Enterprise AI Architecture: Beyond the Model

Claude Code, Ask Sage, Palantir AIP, IBM watsonx, Databricks AI, and Custom Development

Comparative architecture and implementation strategy for text-intensive, data-intensive, big-data, and DoD financial audit workloads

🟨 Deployment note: Model availability, authorization, classification, contract terms, and approved use must be validated for each environment.

The central conclusion is simple: Claude is a model family; Claude Code is an agent harness; Ask Sage, Palantir, IBM, and Databricks are enterprise platforms. Their value should not be measured only by whether they offer a Claude model. The important question is which Claude Code mechanisms they adopt directly, which remain available through the real Claude Code client or Claude Agent SDK, and which are replaced by stronger enterprise-native controls.

🟦 At-a-glance recommendation
Use Ask Sage for secure text and document assistance and as a controlled route to Claude Code. Use Palantir AIP when the mission depends on an operational Ontology, governed actions, evidence relationships, and workflow. Use Databricks when the center of gravity is Spark, SQL, large-scale data engineering, machine learning, lineage, and evaluation. Use IBM where hybrid infrastructure, legacy modernization, enterprise integration, and deterministic decision services dominate. Build a custom harness only for the mission-specific layer, and keep the underlying model replaceable.

For DoD audit, the best outcome is a layered architecture rather than a single product: a trusted financial data and evidence plane; deterministic reconciliation and control logic; governed MCP or API tools; a constrained, replaceable model layer; human review; and an immutable record of every query, source, calculation, tool call, approval, and change.

Many enterprise AI comparisons collapse four different layers into one. This produces misleading conclusions such as "Platform X uses Claude, therefore it has Claude Code," or "Platform Y does not have CLAUDE.md, therefore it lacks persistent enterprise context." The layers solve different problems and should be evaluated separately.

Figure 1. Four-layer enterprise AI architecture.

🟥 Architecture rule
Do not put authoritative business logic, financial calculations, evidence provenance, access control, or approval authority inside a system prompt. Keep them in governed services that the agent can call but cannot redefine.

Claude Code is an agentic coding system that reads a repository, edits files, runs commands, connects to external tools, and integrates with development environments. Its importance to enterprise architecture is not limited to coding. It provides a clear reference design for context, reusable procedures, deterministic controls, tool connectivity, specialized agents, and extension packaging. [1]

Claude Code does not natively replace a distributed query engine, financial semantic model, enterprise knowledge graph, record-level lineage system, case-management platform, or audit evidence repository. Those capabilities must come from the surrounding enterprise platform or custom architecture.

There are four materially different forms of Claude adoption. An accurate comparison must identify which form applies to each platform and use case.

🟪 Important distinction
When Claude Code is connected through Ask Sage or to Palantir MCP, it is incorrect to say that CLAUDE.md, Skills, hooks, plugins, subagents, permissions, or LSP are absent. They remain available in the Claude Code client. The better question is whether the enterprise platform natively understands, governs, distributes, and audits those artifacts.

Ask Sage has the most direct verified compatibility with the actual Claude Code user experience. Its Anthropic-compatible endpoint is documented as working with the Claude Code CLI and VS Code extension. This lets the Claude Code client continue to provide its own project instructions, skills, hooks, plugins, subagents, permissions and IDE integration while Ask Sage supplies the controlled endpoint, authentication, model routing and government-oriented deployment environment. [4]

Ask Sage also provides its own datasets, personas, prompt library, plugins, web search, Deep Agent and MCP capabilities. Deep Agent performs iterative research across the web and private RAG datasets. These are useful platform-native capabilities, but they are not necessarily format-compatible with Anthropic Skills, Claude plugins or CLAUDE.md. [5]

• Secure access to multiple commercial models through a common interface.

• Rapid text-heavy and document-heavy deployments without building a complete model gateway and RAG application from scratch.

• Controlled Claude Code use for development teams where Ask Sage is an approved endpoint.

• Government-oriented deployment options and user-friendly configuration for non-developer analysts.

• Fast prototyping of policy research, report analysis, drafting, summarization and knowledge assistance.

