September 2026 Summaries

CData announces Connect AI Gateway, a platform designed to govern interactions between enterprise AI agents, employees, models, and business data by combining model and MCP gateway capabilities with routing, guardrails, observability, policy controls, and end-to-end identity management. Drawing a comparison to the evolution of SD-WAN, the announcement argues that centralized visibility and control are necessary to address growing AI costs, security risks, reliability concerns, and accountability requirements. Built on CData’s existing connectivity to enterprise systems, the gateway is intended to compute answers through source-aware data layers rather than expose raw records to models, which CData says improves accuracy, reduces context sent to models, lowers token costs, and speeds responses. The company also emphasizes delegated user identities, source-system permission enforcement, validated writes, detailed audit trails, and spend attribution. CData plans to expand the platform with more intelligent workflow routing, natural-language policy creation, deeper organizational context, and integrations with security, identity, and observability providers, positioning the gateway as a layer for responsible enterprise AI transformation.

Enterprise AI agents that connect Salesforce and ERP systems require a staged approach covering readiness assessment, model and orchestration selection, secure data connectivity, retrieval design, testing, production monitoring, and governance. Organizations are advised to begin with low-risk, measurable use cases, verify data quality and access controls, and select models and workflow frameworks based on regulatory, latency, deployment, and process-complexity requirements. The proposed architecture uses retrieval-augmented generation to ground responses in current enterprise data, with CData Connect AI presented as a managed MCP-based gateway that accesses live source records while inheriting source-system permissions rather than relying on potentially stale vector indexes. Recommended security controls include least-privilege RBAC, OAuth or SSO, encryption, audit logging, version management for prompts and tools, approval gates for consequential actions, and ongoing alignment with standards such as SOC 2, ISO 27001, and GDPR. The text also emphasizes sandbox and human-reviewed testing, measurement of accuracy, latency, compliance, and unauthorized-access attempts, followed by incremental rollout, alerting, rollback capabilities, and optimization based on operational feedback.

Real-time streaming integration continuously captures and moves data as events occur, delivering sub-second to seconds-level latency for operational needs such as fraud detection, live inventory, cybersecurity, and IoT monitoring, whereas batch ETL processes accumulated data on scheduled intervals and is generally simpler and less costly to operate. Common streaming approaches include event streaming through durable logs such as Kafka, log-based change data capture (CDC) that reads database transaction logs for ordered inserts, updates, and deletes, query-based change polling for sources without log access, and hub-and-spoke architectures that reduce point-to-point integration complexity. Proprietary databases, legacy ERPs, and custom APIs complicate these patterns because they may lack accessible logs and require source-specific handling for pagination, authentication, rate limits, schema changes, and deletion tracking. Standards-based interfaces such as ODBC, JDBC, and REST can abstract these differences through reusable connectors, allowing integration tools to work with diverse sources through consistent access methods. The discussion contrasts hand-coded, serverless, hosted, and pre-built connector approaches, noting that custom development offers control but increases maintenance demands, while managed infrastructure and connector platforms can reduce operational work; CData Sync is presented as one option supporting CDC for selected databases and incremental extraction for many other sources.

CData introduces Connect AI Gateway, an enterprise AI connectivity and governance platform designed to link AI models and agents with business systems through schema-aware connectors, semantic context, access controls, and centralized auditing. The company argues that conventional AI gateways often manage model traffic or tool calls without sufficient understanding of source-system data, definitions, permissions, or business rules, which can reduce accuracy and increase token costs. The gateway resolves user prompts against a context graph, selects and scopes the necessary tools and data sources, routes tasks to suitable models, applies policies at the model, tool, and data levels, and records actions in an audit trail. CData says its approach can reduce unnecessary context and tool calls by performing operations such as joins outside the model, while allowing organizations to use hundreds of MCP servers, third-party servers, and multiple model providers without locking context into a single platform. CData Labs benchmarks cited in the announcement report higher accuracy and substantially lower token use than raw data-access approaches, although the product is currently offered through an early-access program.

