Written by: Jens Eriksvik

Enterprise AI isn’t measured by how much math it can do. It’s measured by how much work it’s trusted to carry. Accuracy and latency tell us what a model could do. They don’t tell us what the organisation actually lets it do. The real milestone is when an AI agent takes on a step that used to require human oversight, and completes it end to end. That shift, its Span of Responsibility (SoR), is the best sign of AI maturity.

Building on Algorithma’s previous articles, we formalise Span of Responsibility (SoR) as the proportion of a workflow entrusted to an AI agent in production, not in theory, not in a test environment, but in live, accountable work.

“The real sign of progress isn’t a smarter agent, it’s the quiet shift where work is handed off without a second thought. The same way trust grows in any team. It shifts the conversation from "How smart is our AI?" to "How much real work is our AI trusted to do?" - which is what ultimately matters for business impact.”

- Jens Eriksvik, CEO

A real example: Purchase order approval

The value of SoR becomes clear when applied to real-world workflows. Consider a PO approval process, a standard sequence in virtually every enterprise, often managed by procurement platforms like Coupa, or SAP Ariba. These systems automate workflow routing and enforce rules, but they do not carry responsibility. Human oversight remains essential at key decision points.

An AI agent designed with a span of responsibility changes that. Instead of simply moving tasks between humans, the agent manages specific validation and decision-making steps within the process.

While the procurement system orchestrates the workflow, the AI agent executes and owns specific validation and decision tasks. These include budget checks, vendor compliance reviews, routing decisions, and even pre-screening approvals for policy adherence. An AI agent can be scoped to work with up-to-date policies and steering documents, without significant system configuration and/or development. All agent-handled tasks are rule-based and logged, supporting auditability and compliance with frameworks such as SOX and ISO 9001.

By automating these early-stage responsibilities, the organisation reduces approval cycle times, minimises manual errors, and ensures consistent compliance across purchases. Most importantly, responsibility shifts from people to systems; incrementally and measurably.

Capability vs. responsibility

This distinction can be made explicit by comparing model capability to SoR. Capability reflects what the AI system could do under ideal conditions or in testing. SoR reflects what the organisation has actually delegated to the agent in live workflows. The table below illustrates this difference:

Responsibility is an operational assignment in an organisation. The gap between capability and real-world responsibility is reflected in industry assessments. While many organisations invest heavily in AI technologies, few have advanced to the point of systematically delegating decision-making and workflow ownership to AI agents.

An AI model may be technically capable of classifying, scoring, or even making recommendations with high accuracy. But until a team decides to delegate that decision-making or task ownership to the agent, defining scope, guardrails, and escalation paths, the model is not carrying responsibility.

Responsibility isn’t baked into code; it’s granted by the organisation. SoR measures that grant and marks the shift from passive outputs to agentic workflows. As recent industry discussions note, agentic AI is not just about chaining tasks, it’s about granting agents the authority to manage workflows and outcomes. Organisations are moving from task-based AI assistants toward agents that own outcomes across business processes, with accountability, guardrails, and operational trust.

This principle builds directly on Algorithma’s earlier work describing how the future of work will be organised. In that view, AI agents don’t just assist, they take on roles traditionally held by human team members, with clear spans of control, accountability, and escalation logic. As responsibility shifts, so does the structure of teams and the nature of collaboration between humans and digital colleagues. SoR offers a measurable framework for bridging the gap between capability and real-world responsibility.

Measuring Enterprise AI maturity: Introducing the SoR maturity model

SoR isn’t a snapshot. It’s a progression. As organisations increase their trust in AI agents, and as agents prove capable of handling broader scopes, the share of workflow responsibility grows. SoR provides a framework for creating digital colleagues.

To make this progression measurable, SoR can be structured into a five-level maturity model. Each level reflects a meaningful increase in responsibility, decision-making authority, and integration into real-world operations.