Ask Sage is best positioned as a secure interaction and model-access layer. It can support document research, NFR and CAP drafting, policy synthesis, auditor-request response, developer assistance and rapid mission prototypes. It should not be the authoritative source for balances, transactions, reconciliation results, data lineage or audit evidence. Those must come from a governed financial data and evidence platform.

Palantir demonstrates the deepest verified integration with Anthropic's programmable agent layer among the reviewed commercial platforms. Palantir documents Claude Code as an IDE integration, provides Palantir MCP and Ontology MCP, and offers explicit agent templates for the Claude Agent SDK with simplified configuration for Ontology MCP, Palantir MCP and the Ontology SDK. [6][7]

This does not make Palantir a reimplementation of Claude Code. Palantir's strategic differentiation is the enterprise platform surrounding the agent: Ontology, object and action security, user permission propagation, operational applications, workflow, cases, branching, lineage and observability.

• CLAUDE.md is replaced at the platform level by governed application context, Ontology definitions, repository context, system instructions and platform policies. The Claude Code file still exists when the external client is used.

• Claude Skills are complemented or replaced by functions, action types, AIP Logic blocks, agent templates, OSDK capabilities and platform applications.

• Claude hooks are complemented by action validation, application restrictions, workflow approvals, branching and platform automation.

• Claude local permissions are complemented by enterprise identity, markings, object/property security and the user's existing Foundry permissions.

• Claude plugins are complemented by Palantir packages, applications, templates, SDKs and MCP services.

• Operational semantic modeling: representing financial, business and audit objects and their relationships.

• Controlled actions: allowing agents to initiate only predefined business actions and workflows.

• Cross-system data integration, case management, remediation tracking and evidence relationships.

• Permission-aware enterprise applications and human-in-the-loop operations.

• Turning analysis into an operational decision, task, case or approved action.

Palantir/Advana is a strong candidate for the operational audit and evidence layer: connecting financial statement line items to accounts, transactions, feeder systems, supporting documents, controls, NFRs, corrective actions, owners and auditor requests. Its value is greatest when the organization needs to navigate relationships and execute controlled remediation, not merely run large SQL queries.

IBM and Anthropic announced a strategic partnership to integrate Claude into selected IBM development tools and enterprise products while retaining IBM security, governance and cost controls. IBM also committed enterprise guidance, reference architectures and open-source assets to the MCP community. [10]

IBM's architecture remains IBM-native. Watsonx Orchestrate acts as an enterprise agent control plane; watsonx.data can expose document libraries through a framework-agnostic remote MCP server; and IBM Bob provides proprietary repository intelligence, software modernization workflows, integration-agent generation and enterprise governance. [11][12][13]

• Java, COBOL, RPG, IBM i and mainframe modernization with deep enterprise context.

• Hybrid cloud and on-premises integration where data and applications cannot be centralized easily.

• Deterministic decision services and business rules for processes that should not be improvised by an LLM.

• Enterprise orchestration of agents from multiple vendors under a common control plane.

• Integration with existing middleware, applications and regulated operational environments.

IBM is most compelling where audit remediation or financial transformation depends on understanding and modernizing legacy code, interfaces and business rules. Claude may help analyze code or explain policy, but IBM's core value is maintaining enterprise process integrity across hybrid and mainframe environments.

Databricks supports external foundation models, custom Python agents, MCP tools and external-agent registration, but its native architecture is model- and framework-agnostic rather than Claude Code-native. Databricks documents support for agents written with any Python authoring library and integrates them with MLflow tracing. It does not currently document CLAUDE.md, Anthropic Skills, Claude hooks, Claude plugins or LSP as native platform constructs. [8]

The appropriate comparison is therefore not whether Databricks copies Claude Code. Databricks provides the data, query, governance, deployment, lineage and evaluation foundation that an agent needs to work reliably at enterprise scale.

• Terabyte- and petabyte-scale ingestion, transformation, joins, aggregation and machine learning.

• Population-level reconciliation, anomaly detection, fraud and improper-payment analytics.

