CUSTOMER STORY · LEGAL SERVICES · GOVERNED AI
How BeCloud engineered a governed AI workspace for a law firm using Amazon Bedrock and its Glass Box Framework.
Executive summary
A law firm wanted to use generative AI to accelerate legal research and drafting. But adopting AI raised questions that powerful language models alone could not answer. Who initiated a request? Which confidential documents could the system access? Could conflict-restricted information be excluded before reaching the model? Could leadership trace the sources and decisions associated with AI-generated work?
The firm considered ChatGPT and Claude but needed application-specific controls beyond a standalone chat experience. BeCloud engineered a centralized workspace using Amazon Bedrock, Amazon Cognito, and its Glass Box Framework: application-level controls designed to make AI activity visible, governed, and traceable.
The solution combines identity-based access, conflict restrictions, shared firm prompts, source tracking, grounding checks, and tamper-evident activity records. The objective was to give the firm control over how powerful AI models interact with its people, confidential information, and professional workflows.
The challenge: AI was useful. Accountability was missing.
Attorneys can use generative AI to analyze documents, organize research, summarize complex material, and prepare initial drafts. The firm still needs to protect client information and supervise the work performed on its behalf.
Individual conversations did not provide the consistent workflow the firm needed to determine which information a user was authorized to use, which sources informed a response, and whether the result received appropriate review. Leadership needed answers to five questions:
- Who initiated the AI request?
- What information was the AI authorized to access?
- Were confidential or conflict-restricted documents excluded?
- Which sources supported the generated response?
- What happened to the output after generation?
Enterprise AI platforms offer valuable administrative and security capabilities. This firm also needed restrictions tied to its matters and conflicts, shared workflows, and a centralized record connecting requests, sources, and review decisions. BeCloud approached those requirements as an application-engineering challenge.
The solution: put a control plane around the model
The architecture places the firm’s business rules, identity controls, and accountability mechanisms at the application layer surrounding Amazon Bedrock. BeCloud calls this approach the Glass Box Framework.
The framework does not modify or retrain the underlying foundation model. It controls what information enters the workflow, records how the application processes requests, and evaluates generated responses against available source material. The design addresses six requirements.
1. Every AI request has an identity
Amazon Cognito provides authentication and connects requests to authenticated user identities. The application associates activity with a named user and workflow ownership context, creating a foundation for centralized accountability. Authorized leadership can examine activity by user and action.
2. Conflict restrictions are enforced before retrieval
Consider an attorney working on Matter A who is restricted from documents associated with Matter B. A foundation model does not inherently understand the firm’s conflict database or matter-specific rules.
BeCloud applies authorization checks at document listing, opening, attachment, and retrieval boundaries. Documents restricted for a user are excluded from the context supplied to the model. The objective is to prevent unauthorized information from entering the request, rather than asking the model to ignore it after receiving it.
A prompt telling AI not to disclose confidential information is not a substitute for enforcing access permissions before retrieval.
3. The firm controls its prompts and workflows
Individual subscriptions can lead employees to develop disconnected prompts, templates, and working practices. BeCloud implemented shared firm-defined prompts inside the controlled application. Recurring research and drafting activities can use standardized instructions within the same identity, authorization, and logging framework.
4. Security controls operate around model requests
The application incorporates enabled guardrail and prompt-injection scanning capabilities into request processing. These controls are designed to identify potentially problematic inputs and reduce exposure to certain forms of unsafe or manipulated content.
Amazon Bedrock provides foundation-model access without requiring the firm to manage the underlying model infrastructure. BeCloud’s application supplies business-specific controls. Model safeguards address certain AI interaction risks; application authorization governs which users and documents may participate. Neither replaces the other.
5. Generated answers are checked against their sources
An answer can sound authoritative while lacking adequate source support. BeCloud incorporated an independent grounding check to examine whether claims are supported by the source material available to the workflow. Unsupported claims identified by the check are flagged for attention.
The application also assembles references from the documents and retrieval results actually used, including supported external sources. That provides more useful traceability than a model-generated bibliography alone.
The check does not guarantee accuracy or eliminate hallucinations. A supported claim can rely on an incorrect source, and automated checks can miss errors. The purpose is to make questionable outputs easier to identify and verify. Attorney review remains essential.
6. The application maintains a tamper-evident activity trail
Centralized records connect requests with user identities, source information, system actions, and recorded review decisions. Hash chaining helps detect alterations to the recorded sequence. Authorized partners and administrators can filter the trail by user and action.
Tamper-evident logging is not an immutable archive or independent audit certification. It is an engineering control that improves the ability to detect changes and investigate activity.
What the law firm gains
- Attorneys: a centralized research and drafting environment with source references, visible access restrictions, and flags for potentially unsupported claims.
- Managing partners: a firm-level view of AI activity rather than critical workflow information scattered across individual accounts.
- Administrators: identity-based access controls, centralized prompt management, and searchable activity records.
- The practice: a consistent set of technical controls designed around confidentiality and oversight requirements.
The result connects AI capabilities to operating responsibilities. It is not a claim that the application itself establishes professional or regulatory compliance.
Why not simply use ChatGPT or Claude?
ChatGPT and Claude are powerful platforms, and their enterprise offerings include administrative and security capabilities that may be sufficient for many organizations. This law firm wanted its own application to enforce restrictions tied to internal matters and conflict rules, manage shared workflows, and connect AI requests with sources and review activity.
Those requirements called for integration between the model and the firm’s business systems. BeCloud chose Amazon Bedrock as the model-access layer and engineered the surrounding application. The model generates the answer. The application governs the conditions under which that answer is produced.
Governance that travels with the application
As foundation models become more capable and accessible, the operating value comes from integrating them into real work. Authentication, authorization, data boundaries, workflow controls, source traceability, and operational evidence require application engineering.
The principle extends beyond legal services. Healthcare organizations need patient-information boundaries and auditable access. Financial organizations need controlled information handling and supervisory workflows. Government agencies need authorization, accountability, and records-management controls. Their requirements differ, but the architectural principle is consistent: govern the workflow around the model.
The outcome: greater visibility and accountability
BeCloud’s engagement delivered a governed AI application designed around the law firm’s identity, confidentiality, source-traceability, and oversight requirements. The firm has a centralized architecture connecting AI activity to its access rules and workflow records.
The solution does not eliminate AI risk or replace professional judgment. It adds technical controls to help manage risk and records to support oversight. No quantified productivity or accuracy improvement is claimed in this case study.
Powerful AI is increasingly accessible. Accountable AI requires deliberate engineering.
What should AI be allowed to read in your organization?
Start with BeCloud’s AI Readiness Assessment to explore your workflows and next steps. For matter-specific information boundaries and governance requirements, talk with our team.
Start the AI questionnaire →
Discuss your AI governance requirements →
About BeCloud
BeCloud helps organizations design, secure, and operate cloud infrastructure, custom applications, and AI-enabled business systems. Our approach combines enterprise architecture, security engineering, and operational governance.