The SoR Maturity Ladder

These thresholds reflect real-world tipping points. At Level 1, human supervision begins to decline as the agent takes on consistent, low-risk tasks. By Level 2, automation starts delivering measurable ROI by displacing significant manual effort across an entire process. At Levels 3 and 4, responsibility expands beyond individual tasks to decision-making and coordination across functional boundaries, requiring governance mechanisms equivalent to those used for human roles.

Progress along the maturity ladder is not binary. Many organisations will have agents operating at different levels simultaneously, depending on the function, risk tolerance, and complexity of the workflow. The maturity ladder makes responsibility growth visible, highlighting not just technical capability upgrades but the actual delegation of work. It also aligns technical deployment with operational trust, ensuring that governance, audit practices, and retraining evolve in step with the agent’s expanding role.

Traditional AI maturity frameworks focus on organisational readiness or technical capability but rarely capture operational responsibility. Models often emphasise strategic posture and intent but lack task-level resolution. Other approaches focus on tech stack maturity and compliance but overlook workflow ownership. The same pattern is reflected in recent analyses like the MIT Sloan Management Review’s AI Maturity Levels, which assess strategy, data, and culture but stop short of measuring how much work is actually entrusted to AI agents. In contrast, SoR measures the percentage of real work actually delegated to AI agents, providing a metric that is measurable, auditable, and based on live operations. As a result, SoR tracks what the agent owns, not what the organisation hopes it might do.

Implications of SoR

As Span of Responsibility increases, the impact extends well beyond the agent itself. Responsibility growth reshapes how systems are designed, how teams operate, and how governance and risk are managed. More importantly, it redefines how organisations distribute work, accountability, and decision-making between human and AI team members.

SoR is not just a metric. It’s a lens for understanding the operational and organisational transformation required to scale AI responsibly.

As Span of Responsibility increases, AI adoption stops being an IT initiative and becomes an organisational transformation. Teams must evolve from supervising tools to collaborating with digital colleagues. Decision rights, accountability structures, and skill requirements will shift accordingly. Governance will move from periodic oversight to continuous, integrated monitoring. Critically, AI literacy must become a core organisational skill, not only for technical teams but for managers and frontline staff who will increasingly assign, supervise, and collaborate with AI agents.

Scaling SoR is not just about deploying more advanced agents. It’s about redefining how work is assigned, how success is measured, and how humans and AI share responsibility across the enterprise.

How to operationalise SoR

Recognising SoR is only the first step. To use it as a practical management tool, organisations must embed SoR into how they assess, design, govern, and scale their AI initiatives. SoR isn’t just a maturity model, it becomes a planning framework and a management discipline.

Operationalising Span of Responsibility requires more than project-level thinking. It must become part of the enterprise’s operating model. Leadership should set responsibility growth targets, not just capability milestones, ensuring that AI adoption aligns with broader business goals. AI teams must work closely with process owners to scope, monitor, and adjust SoR boundaries as agents evolve. Governance bodies should incorporate SoR into oversight reviews and compliance reporting, making responsibility growth a formal part of risk and performance management.

Equally important is workforce enablement. As AI agents take on more work, human team members will shift into supervisory and escalation roles. Developing AI literacy across the organisation is essential, not only for technical teams but also for managers and frontline staff who will assign, supervise, and collaborate with AI agents. Responsibility growth is not just a technical progression. It’s an organisational transformation that requires trust, clear accountability, and shared understanding between humans and their digital colleagues.

Measure what you delegate

SoR moves the discussion from model capability to delegated accountability. By mapping SoR in a single live workflow, procurement, incident triage, or contract drafting, organisations gain a factual view of where trust already exists and where manual checks still dominate. Review the data in the next planning cycle; it will indicate whether controls are sufficient or responsibility can safely expand.

In practice, the rhythm is simple: delegate, measure, adjust. Fold SoR reviews into regular operational reporting and the metric becomes a quiet, consistent driver of both efficiency and governance.

When the agent takes over: Measuring enterprise AI by work owned, not math done