• Lakehouse governance, data access control, automatic lineage and reusable analytical data products.

• ML and agent tracing, evaluation, monitoring and cost/performance optimization.

• Open data and SQL/Spark orientation that can support multiple downstream applications and models.

Databricks is a strong computational and analytical foundation for normalizing ERP and feeder data, constructing transaction populations, performing large joins, reconciling balances, training anomaly models and producing curated exception tables. Palantir or another operational platform may then turn those analytical results into cases, evidence relationships and approved actions.

Direct = explicitly supported as a native or documented integration. Client = supplied by the actual Claude Code client. Analog = proprietary platform capability serving a similar function. Not confirmed = no current official evidence of exact support.

The word "knowledge" describes fundamentally different mechanisms across these products. A model context file, a vector index, a SQL semantic layer and an operational Ontology are not substitutes for one another.

🟩 For financial management
A canonical financial semantic layer should define funds, appropriations, fiscal years, TAS, USSGL, SFIS attributes, balances, obligations, expenditures, disbursements and financial statement line items. These definitions belong in governed data models and deterministic functions, not only in prompts or vector documents.

An LLM should not calculate a material balance by reading thousands of transaction rows in its context window. SQL or Spark should perform the aggregation; a rules engine should test compliance; a statistical or ML model should score anomalies; and the LLM should select approved tools, explain results, identify missing evidence and guide follow-up. This separation improves scale, reproducibility, cost and auditor confidence.

Harness engineering is the discipline of making an LLM operate as a bounded, observable and testable system. The strongest model can still fail if the harness retrieves the wrong context, chooses unsafe tools, loops indefinitely, loses state, bypasses permissions or produces conclusions that cannot be reproduced.

• Curate authoritative collections and preserve document version, effective date, owner, security marking and supersession status.

• Use hybrid retrieval, metadata filtering and reranking rather than relying only on vector similarity.

• Require sentence- or claim-level citations and test whether the cited passage actually supports the answer.

• Isolate retrieved content from system instructions and defend against prompt injection embedded in documents.

• Measure retrieval coverage, citation support, abstention quality and human correction rate - not only response fluency.

• Create governed SQL, Spark or service functions for every material calculation and reconciliation.

• Expose narrow tools such as get_balance, trace_to_transactions, retrieve_support, run_reconciliation and create_case.

• Return structured outputs with row counts, filters, effective dates, data versions and query identifiers.

• Keep agent access read-only by default; require explicit approval for case updates, workflow changes or postings.

• Use the LLM to plan, interpret and explain - never to substitute for a reproducible calculation engine.

For population-scale analysis, the correct pattern is raw data to distributed processing to curated analytical products to governed tools to AI explanation. Raw transaction populations should not be placed directly into an LLM context. The agent should invoke pre-approved distributed queries, receive summarized and traceable results, and drill into selected exceptions only when necessary.

GAO reported in May 2026 that DoD's revised audit approach increases emphasis on technology, including AI, focuses on material line items and evidence supporting account balances, and directs the Department to maximize AI, automation, data analytics and Advana to resolve audit issues. GAO also emphasized continuing concerns about reliable financial information, internal controls, IT weaknesses, transparency, fraud risk and sustainability. [16]

Figure 2. AI-enabled DoD financial audit target architecture.

🟥 Non-negotiable audit principle
The AI may retrieve, classify, explain, recommend and orchestrate. The governed data platform must calculate, reconcile, authorize, record and prove.

A broad GL Extended table can be a valuable analytical spine because it brings many accounting attributes into one queryable population. It is not, by itself, an audit-ready evidence model. It may lack the full sequence of subledger events, interface-control evidence, source-document relationships, status history, reversals and corrections, user approvals, and business-event context needed to explain why an accounting entry occurred.

The recommended design is to retain the broad analytical table for speed while adding governed links to source transactions, subledger events, supporting documents, system/interface metadata, control results and audit cases. Palantir Ontology is one way to represent those relationships; a Databricks-centered implementation can use curated relational models, lineage, metric definitions and a separate case/evidence application.

