# Agentic AI: The Five Levels from Assistance to Autonomous Systems Author: Marin Niehues, Niehues Consulting · Published: 2026-09-07 · AI Adoption & Operations Source: https://www.niehuesconsulting.de/en/blog/agentic-ai-maturity-levels/ Agentic AI matures in organisations across five levels, from simple assistance to an autonomous system that handles workflows on its own on a continuous basis. With each level, more work passes from people to the system, while the demands on data and control rise. This article describes the levels, shows when each one pays off and provides a self-assessment for your own organisation. Agentic AI changes the division of labour, because a task is delegated to a system that returns a result people can check. How this step fits into AI implementation as a whole is set out in the [guide to AI implementation](/en/blog/ai-implementation-guide/). Because questions of human oversight and liability are involved: This article provides an overview and does not constitute legal advice. **The maturity of agentic AI is measured by how much responsibility for a workflow a system takes on. The economic leap happens where work is delegated, which means from level 3 onwards.** ## What is agentic AI? Agentic AI refers to AI systems that receive a goal, plan the steps towards it themselves and use approved tools in company systems to get there. An assistant answers a question. An agent completes a case, such as a complaint from receipt through to the credit note. The vendor Anthropic describes workflows as systems in which models and tools are orchestrated through predefined code paths, whereas agents direct their own process and tool use. A research group at the University of Washington describes autonomy through the role of the human, from the operator who directs every step to the observer who monitors. The maturity model below reflects the same shift in the language of operations. An overview of the term is given in the article on [agentic AI](/en/blog/agentic-ai-workflows/). ## The five levels of agentic AI The levels build on one another. Without reliable workflows, agents fail for lack of interfaces and test criteria. ### Level 1: Assistance A person asks, the AI answers. Most organisations that use AI today are at this level. - **Technology:** Chat interfaces and copilots in office and development tools, complemented by search over internal documents using retrieval augmented generation. - **Benefit:** Time savings for individuals on drafts and research, hard to measure for the organisation. - **Prerequisites:** Approved enterprise tools and a usage policy staff are trained in. - **Risk:** Unchecked answers and confidential data in private tools. Responsibility lies with the person who adopts the result. ### Level 2: Workflow AI works at fixed points in a defined workflow. It routes incoming items and drafts replies, while the workflow itself remains fixed in code. - **Technology:** Model calls at fixed process steps, linked as prompt chains or with routing by case type. Connection to business systems runs through APIs. - **Benefit:** Shorter lead times in high-volume workflows such as invoice checking, with predictable behaviour. - **Prerequisites:** A documented workflow with accessible systems and checks on every step. - **Risk:** Low while inputs follow the expected pattern. New case types make the workflow stop or sort incorrectly. For many organisations, level 2 is the right entry point. Anthropic recommends starting with the simplest solution and increasing complexity only when it measurably improves the result. ### Level 3: Agent The agent receives a goal instead of an instruction. It plans the steps, uses systems such as CRM or ERP through defined interfaces and delivers a result that a person reviews. - **Technology:** Models with tool access, connected through open standards such as the Model Context Protocol (MCP). Added to this are guardrails and evaluations against test cases in sandboxed environments. - **Benefit:** Entire cases with variable inputs run without manual steps. Scaling begins here, because capacity is no longer tied to headcount. - **Prerequisites:** A named process owner and least-privilege permissions. Every step is logged, with a fixed escalation path. - **Risk:** Wrong actions in live systems and prompt injection. Costs rise, because an agent consumes a multiple of the computing power per case. ### Level 4: Agent team A lead agent breaks a large task down and assigns the parts to specialised subagents working in parallel. One agent researches while a second checks the sources. - **Technology:** Orchestrator-worker architecture. Subagents work with their own context and report back only results. Across system boundaries, agents communicate through protocols such as Agent2Agent (A2A). - **Benefit:** Broad, divisible tasks are completed faster and more thoroughly. According to the vendor, a multi-agent system from Anthropic outperformed a single agent by 90.2 per cent on an internal research evaluation. - **Prerequisites:** Stable individual agents with clearly