CData has announced early access to Connect AI Gateway, a platform designed to govern how AI agents interact with enterprise systems by controlling model routing, tool access, identities, records, actions, costs, and audit trails from a single point. Built on CData’s existing data connectivity layer, the gateway uses source-system schemas, relationships, permissions, and business rules to improve AI accuracy and enforce record-level governance, with the company citing a study in which Connect AI achieved 98.5% accuracy compared with 65% to 75% for other MCP providers. Its architecture combines LLM and MCP gateways, security policies enforced under user identities, a context engine that unifies system metadata, company documentation, and approved institutional knowledge, and live connections to SaaS platforms, databases, warehouses, and on-premises systems. CData argues that direct, governed connections enable agents to retrieve more efficient answer sets, reduce model-token costs, maintain consistent definitions across multiple AI models, and trace outputs or actions back to the associated prompt, model, tools, records, and policies.

Meta Muse, described as a personal AI agent launched in September 2026, can be extended beyond its built-in consumer-focused app connectors by using CData Connect AI as a custom MCP connector for enterprise systems such as Salesforce, HubSpot, NetSuite, ServiceNow, data warehouses, and productivity tools. Users need a configured CData Connect AI data source, the shared MCP endpoint URL, and a personal access token converted into a Basic authentication string, then can add the connector through a Muse chat and authorize it using the Authorization header. Once connected, a single endpoint can provide Muse access to multiple configured sources, enabling plain-language requests such as identifying inactive sales deals, reviewing customer interactions, checking inventory, or finding overdue critical incidents. The post emphasizes that connections use HTTPS and token authentication, credentials are stored separately from chats, tokens can be revoked or rotated, access can be scoped through custom MCP tools, and setup is intended to take about 10 minutes without coding.

LiteLLM is an open-source, self-hosted LLM proxy that provides an OpenAI-compatible interface for more than 100 model providers, with routing, cost tracking, rate limiting, budgets, caching, and broad observability integrations, while enterprise features such as SSO, advanced RBAC, and audit logs require a paid license. CData Connect AI is a managed enterprise AI and MCP gateway focused on live access to hundreds of business data sources, using passthrough identity to preserve each user’s source-system permissions and adding tool-level trust controls, audit capabilities, and a semantic Context Engine for business definitions and cross-source data relationships. The comparison argues that LiteLLM is best suited to teams prioritizing flexible multi-provider model routing and self-managed infrastructure, whereas CData Connect AI targets organizations needing governed agent access to regulated enterprise data, identity-aware permissions, and managed deployment. The platforms can be complementary, with LiteLLM handling model-layer routing and CData handling MCP-based data access, though the text also notes different compliance coverage, operational requirements, and a March 2026 LiteLLM package supply-chain incident.

Model routing and enterprise data access are presented as separate layers of an AI application stack: routing gateways manage LLM selection, fallback, costs, rate limits, caching, and observability, while data-access layers connect agents to live business systems such as Salesforce, Snowflake, SAP HANA, and PostgreSQL. The passage argues that routing has become relatively standardized through OpenAI-compatible APIs and increasingly competitive open-source models, whereas enterprise data access remains difficult because each source has distinct authentication, schemas, business definitions, permissions, and API limits. Model Context Protocol standardizes agent-to-tool communication but does not solve the integration logic within each tool, creating an N-by-M connector burden as organizations add agents and data sources. A dedicated data-access layer is described as providing semantic schema resolution, source-system permission passthrough, per-query audit records, and MCP-native access, helping agents return contextually meaningful results without exceeding users’ access rights. The discussion also links this layer to compliance needs under GDPR Article 30 and the EU AI Act’s logging provisions, since model-routing logs reveal model usage and cost but not what enterprise data an agent accessed. CData Connect AI is positioned as a managed implementation intended to replace many custom integrations with governed connections to enterprise data sources.

CData Connect AI is presented as a managed enterprise AI gateway based on the Model Context Protocol that connects AI agents to production data while controlling authentication, agent-level access scopes, and request-level auditing. Its passthrough identity model forwards each user’s credentials to supported sources such as Salesforce and BigQuery at query time, allowing existing source-system permissions, including sharing rules and row- or column-level security, to determine what an agent can access rather than relying on shared service accounts. Organizations can use the platform for natural-language business-data assistants, scheduled data-quality and anomaly monitors, or a centralized data layer shared by multiple internal agents, with workspaces and toolkits limiting each agent’s available data and tools. The text recommends a gradual deployment approach that begins with read-only access, progresses to human-approved actions, and reserves autonomous writes for low-risk or reversible operations. Connect AI also supports hundreds of connectors, standard MCP-compatible models and frameworks, SIEM-exportable audit logs, and, as of March 2026, SCIM 2.0 synchronization with Okta, Microsoft Entra ID, and Ping Identity alongside custom OAuth application support.