As of July 13, 2026, the DoD-Anthropic procurement and litigation environment remains unsettled. Current reporting indicates that Anthropic was designated a supply-chain risk for Pentagon contracting and was not included in the May 2026 agreements with seven other AI providers. The architecture should therefore support model substitution and validate model availability for each contract, network, classification and use case. [17][18]

A custom solution should borrow Claude Code's engineering patterns without making Claude-specific artifacts the only way the mission system operates. Mission logic and evidence should remain accessible to other approved models and agents.

• Enterprise identity and access management when an approved service already exists.

• A new distributed query engine or data catalog when Databricks, Palantir or another approved platform already provides it.

• A custom vector database solely because an LLM feature requires retrieval.

• A proprietary tool protocol where MCP or OpenAPI can meet the requirement.

• A model-specific evidence store that becomes unusable when the model changes.

• An autonomous financial posting agent before read-only workflows, evidence controls and human approvals are proven.

1. Do not select a platform based on which one advertises Claude. Select the authoritative data, semantic, workflow and control layers first.

2. Use the real Claude Code client when repository-centric software engineering is the problem. Use the Claude Agent SDK when its harness patterns are needed inside a custom application.

3. Treat Ask Sage as a secure model and text-assistance layer, not as the transaction or evidence system.

4. Treat Palantir AIP as an operational semantic and workflow platform whose differentiation is Ontology, permissions, actions and evidence relationships.

5. Treat Databricks as the analytical and computational foundation for large structured data, reconciliation, ML, lineage and evaluation.

6. Treat IBM as the preferred specialist where legacy modernization, hybrid integration and deterministic enterprise decisioning are central.

7. For DoD audit, begin with trusted data, material-line-item semantics, reconciliation and evidence provenance. Add AI after the deterministic foundation exists.

8. Keep model access replaceable because technical performance, cost, authorization and procurement conditions can change faster than mission systems.

9. Default to read-only agent tools. Introduce actions only through predefined workflows, explicit approval and segregation of duties.

10. Evaluate the complete system - retrieval, data, tools, calculations, permissions, citations and human workflow - rather than the model alone.

🟦 Recommended Advana FM pattern
Governed financial and evidence data + distributed reconciliation and analytics + operational Ontology/case management + narrow MCP tools + replaceable approved models + read-only agent by default + human-approved actions + immutable execution record.

Product capabilities and availability change rapidly. The following official documentation and current reporting support the principal factual statements in this report. Revalidate before acquisition or deployment.

1. Anthropic - Claude Code overview: <u>Open source</u>

2. Anthropic - How Claude remembers your project: <u>Open source</u>

3. Anthropic - Create plugins: <u>Open source</u>

4. Ask Sage - Anthropic-compatible endpoints and Claude Code integration: <u>Open source</u>

5. Ask Sage - Data, Deep Agent and MCP: <u>Open source</u>

6. Palantir - July 2026 announcements, including Claude Agent SDK templates: <u>Open source</u>

7. Palantir - June 2026 announcements, including Palantir MCP and Ontology MCP GA: <u>Open source</u>

8. Palantir - Ontology MCP overview: <u>Open source</u>

9. Palantir - AI FDE overview: <u>Open source</u>

10. IBM and Anthropic partnership: <u>Open source</u>

11. IBM Cloud - Agentic AI workflow with watsonx Orchestrate and MCP: <u>Open source</u>

12. IBM watsonx.data remote MCP server: <u>Open source</u>

13. IBM Bob AI coding agent: <u>Open source</u>

14. Databricks - Use agents on Databricks: <u>Open source</u>

15. Databricks - Unity AI Gateway and Unity Catalog lineage: <u>Open source</u>

15a. Databricks - Unity Catalog data lineage: <u>Open source</u>

16. GAO-26-109115 - DOD Financial Management: Questions Associated with New Financial Audit Approach: <u>Open source</u>

17. Reuters - Pentagon supply-chain-risk designation of Anthropic, March 5, 2026: <u>Open source</u>

18. Reuters - Pentagon agreements with seven AI providers, May 1, 2026: <u>Open source</u>