That is not a slogan about our intentions; it describes where the decision rights sit. The AI recommends. Your firm's own risk policy, codified into executable rules, decides. A named human approves, and nothing material happens without one. Every step is recorded against the exact system state that produced it.
Roles under the EU AI Act
The AI Act allocates obligations by role. For JUSTE the allocation is:
| Provider | JUSTE AI Ltd. We develop the system and place it on the market, and we carry the provider-side obligations: technical documentation, risk management, logging capability, transparency information and instructions for use. |
| Deployer | Your firm. A regulated entity using JUSTE under its own authority to perform AML and CTF obligations is a deployer, responsible for using the system according to its instructions, assigning competent human oversight, controlling input data, monitoring operation and retaining logs. |
| Affected persons | The people and entities screened, classified or subject to due diligence through the system: your clients and prospective clients. |
| Model providers | JUSTE orchestrates general-purpose AI models from third parties. Obligations attaching to those models rest with their providers. Our provenance records identify the model behind every material output. |
The practical consequence for a deploying firm: you do not need to produce technical documentation for the AI system itself — that is our job. Your job is to show that you deployed it knowingly, configured it under your own risk policy, supervised it with named competent people, and kept the records.
What the system does
JUSTE is an agentic operating system for AML compliance. A workforce of specialised AI agents executes screening, classification and due-diligence workflows under rules codified from your senior-management-approved risk policy, with defined human decision points. Each deployment is anchored to your own policy; the system does not apply a generic risk appetite.
- Screening of natural persons and legal entities against sanctions, politically exposed person, adverse media, sector and territorial risk sources, with entity cross-referencing for KYB.
- Codification of your written risk policy into executable rules that govern, and where required override, AI risk recommendations — with each override recorded.
- Generation of customer and enhanced due-diligence questions targeted at the specific risk drivers screening surfaced, with a supervising agent evaluating the quality and completeness of responses.
- Production of an inspection report for every file: what was screened, what the AI found, what the policy required, where policy overrode AI, what the human did, and the supporting evidence with timestamps.
What it does not do
- It is not a decision-maker of last resort. High-risk classifications trigger a hard system block requiring authentication by your MLRO or nominated senior authority. Refusals and weak enhanced due-diligence answers escalate to senior management for a documented human decision.
- It does not file suspicious activity reports or make tipping-off judgements. Those remain human, firm-level acts informed by the system's output.
- It is not a general-purpose chatbot. Agents operate only within codified workflows; free-form use of the underlying models is not part of the deployed system.
- It is not a substitute for your risk policy or for senior management responsibility. The policy governs the system; the system does not generate your risk appetite.
- It does not give legal advice or determine that a business relationship is lawful.
Human oversight
Oversight in JUSTE is structural rather than optional. The system is built so that the people overseeing it can understand its output, can intervene, and in defined situations must intervene before anything proceeds.
| AML Team Lead | Runs day-to-day screening and simplified or customer due diligence. Reviews recommendations, confirms or queries classifications, manages document collection. Cannot proceed past a high-risk block. |
| MLRO or nominated senior authority | The only role able to unlock a high-risk file. Authentication is required by the system itself — a hard block, not a convention. Reviews enhanced due-diligence outcomes, decides escalated files and owns refusals. |
| Senior management | Owns and signs the risk policy that governs the system. Receives escalations where enhanced due-diligence answers are refused or inadequate. Cannot pressure a fee earner into bypassing the policy without that becoming a recorded, attributable event. |
The intervention points are: review — every classification reaches a human with its reasoning and provenance before it affects the client relationship; override — policy overrides the AI automatically where the rules require, and humans may additionally adjust, with who, when and why recorded; stop — any file can be held, and high-risk files are held by design; escalate — weak or refused answers route to senior management, and the system does not resolve these itself.
Known limitations
These are inherent to the task rather than defects, and every deploying firm should understand them.
