From Automation to Autonomy

What evidence earns the next degree of autonomy?

A system can summarize a difficult case, recommend an answer, and prepare an action while a person still carries the decision. Another system may execute a narrow action reliably without individual approval. Calling both “autonomous” conceals the operating difference that leaders need to understand.

The practical question is what the business can responsibly entrust to the complete operation. That includes the work before and after the AI contribution, the people handling exceptions, and the consequences when the normal path fails.

My perspective

A stage of maturity should describe an operating commitment the business can keep. I would judge progress by the work, decisions, and changes an operation can reliably manage within its authority. The next investment should resolve the dependency that prevents the next useful commitment. Sometimes that means more autonomy. Sometimes it means better knowledge, a simpler process, or a stronger human decision.

Anthropic’s February 2026 study of Claude Code and API activity illustrates why permission settings alone are insufficient. Experienced Claude Code users approved actions automatically more often and also interrupted more often. These observations concern particular products and users; they do not establish enterprise reliability. For leaders, they reinforce the need to examine actual supervision and intervention alongside delegated authority. [1]

The Automation-to-Autonomy Maturity Curve

Use this Framework when deciding whether to connect fragmented work, introduce decision assistance, delegate a bounded decision, or strengthen adaptation. The unit of assessment is one defined workflow and its consequential decisions, with a named Outcome owner. A department may contain several different patterns.

The curve preserves five capability shifts: Task Automation, Workflow Orchestration, Intelligence-Assisted Operations, Bounded Autonomy, and Adaptive Operations. The observable progression is from executing a task to coordinating work, supporting interpretation, exercising permitted judgment, and improving that bounded operation through verified learning.

This is an editorial maturity model, not a validated score or an inevitable destination. The stages describe increasing operating capability, not equal steps in value, elapsed time, or investment. Required capabilities can be developed together; the curve does not prescribe five separate projects.

Adaptability and authority remain distinct. People can lead effective learning at any stage. Stage 5 here names sustained adaptation of a bounded operation, extending Stage 4; it does not claim that every adaptive service must delegate its decisions. Record valuable human-led adaptation separately when the appropriate decision pattern remains Stage 2 or 3.

What must change at each stage?

Stage 1: Task Automation

A defined activity runs through established rules: transferring a record, checking a required field, or producing a standard notification. People specify the rule and handle cases outside it. Evidence includes accepted outputs, visible failures, and a workable recovery route.

This can be a durable destination for stable work. Its limit becomes material when a faster task leaves the overall outcome constrained by waiting, re-entry, or unowned handoffs.

Transition test: move toward Workflow Orchestration when the business can define the complete path, its owner, and its exception routes. Demonstrate that a failed or delayed step becomes visible and reaches someone able to act. Connecting task automations without resolving responsibility does not establish Stage 2.

Stage 2: Workflow Orchestration

Work is coordinated across systems and teams. Its status, dependencies, handoffs, and completion criteria are visible. Rules can route work and trigger actions; predefined execution alone does not establish context-dependent autonomy.

The leadership question is whether the whole service improves. Check elapsed time to an accepted outcome, unresolved exceptions, repeated transfers, and recovery after an interruption. A local throughput gain can coexist with more work in the next team.

Transition test: introduce Intelligence-Assisted Operations when interpretation is a material constraint and relevant knowledge can be supplied and checked. Test whether assistance improves the human decision after review and correction effort are included. If stable rules solve the problem adequately, continued orchestration may be the better commitment.

Stage 3: Intelligence-Assisted Operations

AI interprets information, assembles evidence, or recommends a response within the workflow. For the decision being assessed, a person retains approval and accountability. Automated preparation or routine routing can coexist with that human decision.

The meaningful evidence is the quality of the combined human-system result. Can reviewers identify missing or conflicting evidence? Can they reject a recommendation? Does the workflow remain useful when the model is uncertain? A fluent explanation does not answer those questions.

