--- title: >- The three levels of AI delegation, assistive, advisory, autonomous, and why each raises the evidentiary bar description: >- As AI moves from explaining options to acting on a customer's behalf, the burden of proof rises with it. A governance framework for the emerging proxy economy, built around three levels of delegation. author: murilo date: '2026-06-11' tags: - ai-governance - proxy-economy - fca - decision-traceability categories: - ai-governance --- One of the more useful frameworks to emerge from recent regulatory thinking on artificial intelligence is the idea that AI systems delegate decision-making to differing degrees. The FCA, in its examination of what it calls the emerging proxy economy, has described three levels of delegation: assistive, advisory, and autonomous. The framing is helpful not only as a way to classify systems, but because it reveals something governance teams often miss. The level of delegation determines the evidentiary bar. As a system moves from explaining options to acting on a customer's behalf, the standard of proof a firm must be able to meet rises with it. This article works through the three levels, sets out what changes at each, and argues that the evidentiary bar is the right lens through which to govern AI, more useful, in practice, than abstract debates about model risk. ## Level one: assistive AI Assistive AI explains, compares, and surfaces information for a human decision-maker. A chatbot that helps a customer understand the difference between two pension options is assistive. So is a tool that summarises a policy document, or one that retrieves relevant information for an adviser to consider. The defining characteristic is that a person remains squarely in the decision. The AI informs; the human decides. At this level the consequences of an incorrect output are real but bounded. If the assistive system surfaces something misleading, a competent human in the loop has the opportunity to catch it before it affects the customer. The evidentiary bar is correspondingly modest. A firm should be able to show what information the system provided and that it was reasonable, but the decision itself, and the accountability for it, rests with the person who made it. This is not an argument for complacency. Assistive systems shape decisions even when they do not make them, and a system that consistently frames options in a particular way can steer outcomes. But the burden of proof at level one is meaningfully lighter than at the levels above. ## Level two: advisory AI Advisory AI nudges, recommends, and guides. An AI suitability engine that recommends portfolio adjustments based on a customer's risk profile is advisory. So is a system that recommends a particular product, or flags a customer as suitable or unsuitable for a course of action. The human may still confirm the decision, but the AI has now expressed a view, and that view carries weight. The evidentiary bar rises sharply here, for a specific reason. When a system recommends, the firm has to be able to show that the recommendation was appropriate for that customer. This is the territory of suitability, and of the Consumer Duty's expectation that firms deliver good outcomes. If a regulator or an ombudsman asks why the system recommended what it did, "the model suggested it" is not an answer. The firm must be able to produce the basis for the recommendation: the inputs that described the customer, the output, the model version, and the context that made the advice suitable at the time. A subtle risk appears at level two: the human confirmation can become a formality. If a person is confirming hundreds of model recommendations a day, the meaningful decision has effectively been delegated to the model, even though a human signs it off. Governance has to account for the system as it actually operates, not as the org chart describes it. That means the evidence has to stand on its own, because the human sign-off may not bear the weight the firm assumes it does. ## Level three: autonomous AI Autonomous AI acts within defined boundaries without human intervention for each decision. An algorithmic system that executes orders on market signals, or one that automatically rebalances a portfolio, operates at this level. There is no per-decision human checkpoint. The system acts, and the action has effect. At level three the evidentiary bar is at its highest. The firm cannot rely on a human in the loop to catch errors, because there is none for the individual decision. The entire weight of accountability falls on the firm's ability to reconstruct, explain, and defend each decision the system made, after the fact, at scale, and potentially years later. When the system is acting on a customer's behalf, the firm is answerable for every one of those actions as if it had made them itself, because in a regulatory sense it has. This is what makes the proxy economy a governance challenge rather than merely a technical one. A system acting autonomously on behalf of consumers generates a continuous stream of consequential decisions, each of which the firm may one day need to account for. The volume alone defeats any approach based on reconstructing decisions from fragmentary logs. ## The bar rises with the delegation The pattern across the three levels is consistent. At level one, a human decides and the evidence supports them. At level two, the model recommends and the firm must evidence that the recommendation was appropriate. At level three, the model acts and the firm must evidence every action without the safety net of human review. The further a firm delegates, the more it must be able to prove. This has a practical implication for governance. The question a firm should ask of any AI deployment is not simply "how good is the model" but "what level of delegation is this, and can we meet the evidentiary bar that level demands?" A firm comfortable deploying an assistive chatbot may be nowhere near ready to evidence the decisions of an autonomous system, and the gap is not closed by improving the model. It is closed by building the capability to capture and produce a complete record of every decision. ## Governing for the highest level you deploy Most firms do not operate at a single level. They run assistive tools, advisory engines, and increasingly autonomous systems side by side, and the mix shifts toward greater delegation over time. The prudent approach is to build evidence infrastructure for the highest level of delegation the firm deploys, and to apply it uniformly. A system that captures complete decision provenance for an autonomous trading engine will comfortably cover an assistive chatbot; the reverse is not true. Aegis Trace was designed around exactly this principle. A single integration captures complete provenance for every AI decision, assistive, advisory, or autonomous, linking each output to its inputs, model version, and context, with tamper-resistant storage and regulator-ready exports. As a firm's systems move up the delegation ladder, the evidence layer does not have to be rebuilt to keep pace. ### Match your evidence to your level of delegation. Request early access to Aegis Trace and our technical documentation. [Request Access →](/request-access)