- Source dependence. Screening quality depends on the coverage and currency of the underlying sources at the moment of screening. A nil result evidences what those sources contained at that moment; it is not a guarantee about the subject.
- Probabilistic components. Adverse media analysis and risk recommendations involve probabilistic models. They are recommendations into a governed process, not determinations — which is precisely why the policy layer and the human decision points exist.
- Identity resolution. Name matching across jurisdictions, scripts and transliterations produces false positives and false negatives. Medium and high-risk workflows are designed to surface and resolve these rather than assume them away.
- Policy quality in, governance quality out. The system enforces your policy as codified. If the policy is outdated or unsigned, it will faithfully enforce an outdated policy.
- No legal determinations. Output is evidence and structured recommendation for your own decision.
Provenance and versioning
Every material output carries provenance metadata: the model used, the agent, the prompt version and the policy version in effect. Changes to agents, prompts or models are versioned, which means any historical decision can be re-examined against the exact system state that produced it. Before reaching human authority, each material output passes through a producing agent, a quality-assurance agent and a supervising agent.
Data, training and location
Customer data and end-client data are never used to train, fine-tune or otherwise improve our models or those of our sub-processors, in any form. This is a contractual commitment, not a configurable setting, and it applies without exception.
Processing takes place predominantly in the United Kingdom and the European Economic Area. Certain sub-processors, including some AI model providers, process data in the United States under appropriate safeguards. The current sub-processor list is published at juste.ai/security-and-trust.
For data protection purposes, a deploying firm is the controller of client personal data processed for AML purposes, on a legal-obligation basis, and JUSTE acts as processor under the Data Processing Agreement. Our Privacy Policy describes both roles, including the screening data we hold as controller.
Regulatory classification
Whether a given AML risk-classification system falls within Annex III of the AI Act is a legal question that depends on the use case and on guidance still being developed. Draft, non-binding European Commission guidelines published in May 2026 indicate that AI systems intended for AML and CTF purposes will not ordinarily fall within Annex III unless their purpose is functionally linked to evaluating creditworthiness or establishing a credit score — a position consistent with JUSTE's scope. The same draft guidelines note that agentic systems whose interacting components jointly produce outputs materially influencing a decision are assessed holistically, which is directly relevant to our producer, quality-assurance and supervisor agent architecture.
Our working position is deliberately conservative. The platform is built and documented to Annex III high-risk standards regardless of where the classification finally settles, so that neither JUSTE nor a deploying firm depends on a favourable interpretation. Decision provenance, human oversight points and logging are native to the architecture rather than a retrofit.
Prohibited-practice and AI-literacy obligations have applied since February 2025; no JUSTE function falls within the prohibited practices list. General-purpose AI model obligations have applied since August 2025 and rest with the model providers. Deploying firms should record their own classification view with their counsel's input, and should treat draft guidance as persuasive rather than final.
Documentation for your audit
Supervisors increasingly ask a specific question: show me how your policy governed this decision about this client. Not "show me your screening tool", and not "show me your policy document". Every JUSTE inspection report is built to answer that question per file, evidencing what was screened and when, what the AI recommended with model attribution, what the policy required, where policy overrode AI, what the named human did, and what evidence supports the outcome.
Firms deploying JUSTE receive the full AI System Transparency and Audit Support pack, which includes a fill-in deployment record that converts the generic document into your firm's own signed, standing answer to an auditor.
Ask us the full pack at hello@juste.ai.
Reporting a concern
Suspected malfunctions, implausible outputs or anything you regard as a serious incident should be reported to compliance@juste.ai without undue delay. We investigate, and where provider-side reporting duties to authorities are engaged, we discharge them.
Material changes to agents, models, prompts or the policy engine are versioned and notified. When you reissue your risk policy, the new version is codified and takes effect from a recorded date; prior decisions remain governed by, and auditable against, the version then in force.
It describes the system and the legal framework in good faith. Each firm remains responsible for its own regulatory compliance and should confirm its position with qualified counsel, particularly on classification and territorial scope.