监管行业的四种 AI 架构模式

一个用于确定智能体可以在何处进行推理、生成、推荐或采取行动的决策框架。
为什么是受治理的数据、数据血缘、强制执行、评估和审计证据——而非仅靠模型本身——让企业级智能体具备可部署性。
企业级 AI 无法取代所有确定性控制。这四种模式帮助架构师判定智能体可以在何处推理、行动,并始终受控制机制或人类权威的约束。
企业级 AI 背后的架构问题
企业正在向建立在固定规则和可重复流程之上的系统加入智能体。智能体能够对复杂问题进行推理,但企业系统——尤其是受监管行业的系统——必须保持可控且可审计。问题不在于选择其中哪一个,而在于如何让两者协同工作。
我们撰写这一系列文章,是因为客户反复问我们同样的问题:世界各地的企业如何应对与我们相似的约束?企业级 AI 落地的真实情况是怎样的?其他组织如何应对这些细微差别?当需求模糊、复杂且高度特定于某一机构时,Databricks 如何帮助解决这些问题?
作为前线部署工程师(Forward Deployed Engineer),我们就工作在这一边界地带。我们把模糊的业务问题转化为可投入生产的系统。虽然具体实现会因机构而异,且我们无法披露具体的实现细节,但我们可以讨论普适的模式与实践。随着智能体成为大型企业系统的一部分,有两个问题反复出现:
确定性必须保留在哪些环节(例如,当行动不可逆,或错误会带来非对称风险时)?
在哪些场景可以赋予智能体权限,尤其是在行动可逆或补救成本较低的情况下?
答案并非对每个用例都一样。在我们于实际场景中构建和现代化改造的各类系统里,正在浮现出四种模式。本文是第一篇,介绍这一框架。接下来的四篇文章将展示每种模式在实践中如何运作。
图 1. 在受监管企业系统中部署智能体的决策框架。
模式 1:确定性控制与智能体式推理
当结果必须严格确定且可复现,或者当错误不可逆且补救成本高昂时,使用这一模式。在这里,确定性意味着结果可以在一个明确、已版本化的契约下复现,该契约将数据、规则、容差、配置和执行上下文一并记录。
金融机构拥有大量控制措施。对账就是一个很好的例子。控制措施必须计算余额、应用已批准的规则,并产生可复现的 PASS 或 FAIL。智能体不能重新解释容差,也不能把自己的解释变成一项通过的控制。当对账失败时,智能体可以收集证据、检查数据血缘和批处理状态、将差异与既往事件进行比较、测试一组有边界的可能原因,并建议下一步应做什么。确定性控制仍然决定对账是否通过。智能体则帮助解释它为何失败。
智能体可以调查该控制措施。它不能成为该控制措施。
这一模式也适用于监管计算、策略执行、数据质量门禁和系统完整性检查——凡是结果必须精确、可复现且可审计的场景。
模式 2:智能体式执行与确定性评判器
当可能存在多个有效答案、但结果可以被客观检验时,请使用这一模式。
遗留系统迁移或现代化改造就是一个很好的例子。智能体可以分析现有的工作负载并生成目标实现。它每次生成的实现未必完全相同。这并不一定是问题。真正重要的是结果是否正确。
确定性评判器可以检验代码能否编译、模式是否合规、数据是否对账一致、控制总数是否匹配、治理要求是否满足,以及性能是否达到目标。如果结果未通过,智能体会收到相关证据,诊断问题,修正其工作,并再次运行验证。
通过该评判器,只能证明候选方案满足既定的验收契约,并不能证明该契约覆盖了所有生产环境行为。测试覆盖率、贴近生产的边缘案例以及定期评审仍然很重要。
当你能够确定性地评估智能体的输出时,就不要把智能体本身做成确定性的。
验证器不仅是一种控制手段,它还是让智能体得以改进结果的反馈回路。这一模式适用于迁移、代码生成、转换、测试和修复。
模式 3:智能体式推理与人类权威
当答案需要判断、且无法通过确定性测试证明其正确时,使用这一模式。
金融犯罪调查就是一个例子。智能体可以关联交易、文档、政策、关系以及历史模式。它可以汇集证据、识别可能的解释、提出尚未解决的问题,并建议处置方案。当一项决定会显著影响某个人,或者当你必须向监管机构、审计方、法院或受影响的人说明结果时,法规往往也会提出这样的要求。
但建议并非决定。当结果影响重大,且不存在确定性的方式来确立正确性时,决策权必须留在可问责的人手中。
人工审核应位于决策边界,而不是嵌入每一个步骤之中。
智能体的推理过程往往还必须可解释。每一条建议都应能追溯到其背后的证据、来源和工具。这使调查人员能够质疑其逻辑,也使审计方或监管机构能够还原智能体是如何得出该建议的。在法规赋予受影响个人获得解释的权利时,该解释必须针对具体决定。它涵盖主要因素、智能体的角色以及人的角色。“模型把它标记出来了”并不足以说明问题。调查人员要记录最终裁定的理由,包括任何对智能体结论的推翻。
智能体减少了调查工作量,并提高了所呈证据的一致性。调查人员仍对最终决定负责,而该决定可能是主观的、需要细微权衡的,而非确定性的。
模式 4:有限自主性的智能体式执行
当智能体需要采取行动,但只能在明确界定的运行边界内行动时,使用这一模式。
设想一个 Governance Intelligence 模块,负责管理对受治理企业数据的访问。用户描述他们需要什么数据、为什么需要以及需要多长时间。智能体解读该请求,识别合适的数据产品,确定所需的最小访问权限,并收集相关的身份、策略、所有权、分类、血缘和使用上下文信息。
对每个标准请求都要求人工审批会造成延迟,却未必能降低风险。赋予智能体不受限制的授权权限会制造不可接受的控制缺口。答案是有限自主权。
运行边界定义了:
哪些数据产品和访问级别有资格进行直通式处理;
必须通过哪些身份、用途、地理、法律实体、驻留和职责分离规则;
可授予的最大访问时长和最小访问级别;
智能体可以发起哪些已批准的 Unity Catalog 操作;
哪些敏感、模糊或例外请求需要人工审批;以及