Secure real-time AI can help finance teams use current Salesforce, CRM, transaction, ledger, KYC, AML, and customer-support data for activities such as fraud detection, credit decisions, portfolio updates, and compliance case management, while enforcing permissions, privacy, and regulatory requirements at the moment of each request. Unlike batch analytics, which relies on periodically refreshed data, real-time AI requires identity verification, permission-aware retrieval, data masking, audit logging, and governance controls before information reaches a model. Effective implementation begins by mapping data sources, applicable regulations such as GLBA and GDPR, and accountable data owners, then unifying access through Salesforce Data 360 or external managed MCP platforms such as CData Connect AI. Salesforce’s Einstein Trust Layer supports capabilities including dynamic grounding, prompt protection, masking, toxicity detection, zero data retention, and transaction auditing, though teams must verify which protections apply to particular AI features. AI agents that can initiate workflows or flag transactions require staged testing, human approval for consequential decisions, and production safeguards such as zero-trust access, micro-segmentation, centralized identity management, continuous monitoring, incident playbooks, and periodic control reviews.

Meta Muse, launched in September 2026 as a personal AI agent, can be connected to enterprise systems that are absent from its built-in consumer-focused connector list by using CData Connect AI as a custom MCP connector. Users with a configured Connect AI data source create a personal access token, form an Authorization API key, and provide Muse with CData’s MCP endpoint, allowing one connector to access sources such as Salesforce, HubSpot, NetSuite, ServiceNow, data warehouses, productivity tools, and finance platforms. Once connected, Muse can answer plain-language questions about sales pipelines, customer interactions, inventory, and support incidents, subject to the user’s configured data permissions. The setup is presented as a roughly 10-minute, no-code process, while CData Connect AI offers controls such as HTTPS, token authentication, revocable personal access tokens, optional IP allowlisting, and the ability to limit agent access to selected tables or columns.

CData Connect AI is presented as a managed Model Context Protocol platform for connecting Claude to live Oracle data while preserving existing access controls, auditability, and database-resident processing. The setup requires a reachable Oracle database with read permissions, a Claude plan supporting custom connectors, and a Connect AI account; users configure an Oracle source, create a hosted MCP endpoint, connect it through Claude’s connector settings, authenticate through delegated OAuth or SAML, and test access with read-only or business queries. The service is intended to apply Oracle role-based permissions separately for each user through identity passthrough, avoid shared service accounts, and log requests in both Connect AI and Oracle’s unified audit trail. It also describes support for Claude web, desktop, mobile, Claude Code, and the Anthropic Agent SDK, along with troubleshooting for incorrect connection strings, unknown service names, inaccessible endpoints, disabled connectors, and missing Oracle privileges.

As enterprises add Model Context Protocol (MCP) servers for systems such as Salesforce, Snowflake, and ticketing platforms, independently managed authentication, tool namespaces, permissions, and credentials can create ungoverned server sprawl and limited visibility into agent activity. MCP servers are independent JSON-RPC processes that expose tools, resources, and prompts to LLM clients, and managing them at scale requires centralized routing, identity controls, policy enforcement, and auditability. While direct configurations may suit small local deployments and simple proxies can forward traffic, a gateway provides a centralized control plane for authentication, authorization, routing, logging, session management, and namespaced tools. The guidance recommends inventorying existing servers, consolidating access through a gateway, using OAuth 2.1 and identity passthrough for HTTP-based services, applying role-based and workspace-specific tool exposure, and reviewing tool descriptions for security risks. It also notes that selective tool loading, caching, batching, health checks, retries, circuit breakers, and secure credential storage can reduce context-window costs and improve reliability. Organizations can build and maintain these capabilities themselves or use managed platforms such as CData Connect AI, with the choice depending on engineering capacity, connector requirements, compliance needs, and deployment speed.