NIST’s Generative AI Profile warns that laboratory and benchmark evaluations may not transfer to real operating conditions. Its suggested practices include testing with end users and operators and monitoring outcomes of human-AI configurations. [2]

Transition test: before Bounded Autonomy, define one decision the system may make and act on without individual approval. Demonstrate acceptable performance within that scope, enforceable permissions, detection of boundary failures, timely escalation, and a rehearsed fallback. Approval removed because reviewers are overloaded is an authority change without sufficient evidence.

Stage 4: Bounded Autonomy

The system may use context to select and execute a permitted response. Its operating envelope states the eligible cases, available actions, limits, escalation conditions, and prohibited decisions. A named Outcome owner remains accountable; a system cannot assume organizational accountability.

Permission must hold in the systems that execute the action. A prompt describing a restriction is insufficient if a connected tool still allows the prohibited transaction. Oversight needs trustworthy records and people who can intervene within the time the consequence allows.

Judge Stage 4 using ordinary operation and adverse cases. Review what actually happened, including unauthorized attempts, missed escalations, unresolved work, and the burden on supporting teams. A successful pilot is evidence for a limited commitment; it is not evidence for every case or volume.

Transition test: claim Adaptive Operations only when the business can trace repeated operational signals through an authorized change to a verified outcome and retained learning. More dashboards, faster model updates, or an unused feedback queue do not establish that capability.

Stage 5: Adaptive Operations

The bounded operation detects material change, interprets its implications, coordinates an appropriate response, and establishes whether the response helped. Lessons become maintained knowledge, revised work design, tested controls, or a different investment decision.

Routine adjustments may occur within preapproved limits. Changes to decision rights, policy, consequential logic, or exposure require the designated owner’s review and renewed evidence. Adaptive Operations does not mean unrestricted self-modification or permission to rewrite the operation’s objectives.

NIST’s March 2026 monitoring report identifies the need to observe deployed behavior and unexpected consequences, while documenting gaps in monitoring methods and practice. It supports continued scrutiny after deployment, not a claim that a monitoring platform makes an operation adaptive. [3]

Renewal test: show which change improved which outcome, against what baseline, and whether the benefit held without shifting harm or effort elsewhere. When knowledge, controls, or recovery cease to be dependable, reduce the affected authority and revisit readiness. Maturity is maintained through evidence.

Six requirements constrain the next commitment

The original curve names Knowledge, Governance, Process, Technology, People, and Measurement as requirements across the journey. Use them to locate the constraint on the next transition. They are not six additional stages, and they do not need to reach an identical level of sophistication.

Interpret these together. Poor knowledge can increase review demand; an overstretched review team can turn an approval into a rubber stamp. Faster execution can increase an exception queue beyond the fallback team’s capacity. Funding the visible AI component without the dependency leaves the operating commitment incomplete.

A critical gap cannot be cancelled by strength elsewhere. Treat it as a reason to hold or narrow the affected scope. Conventional automation, selective assistance, and retained human decisions remain legitimate design choices.

The same curve can justify different destinations

Hypothetical design scenarios. These examples test the model’s reasoning. They describe proposed choices and evidence requirements, not client cases or achieved benefits.

Internal service request routing

An IT service desk uses AI to recommend the team that should receive a standard request. Staff review each recommendation. For routing, that is Stage 3, even though the ticket system already coordinates the wider workflow.

The Outcome owner considers Stage 4 for a defined set of routine requests. The system would assign only approved queues; access grants, record deletion, and security incident decisions remain outside scope. Ambiguous requests go to a staffed triage queue. Queue owners maintain the routing knowledge, and the Outcome owner can disable automatic assignment.

The next commitment is a limited trial after shadow evaluation and recovery rehearsal. Measure correct assignments, time to accepted ownership, repeated transfers, unresolved requests, and total triage effort. A shorter initial assignment time is only an expected benefit until the complete outcome improves.

Stage 5 would require demonstrated learning: recurring misroutes lead to an approved correction in service knowledge or routing logic, followed by evidence that fewer requests circulate between teams. The next decision is whether that improvement holds, not whether to grant unrelated access permissions.