cut tasks. A review instance consolidates the partial results. - **Risk:** Errors propagate between agents, and root-cause analysis becomes more demanding. According to the same source, computing effort is around 15 times that of a chat. How several agents are steered is covered in the article on [multi-agent systems](/en/blog/multi-agent-systems-orchestration/), and the standards behind them in the article on [MCP and A2A](/en/blog/mcp-and-a2a-for-decision-makers/). ### Level 5: Autonomous system The system runs continuously and is triggered by schedules or events. It works within clear guardrails and escalates to people as soon as a case falls outside its authority. Leadership and accountability remain with people. - **Technology:** Continuously running agents with triggers and monitoring in operation. Approval and escalation points are firmly defined. - **Benefit:** A field of work is handled around the clock, and people concentrate on exceptions. - **Prerequisites:** Several weeks of stable metrics at level 3 or 4 and an operating model with on-call cover. - **Risk:** Errors affect many cases before they are noticed. Level 5 is reserved for work with low damage per error or full reversibility. *Interactive tool: Use case check, maps agentic approaches (workflow, agent, agent team) to each business function and bottleneck, with a metric and evidence from practice. Available at https://www.niehuesconsulting.de/en/blog/agentic-ai-maturity-levels/* **What vendor data shows about agentic AI** 90.2% | Performance advantage of a multi-agent system over a single agent on an internal research evaluation (vendor figure, Anthropic) 4x | Computing effort of an agent compared with a chat, around 15x for multi-agent systems (vendor figure, Anthropic) 77% | Share of analysed enterprise usage via the API showing automation patterns, meaning the task is fully handed over (vendor figure, Anthropic) 10,000+ | Active public MCP servers at the time the standard was handed over to a foundation under the Linux Foundation (vendor figure, Anthropic) ## Maturity self-assessment: where does your organisation stand? Your organisation stands at the highest level whose prerequisites it fully meets in at least one production workflow. Pilot projects do not count. **Checklist: maturity of agentic AI** - Level 1: Staff use approved AI tools under a written policy. - Level 2: At least one workflow uses AI at fixed process steps, with a measured error rate and lead time. - Level 3: An agent completes cases in connected systems. Process ownership, step log and escalation path are defined. - Level 4: Several specialised agents work under a lead agent, and a review instance consolidates the partial results. - Level 5: A system works continuously and is event-driven. People intervene only in exceptions, and the metrics appear in regular reporting. Many organisations place themselves at level 3 because an agent is running in a pilot, yet they do not meet the prerequisites of level 2 in regular operation. ## Which level pays off for your organisation? The level that pays off is the lowest one that completes the workflow in full. Each higher level costs more computing power and more operating effort, and it is worthwhile only if it creates value the level below cannot deliver. **When each level pays off** Criterion | Levels 1 and 2 | Levels 3 to 5 | With Niehues Consulting Inputs | Structured or individual requests | Variable and spread across several systems | Structured where possible, agent only for the variable steps Volume | From a few cases per week | Several hundred cases per month as a guideline | Volume and cost per case calculated in the business case Damage per error | Low, a person checks every result | Limited and catchable before it reaches the customer | Autonomy per workflow tied to error rate and level of damage Cost per case | Low and easy to plan | Multiple computing effort, fluctuating | Cost per case as a metric in regular reporting Accountability | With the person using it | With a named process owner | Process ownership and escalation path set before the build Three criteria decide the move from level 2 to level 3. First, variability: if rules capture the workflow fully, the workflow stays cheaper. Second, volume: an agent pays for itself only at several hundred cases per month. Third, the cost of errors: an error must be caught before it reaches customers. Level 4 adds a fourth criterion, a task divisible into independent parts. ## How much human control does agentic AI need? Agentic AI needs named human accountability at every level, and the form of control changes with the level. At levels 1 and 2, a person checks every result (human in the loop). From level 3, a person approves critical steps and reviews samples. At level 5, the person monitors the system through metrics and intervenes in exceptions (human on the loop). Moving up a level follows metrics agreed in advance: the error rate in the sample, the share of cases without intervention and the time to correct a detected error. If they hold the target over several