An agentic control plane is presented as the enterprise layer that governs AI agents’ identities, policies, tool access, actions, and audit trails, while also providing the often-overlooked connectivity needed to reach systems such as SAP, Salesforce, and NetSuite. The discussion distinguishes this broader function from agent frameworks, which build agent logic, multi-agent models, which coordinate specialized agents, and API gateways or iPaaS tools, which primarily manage individual requests. It identifies connectivity through the Model Context Protocol, role-based access control with OAuth or SAML identity passthrough, multi-agent orchestration, and centralized observability as core architectural components. Organizations are said to need this capability as they deploy multiple agent frameworks, connect agents to multiple back-end systems, face compliance requirements, and seek to reduce credential sprawl. The proposed architecture separates reasoning frameworks, governance and identity controls, connectivity infrastructure, and enterprise systems, with federated requests allowing agents to query several sources under a user’s existing permissions. The text argues that managed connectivity can shorten implementation compared with building enterprise-grade integrations internally and positions CData Connect AI as a service providing managed MCP connections, identity passthrough, policy enforcement, and SIEM-compatible audit logging.

Multi-agent systems using the Model Context Protocol (MCP) allow specialized AI agents to access shared enterprise tools and data through a standardized interface, reducing the need for separate integrations but introducing operational challenges at scale. MCP standardizes tool discovery and invocation rather than agent-to-agent coordination, so production deployments also require orchestration patterns such as centralized supervisors for clearer routing or direct swarms for lower latency. Reliable deployments depend on centralized identity and authorization using OAuth 2.1, identity passthrough, role-based access controls, distributed tracing, version management, curated tool catalogs, and audit trails. Common production failures include tool sprawl, token bloat from excessive context and data retrieval, permission drift, fragile upstream APIs, retry cascades, and agent loops; recommended mitigations include gateways, query pushdown, recursion guards, fault isolation, and managed connectors. The discussion contrasts building these capabilities internally, which requires sustained identity, connector, observability, and protocol-maintenance expertise, with using managed platforms such as CData Connect AI, whose performance and efficiency claims are presented as vendor benchmarks.

Connecting on-premises data to cloud AI assistants such as Claude and ChatGPT requires overcoming network isolation, legacy-system limitations, and regulatory controls while ensuring that access, credentials, and activity remain governed. The Model Context Protocol (MCP) provides a standardized architecture in which an AI host communicates through an MCP client and server, with connectors translating approved requests to secured databases, APIs, and business systems without directly exposing those systems. Recommended implementation steps include classifying data, selecting an appropriate deployment model such as private links, VPNs, reverse proxies, or hybrid hosting, building narrowly scoped connectors, and enforcing short-lived credentials, role-based permissions, encryption, monitoring, DLP, and detailed audit logging. Organizations should also test retrieval accuracy and data redaction, establish approval, retention, and incident-response policies, and align controls with frameworks such as SOC 2, ISO 27001, and GDPR. Examples from TELUS and Bridgewater illustrate managed-cloud deployments of Claude, while CData Connect AI is presented as a managed MCP-based option that centralizes authentication, query translation, source-level permissions, and auditing for on-premises integrations.

Enterprise AI agents can produce confident but inaccurate answers when connectors provide only partial access to source systems, as illustrated by Workday compensation data that is unavailable through REST but accessible through WQL and SOAP services. The passage argues that reliable enterprise connectivity requires more than endpoint coverage, encompassing access to all relevant APIs, custom data models, read-and-write operations, source-side query execution, system-specific context, and ongoing maintenance as APIs evolve. Query pushdown is presented as essential for reducing token use and improving accuracy by performing filtering, joins, and aggregations in the source system rather than in an LLM’s context window, while secure write-back requires user-level permissions and audit trails. It positions CData Connect AI, a managed MCP server, as a solution offering broad API access for systems such as Workday, NetSuite, and Salesforce, dynamic schema discovery, source-executed queries, read/write actions, per-user credentials, logging, and managed connector updates. The piece cites public and internal benchmarks to support claims that deeper connectivity and source-level execution improve agent accuracy and efficiency compared with shallow connectors and model-routing gateways.