Insurance claims evidence review

An insurer has a coordinated claims workflow but adjusters spend substantial effort reconciling documents and locating relevant policy wording. The proposed move is from Stage 2 to Stage 3: AI assembles a source-linked evidence pack and highlights omissions or contradictions.

For this scenario, coverage, settlement, and claim-denial decisions remain with authorized people. The claims owner and relevant control specialists establish the applicable requirements, review responsibilities, and acceptable use. Document access and source verification are part of the design.

Test incomplete files, conflicting records, ambiguous wording, and cases requiring specialist judgment. Measure completeness, factual corrections, reviewer effort, and time to an accepted decision. A good average result cannot compensate for missed evidence that changes a consequential decision.

The appropriate commitment may remain Stage 3 while the insurer improves knowledge and review practice. Faster document preparation does not establish readiness to delegate claim decisions. This service can become more adaptive through human-led learning without being labelled Stage 5 under this curve.

Use the curve to make one accountable commitment

Establish the current position

Bring the Outcome owner, an operator, and the relevant knowledge, technology, finance, and Control leads together. Use the actual workflow, representative cases, exceptions, baseline outcomes, and costs. Identify the decisions under assessment and who currently makes them.

Assign the highest stage whose defining behavior and supporting evidence are demonstrated for that scope. Record uncertainty as unknown. Where a workflow mixes patterns, state them separately: “Stage 4 for routing; Stage 3 for resolution.” Do not average them into a department score or count a demonstration as sustained operation.

Choose the next commitment

Name the operating improvement sought, then apply the relevant transition test and the six requirements. Ask which unresolved dependency most limits the proposal. Agree one of four decisions: hold and strengthen, test a bounded transition, extend an evidenced capability, or reduce scope.

Separate expanding usage from expanding authority. A reliable assisted workflow may serve more users while retaining human approval. A narrow delegated action may need further observation before taking on more consequential cases.

Review outcomes and limits

Set criteria before the trial. Include representative and difficult cases, error consequences, missed escalations, and loss of a dependency. Match the observation period to the variation that matters; no universal number of successful cases proves readiness.

Count review, correction, knowledge maintenance, support, and fallback costs alongside delivery costs. Time released is capacity until the business explains how it will use it. Cash savings require spending that actually falls. The next stage must justify its incremental cost and consequence.

The curve supplies the progression decision. The Autonomous Enterprise Blueprint provides the wider execution responsibilities: Envision, Plan, Build, Operate, Govern, Improve. Governance remains active throughout. Neither a stage label nor a completed record establishes regulatory compliance, model reliability, or a guaranteed return.

Leadership takeaway

The next operating commitment

Choose one workflow. Ask its owner to show what the operation can reliably do today, which decision might change next, and the evidence that would justify that change. Fund the limiting dependency and name the condition that would stop or reverse the expansion. Progress is a more capable operation whose authority the business can explain, support, and renew.

Where to go deeper

• Digital Operating Model — Design the connected work and responsibilities behind the transition.
• Bounded Autonomy — Specify the operating envelope before delegating a decision.
• Self-Healing Enterprise — Examine bounded response, recovery, and operational learning.
• The Autonomous Enterprise Blueprint — Connect the next commitment with the whole execution cycle.

Sources and further reading

The five-stage curve, transition tests, and application method are RePerspective Labs editorial synthesis. These sources support specific premises; they do not validate the maturity model. Evidence checked 13 September 2026.

1. Anthropic — Measuring AI agent autonomy in practice
   18 February 2026
   Provider research on Claude Code and API activity. Supports the distinction between permission settings and observed oversight; visibility and product scope limit generalization.

2. NIST — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
   26 July 2024 · AI 600-1
   Voluntary guidance on evaluation and human-AI configurations; see MP-3.4-005/006 and Appendix A.1.4. It does not certify deployment readiness.

3. NIST — Challenges to the monitoring of deployed AI systems: Center for AI Standards and Innovation
   6 March 2026 · AI 800-4
   Practitioner workshops and literature review on monitoring challenges. Supports post-deployment scrutiny while identifying gaps in methods and practice.