate candidate changes Less error and rework in protected evaluations and later cases
New tools leave ownership unclear Assign service owners, release criteria and retirement rules Faster recovery and fewer abandoned dependencies

For each operating loop, measure the share of eligible work completed correctly, elapsed time, human handling time, rework, exceptions and total cost. Include inference, infrastructure, integration, maintenance and expert review. Usage and API spending tell you something about activity. The business result requires its own evidence.

Expand delegation when measured outcomes support it, one case type and workflow version at a time. Some decisions should continue to receive expert review even in a highly automated enterprise. The practical choice is which work the company can now trust the system to carry, and what evidence justifies that trust.

Optional technical companion

The runnable LangGraph examples remain available for readers who want to inspect the mechanisms. They illustrate selected control and persistence patterns; they do not validate the organizational levels. Implementation details and setup are collected here so the essay can be read independently.

Download the implementation examples and setup guide

Start with the example guide. It covers Python 3.11+, installation, execution, optional live calls and each example's limitations.

File What it illustrates
requirements.txt Pinned example dependencies
common.py Typed decisions, model adapter and local records
l0_theater.py A status message without operational work
l1_personal.py Personal drafting
l2_team.py A shared support workflow
l3_connected.py Cross-system evidence and resumable approval
l4_compounding.py Evaluated skill promotion
l5_initiative.py Constrained inquiry and escalation

All integrations are local stand-ins. Default decisions are scripted fixtures; live mode replaces them with structured model calls. The L4 fixture scores are demonstration results, rather than a production benchmark. The L5 sketch does not establish open-ended discovery or a self-driving organization.

LangGraph distinguishes workflows from agents.3 Its interrupts support review and resumption; production systems still need authenticated review and protection against repeated side effects.4 Execution checkpoints and application memory also have different purposes; some example checkpoints are in memory only.5

The broader design follows Enterprise automation: from theory to practice: use predictable flows for recurring cases, agents for open-ended investigation, and accountable experts with an AI second opinion for consequential judgment. Product and business developers define the guardrails; engineering implements and verifies them.

References and attribution

The L0–L5 framing and four diagnostic questions originate with Ann Miura-Ko. This essay applies them to an illustrative customer case, with additional emphasis on scoped authority, ownership, verified completion and evaluated improvement. The diagrams are explanatory illustrations, rather than measurements or maturity scores.


  1. Ann Miura-Ko, Everyone wants to be AI-pilled. Most Companies Are Still Level 1, May 1, 2026. The original framework describes the six levels and four questions; the implementation and case interpretation here are the author's own. ↩

  2. SAE International, Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles, J3016. The enterprise analogy does not transfer SAE's technical definitions or certify an organization's autonomy level. ↩

  3. LangChain, Workflows and agents. An orchestrated sequence can remain a workflow even when some nodes call a model. ↩

  4. LangChain, Interrupts. Review and execution controls are responsibilities of the surrounding application as well as its graph. ↩

  5. LangChain, Persistence. Thread checkpoints and application-defined memory serve different purposes; in-memory implementations do not survive a process restart. ↩