Enterprise AI token costs may rise despite falling unit prices because agentic workflows use more tools, data sources, reasoning steps, and repeated context, with Gartner predicting inference costs per agentic workflow could increase more than fivefold by 2028. The passage argues that much of this expense comes from tool definitions, schemas, and large raw query results being loaded into model context, especially when agents must discover and combine data across multiple systems. It distinguishes these issues from the MCP protocol itself, attributing them to how applications expose tools and insert results into context. While dashboards, rate limits, model routing, caching, and prompt compression can control or reduce spending, the proposed approach is to remove unnecessary context before it reaches a model through a data-access layer that filters, joins, scopes, and caches enterprise data. CData promotes its Connect AI platform as such a managed MCP gateway, citing a benchmark in which a custom multi-source tool reduced usage from 183,541 tokens across 22 calls to 4,427 tokens in one call, a claimed 97.6% reduction, while noting that actual savings depend on workflow design, models, usage, and data volume.

AI governance is presented as a framework of policies, controls, and monitoring that helps organizations deploy AI securely, compliantly, and reliably while addressing risks such as data leakage, inconsistent definitions, weak access controls, limited auditability, and unreliable model outputs. Citing 2025 surveys reporting that privacy and security concerns caused many enterprises to abandon or delay AI initiatives, the discussion argues that governance is necessary to move projects beyond pilots and establish trust in production systems. A semantic layer is described as the operational foundation for this governance because it standardizes data definitions, applies access policies, preserves lineage, and supplies AI systems with contextual, compliant data that can reduce hallucinations. The CData Platform is positioned as a tool for creating this governed data layer across cloud, on-premises, and hybrid environments, offering controls for data residency, auditing, transparency, and fine-grained permissions. Its Talk-to-Your-Data interface is presented as an example of allowing employees to query governed enterprise data in natural language while applying security rules and maintaining traceability.

Unstructured data, which includes emails, social media posts, multimedia, documents, and IoT sensor output, represents an estimated 80–90% of business-generated information and lacks the predefined formats of traditional database tables. Its flexibility, diversity, and capacity for detailed contextual insights can help organizations identify customer sentiment, behavioral trends, operational patterns, and other information not readily visible in structured datasets. However, its volume, storage requirements, inconsistent formats, and need for preprocessing create scalability, management, and usability challenges. Semi-structured formats such as JSON and XML offer an intermediate level of organization, while tools including CData Sync, MongoDB, Microsoft Azure, Apache Hadoop, and Elasticsearch support the integration, storage, processing, search, and analysis of unstructured data. Techniques such as OCR, natural-language processing, transcription, machine learning, and distributed computing can help transform raw content into usable business intelligence.

Model Context Protocol (MCP) is presented as an open standard for giving AI assistants live, permission-aware access to enterprise systems such as support platforms, HR libraries, documentation repositories, contracts, and data warehouses, rather than relying on periodically synced copies of information. Its host-client-server architecture supports isolated connections to individual data sources, enabling organizations to apply source-system permissions, passthrough user identity via SSO mechanisms such as OAuth, SAML, and OpenID Connect, and query-level audit logging. The text contrasts this approach with traditional knowledge management repositories, which can provide outdated answers or retain access after permissions change, while noting that MCP complements retrieval-augmented generation by standardizing live access to the systems from which RAG retrieves context. It argues that MCP is especially useful for mid-sized and large organizations with multiple governed data sources, established identity infrastructure, compliance requirements, and existing AI deployments, while simpler static retrieval may suffice for smaller environments. CData Connect AI is described as a managed MCP platform intended to centralize connector maintenance, access scoping, identity controls, and SIEM-exportable auditing across hundreds of enterprise sources.

Secure enterprise AI access to systems such as Salesforce or SAP requires governance controls that preserve each user’s existing permissions, protect data in transit, and create auditable records of every query. The guidance recommends a zero-trust approach built on federated SSO, OAuth or SAML identity passthrough, source-level RBAC, least-privilege access, scoped tools, workspace isolation, and SIEM-integrated per-query logging, rather than shared service accounts, static API keys, or duplicated permission models. It warns that ChatGPT and Claude can connect to unverified MCP servers, creating risks of data exfiltration, and argues that organizations should use a governed MCP endpoint that authenticates users individually and enforces permissions at the source. It also outlines administrative requirements for enabling custom MCP connections in ChatGPT and Claude, troubleshooting common OAuth and network issues, and aligning controls with standards including NIST Zero Trust, NIST’s Generative AI Profile, SOC 2, and ISO/IEC 27001. The source presents CData Connect AI as a managed, audited MCP layer intended to centralize these controls across enterprise data systems.