weeks, the system gains authority, and if they deteriorate, it drops back a level. Four rules apply to governance: - **Permissions:** Each agent receives only the permissions its workflow requires, with write access to financial data behind approval points. - **Log:** Every step is recorded with its rationale as evidence of due care. - **Escalation:** It is defined who takes over a case the system cannot complete, and within what time. - **Shutdown:** A named person can stop the system at any time, and the workflow continues manually. For high-risk systems, the European AI Act requires effective human oversight, and it is good practice for any system acting in live workflows. How approval and monitoring can be designed is described in the article on [human in the loop](/en/blog/human-in-the-loop-vs-human-on-the-loop/). ## The route to the next level with the Niehues Delivery Model [Niehues Consulting](/en/) supports organisations on their way through the levels using the Niehues Delivery Model, from a delivery perspective and into regular operation. Marin Niehues is a management consultant. He builds organisations that deliver and brings AI into operation, in projects ranging from VC-funded scale-ups to DAX 40 corporations and in engagements in more than 20 countries on six continents. The model has three phases with fixed review questions. For agentic AI, three of them matter most: - **Understand:** Which business outcome should change measurably? The answer determines which level is the target. - **Build:** Does the prototype prove its value in the real workflow with real data? An agent starts with approval by a person and collects the metrics that justify moving up. - **Implement:** Who is accountable for operation from day one? Without a named process owner and escalation path, every system remains stuck at pilot stage. The starting point is the [free 30-minute initial consultation](/en/#kontakt), in which maturity and the next step are assessed. **Key takeaways for decision-makers** The maturity of agentic AI shows in prerequisites met in regular operation. Those who choose the lowest level that completes the workflow and tie autonomy to metrics achieve scale without losing control. **Sources** - Anthropic: Building effective agents, vendor publication, 2024 - Anthropic: How we built our multi-agent research system, vendor publication, 2025 - Anthropic: Anthropic Economic Index Report, September 2025, vendor publication - Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation, vendor publication, 2025 - Feng, McDonald, Zhang: Levels of Autonomy for AI Agents, arXiv, University of Washington, 2025 - Bundesverband Deutscher Unternehmensberatungen (BDU): Honorare in der Unternehmensberatung, fee study 2022 ## Frequently asked questions ### What is agentic AI in simple terms? Agentic AI is AI that receives a goal, plans the steps itself and carries them out in company systems. It works like a clerk with a clear job description who hands cases outside their authority over to a person. ### What is the difference between an AI workflow and an AI agent? A workflow follows a path fixed in code, with AI working at defined points. An agent determines its own path and tools in order to reach a goal. Workflows are more predictable and cheaper, while agents cope with variable inputs. ### How do I assess the agentic AI maturity of my organisation? Your maturity level is the highest level whose prerequisites you fully meet in at least one production workflow. What counts is named accountability and measured metrics in regular operation, and pilot projects do not count. ### Which level should an organisation aim for first? For most organisations, level 2 is the right entry point, because it creates the interfaces and metrics every higher level requires. Level 3 follows for workflows with variable inputs and several hundred cases per month. ### When is a multi-agent system worth it? A multi-agent system is worth it when a broad task splits into independent parts, such as research across many sources. With computing effort around 15 times that of a chat according to the vendor, the added value must exceed the extra cost. ### Is a fully autonomous AI system permitted in a company? A level 5 system works on its own but remains under human accountability and oversight. For high-risk systems, the European AI Act requires effective human oversight. Organisations therefore define approval points and a shutdown option. ### What are the risks of AI agents? The main risks are wrong actions in live systems and manipulated inputs through prompt injection. In agent teams, errors that propagate between agents are added. The risks are contained through minimal permissions, step logs and evaluations before any extension of autonomy. ### How much does introducing AI agents cost? Costs consist of development, ongoing model usage fees and operation. External consulting in the German market is typically billed at around 1,000 to 2,500 euros per consultant day. Scope and terms of a project are discussed in the free initial consultation.