访问如何被监控、到期、撤销,以及如何记录以供审计。
在该边界范围内,智能体可以解读请求、汇集证据、选择已批准的工作流,并启动一项有时间限制的授权。在 Unity Catalog 执行授权之前,确定性策略服务会校验每一项必要条件。
不要试图约束每一个决策。要约束运行边界。在这一边界内跨验证周期实施多智能体评判框架,还可以用自动化的多视角安全检查来补充人工监督。
如果请求落在边界之外——例如涉及敏感数据、属性缺失、策略冲突、权限过大,或职责分离方面的例外情况——系统会阻止直接执行,并将其转交给相应的数据所有者、数据管理员或风险审核人员。
现有系统如何为智能体式架构提供支撑
Databricks 平台能力
原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Four AI Architecture Patterns for Regulated Industries)
A decision framework for determining where agents may reason, generate, recommend, or act.
Why governed data, lineage, enforcement, evaluation, and audit evidence-not the model alone-make enterprise agents deployable.
Enterprise AI cannot replace every deterministic control. These four patterns help architects decide where agents can reason, act, and remain subject to controls or human authority.
The architectural question behind enterprise AI
Enterprises are adding agents to systems built on fixed rules and repeatable processes. Agents can reason through complex problems, but enterprise systems-especially in regulated industries-must remain controlled and auditable. The question is not which one to choose. It is how to make them work together.
We are writing this series because customers ask us the same questions: How are enterprises around the world navigating constraints similar to ours? What is the reality of enterprise AI adoption? How are other organizations approaching these nuances? How is Databricks helping solve them when the requirements are ambiguous, complex, and so specific to the institution?
We as Forward Deployed Engineers work at this boundary. We take ambiguous business problems and turn them into production-ready systems. While implementations will vary by institution and we cannot disclose specific implementation details, we can discuss generalized patterns and practices. As agents become part of large enterprise systems, two questions keep recurring:
Where must determinism remain (e.g., when actions are irreversible or errors carry asymmetric risk)?
Where can an agent be given authority, particularly if actions are reversible or low-cost to remediate?
The answer is not the same for every use case. Across the systems we are building and modernizing in the field, four patterns are emerging. This first article introduces the framework. The next four articles show how each pattern works in practice.
Figure 1. A decision framework for placing agents in regulated enterprise systems.
