How ColleagueAI classifies agent autonomy, maps controls to risk, and produces the evidence your auditors and regulators expect.
Agent runtime and customer business content remain in the customer-controlled Microsoft environment. ColleagueAI supplies the agent blueprint, the CAI Score governance pattern, deployment documentation and versioned package updates.
The runtime, identity checks, retrieval and audit trail all run inside your environment.
We package the agent blueprint and its controls so your teams do not start from a blank prompt, a blank workflow, or a blank audit model.
RAG patterns should inherit the user’s existing Entra ID and M365 permissions. The agent should only retrieve or summarize content the user is already authorized to access.
L3 and L4 packages are designed with review, approval, escalation and audit evidence patterns.
AI Act and regulated-use classification depends on customer process, data and deployment context. This is governance evidence, not a legal compliance guarantee.
ColleagueAI is the trust layer for enterprise AI — the governance framework your classified agents run inside. The hard part of enterprise AI isn’t building an agent that works in a demo, it’s governing the sprawl once dozens of them are live: knowing where each one is allowed to act, who’s accountable when it does, and being able to prove it to an auditor. Colleague AI gives you that framework first (the governance blueprint) then fits classified agents into it at the right tier.
The blueprint comes first: every step of your process gets the CAI tier it's allowed to operate at, and each slot is filled only by an agent classified at or below that tier. Agents fill slots in a governed design, they don't get bolted onto a process and hope for the best.
An agent's CAI tier is a hard contract about its autonomy. Low-tier agents only draft and suggest. High-tier agents assist a named human who stays accountable. Safety is structural, not a marketing claim.
Agents run inside your own Microsoft and cloud environment. Colleague AI hosts the governance layer, metadata, configuration, scores, audit trail, entitlements, and partner-registration records. In production deployments, customer operational data is designed to remain inside the customer tenancy. Public website demos, forms, partner registrations, and other submitted website inputs are separate ColleagueAI-hosted flows and should not be used for confidential production data.
We prefer evidence over claims: defined tiers, named accountability, operating controls, and audit records that can be reviewed.
Colleague AI is built by people who have run regulated finance, insurance and SAP operations, and sat through the audits that follow. The CAI Score encodes how those functions actually reason about risk and control, not how a demo looks.
The CAI Score is an open framework: defined tiers, named human roles, logging and review periods. You can see exactly how every agent is classified, and push back if you disagree. A methodology that holds up to scrutiny is one you can actually stand behind.
Agents run in your tenant. Customer business content remains in the customer-controlled Microsoft environment, while ColleagueAI processes separate governance-layer metadata as described in the Trust Center. Actions are logged and attributable, and the evidence is designed to support governance and compliance review.
The higher the tier, the more an agent can do. The controls also increase: human oversight, logging, review points, and named accountability.
Informs and answers. Takes no action and changes no record. The human does the work; the agent just makes it faster to find and understand.
Produces a work product (a document, a query, a report, an outreach message) for a human to review and approve. Nothing the agent makes is used until a person signs off.
Executes routine, low-risk actions inside a bounded workflow, classify, route, fulfil, log. Exceptions and anything unusual are handed to a human. Every action is time-stamped.
Supports decisions and controls in higher-stakes processes, compliance, contracts, security, four-eyes. A named human remains accountable for the call; the agent assists and evidences it. Built for high-risk-process scrutiny.
Acts independently within hard, pre-approved guardrails. Reserved for the highest level of classification, and not used by any agent in this catalogue today.
// This catalogue ships at L2-L4. L4 agents require senior sign-off before go-live.
The AI Act classifies high-risk systems by intended purpose and the Article 6 criteria, not by CAI tier. For Annex III use cases, the CAI Score can structure the governance and evidence assessment, but it does not determine legal classification.
CAI tiers structure governance controls; the AI Act determines legal roles and obligations. The Act splits duties between the provider of a system (whoever develops it and puts it on the market under their own name) and the deployer (whoever runs it under their own authority). In a typical deployment Colleague AI is the provider of the agent package and you are the deployer — but under Article 25 you become a provider yourself if you rebrand it, modify it substantially, or repurpose it into a new high-risk use. The deployer's duties stay with you either way: real human oversight, keeping the agent's automatic logs at least six months, monitoring and reporting incidents, and informing affected people where required. The CAI Score gives both sides the evidence for these duties; it doesn't transfer or discharge them.
Outside Annex III, L1 carries no AI Act obligations beyond existing law. Article 50's disclosure duty (tell the person they're talking to AI) applies only if the tool is customer- or public-facing, not to an internal lookup or drafting aid.
An Annex III system may fall outside high-risk status under Article 6(3) where it does not pose a significant risk, including by not materially influencing decision-making, and at least one statutory condition applies, such as a narrow procedural or preparatory task. Profiling of natural persons remains high-risk. Any reliance on Article 6(3) must be documented before placing the system on the market or putting it into service.
