The 4 trust zones: classifying your AI agents.

Why classify before deploying

Deploying an agent without classifying it is like signing a blank check. Some decisions are reversible and harmless; others commit the company's liability. The LOOP™ framework requires ex-ante classification, in one of the 4 trust zones, before any production deployment.

Classification is not a defensive exercise. It's a velocity lever: once an agent is classified as green zone, teams can use it freely. The clearer the governance upstream, the smoother the usage downstream.

Classifying is liberating. Without a framework, nothing moves. With a framework, everything becomes possible.

Green zone: supervised autonomy

The agent acts alone, its decisions are logged, and spot-checked. Humans don't intervene in real time but retain the authority to disconnect, retrace, or correct.

Typical examples

• Automatic sorting of incoming documents (invoices, CVs, tickets)
• Email categorization (support, commercial, administrative)
• Internal request routing to the right department
• Entity extraction (dates, amounts, names) from structured documents

Criteria: limited risk, reversible actions, low-cost errors, high volume.

Orange zone: active supervision

The agent proposes, but a human validates within a defined period (24h, 4h, real-time depending on the case). The error can be caught before it leaves the organization.

Typical examples

• Client response (validated by a human agent)
• Candidate pre-qualification (validated by a recruiter)
• Legal summary drafting (validated by a lawyer)
• Supplier order proposal (validated by a buyer)

Criteria: visible potential impact, partial reversibility, human available for the validation loop.

Red zone: mandatory human validation

Every decision goes through a human before taking effect. The agent is a co-pilot, never the pilot. This is the zone reserved for decisions that commit the company.

Typical examples

• Financial decisions: wire transfers, spending commitments
• Legal decisions: signatures, contract terminations
• Individual HR decisions: offers, sanctions, appraisals
• Sensitive external communications: press releases, media responses

Criteria: major impact, limited reversibility, binding regulatory framework.

Black zone: prohibited

Some use cases must not be entrusted to an agent, even with human validation. The black zone is not a convenience — it is the ultimate guardrail.

Typical examples

• Medical diagnosis affecting a person's health
• Disciplinary decisions or dismissals
• Processing sensitive data not classified by the DPO
• Scoring high-legal-risk populations (credit, insurance)

The black zone is reviewed quarterly. A use case that was prohibited may be authorized (in red zone) if the regulatory framework evolves.

How to move an agent from one zone to another

Classification is not fixed. An agent deployed in red zone can move to orange zone after 3 months of documented stability. An agent in orange zone can move to green zone after 6 months without incident.

The reclassification process involves:

1. A quantified review of decisions made (volume, error rate, drift)
2. A supervision review (how many times did humans correct?)
3. Validation by the business sponsor and the CAIO
4. Update of the living registry

In practice

Koneetiv deploys the LOOP™ framework in organizations via:

• Claude Ignite for initial classification
• Claude Cockpit for ongoing review
• Ignite AI Act for monitoring

The 4 classification criteria in detail

Classification into a zone rests on four objective criteria. Each is rated on a simple scale (low, medium, high) and the zone is determined by the combination of all four.

Risk

What is the impact of an erroneous decision? Risk is measured on three dimensions: financial (cost of an error), legal (regulatory exposure), reputational (image impact). An agent that recommends a product has low risk; an agent that commits a spend has medium to high risk.

Volume

An agent used 10 times per day doesn't have the same criticality as one used 10,000 times. Volume changes the nature of risk: even a low error probability becomes critical at scale.

Traceability

Can you reconstruct the decision after the fact? Can you explain why the agent chose one action over another? Traceability is essential for regulated use cases (finance, legal, HR).

Reversibility

Can the action be undone? An email sent to a client is barely reversible. A ticket routing is completely reversible. An executed wire transfer is not reversible at all.

Symptoms of poor classification

How to know if your classification is wrong? Four common symptoms:

• A green zone agent generating too many incidents — it should have been orange zone
• A red zone agent consuming too much human time — it could have been orange zone with a threshold
• An orange zone agent always validated without modification — it can probably move to green zone
• An orange zone agent frequently rejected — its prompt needs revisiting or it should move to red zone

The reclassification process

Agent reclassification is governed. It's not an individual decision — it's a collegial one. The typical process:

1. Metrics collection — volumes, error rates, incidents, drift, user satisfaction
2. Committee analysis — business sponsor, CAIO, IT
3. Documented decision — with justification and revocation thresholds
4. Registry update — new zone, new review date, new sponsor if needed
5. Stakeholder communication — users are informed of the change

Pitfalls to avoid

Three pitfalls cause classification to fail:

Classifying by model, not by use

The zone does not depend on the model (Claude, GPT, Llama). It depends on the use of that model. Two agents using the same model can be in two different zones.

Avoiding the red zone

The red zone is not a zone to avoid. It's a zone where humans remain in structured control. Refusing the red zone on principle leads to incorrectly classifying agents as orange zone, at the organization's risk.

Forgetting the black zone

The black zone is not a failure. It's an explicit decision NOT to use AI for a use case. This decision must be documented and reviewed periodically.

Classification as a dialogue tool

Classifying an agent is not just an act of governance. It's a dialogue tool between business teams, IT, compliance, and the board. Each speaks their own language: business talks value, IT talks technical risk, compliance talks rules, the board talks exposure. The 4 trust zones provide a common vocabulary.

We have observed in several organizations that simply introducing LOOP™ vocabulary unlocked debates that had been stalling for months. An agent blocked by the CISO can restart once you agree to classify it as red zone and define the escalation rules. An agent blocked by the business can restart once you agree to classify it as orange zone with enhanced supervision.

How to document a classification

A classification is only useful if it's documented. Here is the minimum information to keep in the living registry:

• Agent name and ID
• Business sponsor (name + role)
• IT lead
• Use case (brief description)
• Zone (green, orange, red, black)
• Zone justification (4 criteria rated)
• Escalation rules (who validates, within what timeframe, what thresholds)
• Classification date and next review date
• Incident history
• Reclassification history

Classification anti-patterns

Three anti-patterns to avoid:

Universal classification

"We classify all our agents as red zone by default." This is false prudence. It overloads humans, blocks simple use cases, and discredits the governance as a whole.

Model-based classification

"Everything using Claude is in zone X." This confuses the tool and the use. Two agents using the same model can be in two different zones.

Fixed classification

"Once classified, we never change." This negates the "Living" principle of LOOP™. Governance must be alive: each agent evolves, improves, or drifts. Classification must follow.

Classification is the first decision in your AI deployment. It determines everything else.

Let's talk about your use cases →