Pattern 1: Deterministic control and agentic reasoning
Use this pattern when outcomes must be strictly deterministic and reproducible, or when errors are irreversible and carry high remediation costs. Here, deterministic means the result can be reproduced under an explicit, versioned contract that records the data, rule, tolerance, configuration, and execution context together.
Financial Institutions have lots of controls. Reconciliations are a good example. The control must calculate the balances, apply an approved rule, and produce a reproducible PASS or FAIL. An agent cannot reinterpret the tolerance or turn its own explanation into a passed control. When a reconciliation fails, an agent can gather evidence, inspect lineage and batch status, compare the break with prior incidents, test a bounded set of causes, and recommend what should happen next. The deterministic control still decides whether the reconciliation passed. The agent helps explain why it failed.
The agent can investigate the control. It cannot become the control.
This pattern also applies to regulatory calculations, policy enforcement, data quality gates, and system integrity checks-anywhere the result must be precise, reproducible, and auditable.
Pattern 2: Agentic execution and a deterministic judge
Use this pattern when more than one answer may be valid, but the result can be tested objectively.
Legacy migration or modernization is a good example. An agent can analyze an existing workload and generate a target implementation. It may not produce exactly the same implementation each time. That is not necessarily a problem. What matters is whether the result is correct.
A deterministic judge can test whether the code compiles, schemas conform, data reconciles, control totals match, governance requirements are satisfied, and performance meets the target. If the result fails, the agent receives the evidence, diagnoses the problem, corrects its work, and runs through validation again.
Passing that judge proves that the candidate satisfied the approved acceptance contract. It does not prove that the contract covers every production behavior. Test coverage, production-shaped edge cases, and periodic review still matter.
Do not make the agent deterministic when you can deterministically evaluate its output.
The validator is not only a control. It is the feedback loop that allows the agent to improve its result. This pattern works well for migration, code generation, transformation, testing, and remediation.
Pattern 3: Agentic reasoning and human authority
Use this pattern when the answer requires judgment and cannot be proven correct by a deterministic test.
Financial crime investigations are one example. An agent can correlate transactions, documents, policies, relationships, and historical patterns. It can assemble evidence, identify likely explanations, surface unresolved questions, and recommend a disposition. This is also often required by regulation where a decision significantly affects a person, or when you have to justify the outcome to a regulator, auditor, court, or the affected person
But a recommendation is not a decision. When the outcome is consequential, and there is no deterministic way to establish correctness, authority must remain with an accountable person.
Human review should sit at the decision boundary, not inside every step.
The agent's reasoning must also often be explainable. Every recommendation should trace back to the evidence, sources, and tools behind it. That lets the investigator challenge the logic and lets auditors or regulators reconstruct how the agent reached its recommendation. Where regulation gives an affected person the right to an explanation, the explanation must be specific to the decision. It covers the main factors, the agent's role, and the human's role. "The model flagged it" is not enough. The investigator records the rationale for the final call, including any agent override.
The agent reduces investigative effort and improves the consistency of the evidence presented. The investigator retains responsibility for the final decision, which may be subjective and nuanced and not deterministic.
Pattern 4: Agentic execution with bounded autonomy
Use this pattern when the agent needs to act, but only within an explicit operating envelope.
Consider a Governance Intelligence module that manages access to governed enterprise data. A user describes what data they need, why they need it, and for how long. The agent interprets the request, identifies the appropriate data product, determines the minimum access required, and gathers the relevant identity, policy, ownership, classification, lineage, and usage context.
Requiring manual approval for every standard request creates delay without necessarily reducing risk. Giving the agent unrestricted authority to grant access creates an unacceptable control gap. The answer is bounded autonomy.
The operating envelope defines:
which data products and access levels qualify for straight-through processing;
which identity, purpose, geography, legal entity, residency, and segregation of duties rules must pass;
the maximum duration and minimum level of access that may be granted;
which approved Unity Catalog actions the agent may initiate;
which sensitive, ambiguous, or exceptional requests require human approval; and
how access is monitored, expired, revoked, and recorded for audit.
Within that envelope, the agent can interpret the request, assemble the evidence, select the approved workflow, and initiate a time-bound entitlement. Deterministic policy services validate every required condition before Unity Catalog enforces the grant.