An L3 agent used in an Annex III context may be high-risk depending on its intended purpose and the Article 6 criteria. Where the system is high-risk, applicable requirements can include risk management, logging and effective human oversight.
L4 is designed for high-stakes use cases where stronger oversight and documentation are appropriate. Where a system is high-risk under the AI Act, applicable requirements may include Article 14 human oversight, technical documentation and, where required, conformity assessment.
Not because Article 5 bans it — unbounded autonomy isn't on that list. The problem is Article 14: a high-risk system no one can meaningfully oversee, override, or stop isn't compliant. That's the actual reason this tier stays reserved and unused in the catalogue today.
// Indicative mapping, not legal advice. Classification depends on the specific workflow, not the tier alone — always confirm Annex III scope and any Article 6(3) reliance, including the profiling exclusion, with counsel before go-live.
Agents run where your business data already lives. ColleagueAI holds the governance metadata, scoring, audit evidence, and token-economy indicators around them.
Agents run inside your Microsoft Copilot Studio, Power Automate and Azure estate. ColleagueAI hosts only the governance layer, scores, policies and audit metadata. No customer business data is processed by the ColleagueAI governance layer. Your business data remains in your Microsoft environment.
The CAI tier defines exactly where a person stays accountable. L4 agents never decide alone, they assist, evidence, and hand the call to a named human. Oversight isn't a setting you can switch off; it's built into the tier.
Every agent action is logged, time-stamped and attributable. When an auditor or regulator asks what happened and who approved it, the answer is already a record, not something you piece together after the fact.
Risk classification, human oversight, logging, and transparency are core to the CAI Score. The framework is designed to support EU AI Act, DORA, and ISO/IEC 42001 evidence work, subject to client implementation and legal review.
Six questions. Get your organisation's AI governance maturity level, and the specific gaps to close before a regulator or auditor asks. Built on the same CAI methodology we classify agents with.
Indicative self-assessment. Your answers stay in your browser, nothing is sent or stored.
Direct answers on the CAI Score, deployment, EU AI Act alignment, and how the trust layer differs from an agent management platform.
Colleague AI is the trust layer for enterprise AI. It classifies AI agents against the CAI Score, a five-tier risk classification (L1-L5), documenting each agent’s controls and producing an audit trail. Agents run inside your own environment; Colleague AI hosts only the governance layer, so enterprises can deploy AI they can defend.
The CAI Score is a governance and risk-classification framework for AI agents, described as “the FICO of AI.” It classifies each agent by risk on a five-tier scale, from L1 (Assist) to L5 (Autonomous), defining exactly how much the agent does on its own and where a named human stays accountable, then evidences it for audit.
L1 Assist: informs only, takes no action. L2 Draft: produces work for a human to approve. L3 Operate: executes routine actions inside a bounded workflow. L4 Decide (supervised): supports high-stakes decisions with a named accountable human. L5 Autonomous: acts within hard, pre-approved guardrails. The higher the tier, the more oversight and logging the framework requires.
Agents run inside your own Microsoft Copilot Studio, Power Automate and Azure environment. Colleague AI hosts only the governance layer, scores, policies and audit metadata. Customer business content remains in the customer-controlled Microsoft environment. ColleagueAI processes separate governance-layer metadata as described in the Trust Center.
It supports your compliance work; it doesn't replace it. Risk classification, human oversight, logging and transparency are the core of the CAI Score, and they map to the EU AI Act's high-risk obligations — but those obligations fall on the provider and the deployer, not on a framework. General-purpose AI duties have applied since August 2025 and the Act's transparency duties since August 2026; the main obligations for high-risk Annex III systems were deferred to December 2027 by the 2026 Digital Omnibus. Every agent ships pre-classified on the CAI Score with the evidence an assessment needs — the final classification and sign-off stay with you and your counsel.
Most platforms observe and orchestrate agents (the control plane) or audit models after the fact. Colleague AI adds the missing layer: a portable risk score and classification for every agent (the CAI Score) plus governed agents that ship pre-classified on the CAI Score. It is governance you can act on, not just another dashboard.
Agent sprawl is what happens when AI agents multiply across an enterprise faster than anyone is tracking them, different vendors, frameworks and clouds, each with its own risk profile and no shared audit trail. Giving each agent a CAI Score classification means every one has a known tier and an evidence record, so you always know what you’ve got running and who owns it.
The catalogue includes 36 governed AI agent packages across five functional pillars: Operations & Service Delivery, Risk, Security & Compliance, Data & Infrastructure, Sales & Marketing, and Corporate (HR, Legal, Procurement, Reporting). They run at CAI tiers L2 to L4, with L4 agents requiring senior sign-off before go-live.
The CAI Score is a practical estimate of deployment and governance complexity. It is not a legal compliance rating and does not replace customer-specific risk review.