The Five Levels of Authority for Building AI Agents

A data center does not fail because an algorithm lacked ambition. It fails when the wrong system gets authority. There, thermal optimization competes with uptime promises.

The first question is not how do you automate cooling. It is should you let an agent act at all. That framing gives building teams a better way to judge building AI agents.

The five levels

The core issue is authority, not software style. Cognitive Corp’s trust chain starts with connectivity, then context, then Governance, then evidence, then lifecycle continuity. If any link is weak, higher authority becomes harder to justify.

• Level 1: Observe

The agent reads telemetry and produces summaries. It does not explain causes or trigger changes. This level depends on a maintained OT asset inventory and usable taxonomy.

• Level 2: Interpret

The agent maps signals to assets, spaces, and relationships. It can explain what a point means in context. This level depends on semantic interoperability, not raw data volume.

• Level 3: Recommend

The agent proposes actions, priorities, or likely faults. A person approves or rejects the next step. This is where Human-in-the-Loop should remain explicit.

• Level 4: Bounded action

The agent can execute narrow actions inside fixed limits. Those limits need defined assets, known context, review paths, and strong rollback discipline. NIST’s AI building work ties authority to conformance, metrics, reliability, cybersecurity, semantics, and digital twins.

• Level 5: Delegated control

The agent holds standing authority over a narrow operational domain. It acts within approved boundaries without waiting for each approval. This is the hardest level to justify because changing conditions can break the original assumptions.

Level 5 is not just a bigger Level 4. It is a different risk class. Once action becomes continuous, missing context stops being a reporting issue and becomes an operational issue.

How to use the ladder

Treat the ladder as an escalation test. Before any step up, ask what new proof exists, who owns the decision, and what evidence stays current after change. If those answers are weak, stop at the lower level.

For example, a REIT can use Level 3 for chilled-water recommendations across several offices. Staff can review each suggestion, compare it with tenant commitments, and approve only what fits the day’s conditions. That is very different from giving the same agent Level 5 control across the portfolio.

What is established now

Most buildings are not blocked by a lack of dashboards. They are blocked by weak inventory, weak naming, and weak context between systems. That keeps many teams at Levels 1 through 3, even when the interface looks mature.

A second reality is easy to miss: Security ≠ Governance. A protected connection does not answer who can interpret tradeoffs, override intent, or act when evidence is incomplete.

That gap becomes obvious in healthcare. In an operating room, energy optimization can collide with sterility requirements. Safety wins, and the agent often lacks that full context.

What is emerging

NIST’s current direction matters because it connects the pieces many teams treat as separate. AI-enabled buildings are being framed alongside conformance, metrics, reliability, digital twins, semantics, cybersecurity, and grid interaction.

That is important for authority decisions. Better predictions do not justify broader permissions by themselves. Evidence has to mature with the level of action.

Cognitive Corp inference

Cognitive Corp’s useful inference is simple: stronger authority requires stronger proof. Connectivity moves data. Context makes the data usable. Governance decides who or what may act. Commissioning creates trusted evidence, and lifecycle management preserves that trust as the building changes.

That logic makes most Autonomous agents a poor fit for instant write access. They should earn authority in stages. A team that cannot explain the jump from Level 3 to Level 4 should not make it.

Open questions

Three questions remain open. First, how should portfolios retest agent authority after a retrofit, software update, or asset replacement? Second, how should conformance travel between buildings when naming and metadata differ?

Third, who owns authority when facilities, IT, and vendors share one operational loop? The answer is organizational, not only technical.

FAQs

What level should most teams start with?

Most teams should start at Level 1 or Level 2. They need inventory, context, and review discipline before broader authority makes sense.

When does Level 3 become valuable?

Level 3 helps when teams want faster triage without giving up approval rights. It fits fault review, prioritization, and operating recommendations.

What separates Level 4 from Level 5?

Level 4 acts inside a narrow box with clear limits. Level 5 keeps standing authority when conditions change, so the evidence burden is much higher.

Why does semantic context matter so much?

Without shared meaning, an agent reads points but misses relationships. It sees data, not the system the data belongs to.

Does strong cybersecurity justify write access?

No. Cybersecurity protects the connection and the system. Governance decides whether an agent should act, and under what limits.

Final thought

The right first upgrade is not more automation. It is better permission.