Do not try to constrain every decision. Constrain the operating envelope. Implementing multi-agent judge frameworks across validation cycles within this envelope can also complement human oversight with automated, multi-perspective safety checks.
If the request falls outside the envelope - for example, because it involves sensitive data, missing attributes, conflicting policies, excessive privilege, or a segregation of duties exception- the system blocks straight through execution and routes it to the appropriate data owner, steward, or risk reviewer.
How Existing Systems Power Agentic Architecture
Databricks platform capabilities
Databricks provides the governed foundation across the four patterns. Omnigent can compose agents and agent harnesses without making any one agent the architecture. Unity Gateway governs access to models, tools, MCP servers, and external APIs. Unity Catalog governs enterprise data and enforces approved access including residency and cross-border transfer rules. Secure connectivity reaches on-premises, cloud, and SaaS systems without bypassing these control points. MLflow preserves traces and evaluation evidence that also serve as audit records for regulators. See the attached diagram for a reference architecture showing how agent calls flow through governed control points to customer systems across on-premises, cloud, and regulated jurisdictions.

Traditional modernization changes where and how technology runs. AI transformation changes the execution model itself. Systems that primarily followed predefined instructions can now interpret, reason, generate, and act.
That does not mean discarding the deterministic systems enterprises have spent decades building. In many cases, those systems become the foundation of the agentic architecture:
Policy engines become control planes.
Validation frameworks become deterministic judges.
Acceptance criteria become agent evaluations.
Operational limits become autonomy boundaries.
Governed data, lineage, and audit trails provide the context and evidence agents need.
The opportunity is not simply to replace legacy systems with agents. Across all four patterns, what makes an agent deployable is the ground it stands on: governed enterprise data, verifiable lineage, enforced access, measurable quality, and a reproducible audit trail.
This is especially important in regulated industries, where a weak architectural boundary can lead to control failures, penalties, or delayed regulatory commitments. A well-designed boundary can reduce investigative effort, shorten modernization cycles, and enable high-value automation without transferring inappropriate authority to the model.
AI-native architecture is about deciding where determinism, intelligence, and authority should live.
Choose the pattern before choosing the agent
Before designing an agentic solution, ask three questions:
What may the agent reason about?
What may the agent be allowed to do?
What must remain outside the agent’s authority?
Those answers determine whether the solution needs a deterministic control, a deterministic judge, a human decision boundary, or an explicit operating envelope. The model comes after that architectural decision-not before it.
The patterns are not mutually exclusive. A production system may use a deterministic control for one decision, human authority for an exception, and bounded autonomy for an approved remediation. The important point is to make each boundary explicit.
As systems mature in production, these boundaries can adapt: deterministic judges refine their evaluation rules using accumulated usage data, while human authority boundaries can shift as feedback establishes confidence in agent reasoning.
The next four articles take each pattern from architecture to implementation:
Part 2: Deterministic control and agentic reasoning
How agents can investigate exceptions around deterministic financial controls without weakening enforcement or auditability.
Part 3: Agentic execution and a deterministic judge
How agents can generate and correct implementations while deterministic validation decides whether the result is acceptable.
Part 4: Agentic reasoning and human authority
How Lakewatch can combine monitoring signals, bank data, and specialist network intelligence into a reviewable evidence package while an accountable investigator retains the consequential decision.
Part 5: Agentic execution with bounded autonomy
A design blueprint for agents to coordinate governed data access inside a policy-defined operating envelope, with deterministic authorization, Unity Catalog enforcement, and human resolution of material exceptions.
Classify your use case using the four-pattern framework, then continue to the implementation article that matches the boundary your system requires. For regulated use cases that cross multiple patterns, contact your Databricks account team or FDE engagement manager to map the control, reasoning, and authority boundaries before selecting the agent architecture.
Molly Nagamuthu is a Principal Forward Deployed Engineer at Databricks. She works with enterprises on complex data and AI transformation initiatives, particularly with global banks where modernization must coexist with governance, controls, and operational rigor.
In this series, Molly and other Forward Deployed Engineers share implementation patterns emerging from work in the field as enterprises move from deterministic architectures toward AI-native systems.
Get the latest posts in your inbox
原文为英文,正文由自动抓取并保留原文。
原文发布于 Databricks Blog(Four AI Architecture Patterns for Regulated Industries),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。