Programmatic replication management uses code, policy-as-code, and automation to coordinate data replication and operational controls across AWS, Azure, Google Cloud, and on-premises systems, reducing configuration drift, manual effort, and recovery risks. Successful implementations begin by defining recovery point and recovery time objectives, throughput, cost limits, and workload classifications, then selecting centralized or federated control-plane designs and appropriate synchronous, asynchronous, CDC, or batch replication modes. Infrastructure-as-code tools such as Terraform and Pulumi, combined with policy engines like Open Policy Agent and Sentinel, can standardize provisioning, security, compliance, and auditing. Automated lifecycle processes should address scaling, credential rotation, configuration remediation, and failover, while centralized observability and FinOps monitoring track replication health, latency, integrity, and costs. The guidance also emphasizes regular failover and disaster-recovery testing, consistent IAM, encryption, data residency, and audit controls across clouds, and scalable practices such as stateless connectors, separated control and agent scaling, resource tagging, and incremental expansion. CData Sync is presented as a platform offering connectors, CDC, scheduling, RBAC, audit logs, encrypted movement, and hybrid deployment capabilities for governed multi-cloud replication.

A benchmark of 22 AI models across 1,034 runs using live CRM, warehouse, and ITSM data found that moving business logic, data boundaries, and action controls into a governed data layer substantially reduced differences in model accuracy while leaving large cost differences. In composite read tasks, 17 models achieved identical perfect results through curated custom tools, including lower-cost models, while cost per correct answer differed by up to 178 times; a second read task showed a similar pattern. On raw data access, models frequently failed to infer required business logic, and no model completed all write-task runs safely, with some creating large numbers of unauthorized records. Restricting accessible tables and columns improved results but did not fully prevent incorrect actions, whereas server-side validation embedded in custom tools eliminated unauthorized writes across all tested models. The findings argue that governed data-layer tooling can make model selection primarily a cost and compatibility decision for these tasks, while emphasizing that write operations require validation beyond simple access scoping.

Model Context Protocol (MCP) standardizes how AI applications connect to SAP data and actions, reducing the need for many point-to-point integrations by separating AI hosts and clients from MCP servers that expose SAP OData entity sets as controlled tools. An SAP MCP server can often be built with minimal or no custom code using Node.js and odata-mcp-proxy, with JSON configuration defining destinations, API paths, CSRF requirements, and entity sets that generate CRUD tools automatically. Secure deployment requires OAuth scopes, role-based access controls, tenant and audit alignment, and service-specific CSRF settings, while SAP BTP provides Destination, XSUAA, and Connectivity services for credentials, authentication, and on-premises access. Servers should be tested locally and in BTP, with MCP clients using supported transports such as stdio or Streamable HTTP to enable live SAP queries and multi-step workflows. MCP is most appropriate for real-time business questions, inventory or order checks, and governed cross-system agent workflows, whereas replication or semantic search may better suit static or batch-oriented data. Ongoing management includes consolidating related integrations where practical, maintaining simple version-controlled configurations, auditing permissions, and isolating failures across configuration, authentication, and client connection layers. Organizations needing less customization can instead use managed platforms such as CData Connect AI to provide centrally governed AI access to SAP data without operating the underlying MCP infrastructure.

CData’s Episode 11 of “Vibe Querying with MCP” demonstrates how its Google Drive MCP Server connects Claude to unstructured business documents, allowing users to retrieve, analyze, create, and update files through natural-language prompts. Using OAuth-based access and automatically discovered procedures such as ListFiles, GetDocumentContent, UploadFile, and UpdateResource, the setup avoids direct API knowledge, scripts, or manual file handling. The episode illustrates two workflows: converting a technical CData Sync whitepaper into an executive-ready summary saved as a new Google Drive document, and analyzing a sales-call transcript to identify customer pain points, priorities, recommended follow-up actions, and then append the analysis to the original file. The approach, described as “vibe querying,” is presented as a way to make document knowledge more accessible, reduce time spent on manual review, support iterative analysis, and help teams turn static files into interactive sources of business intelligence.

The Model Context Protocol (MCP) is an open-source standard that connects AI tools and agents with enterprise data, and Stripe offers two contrasting MCP approaches for accessing payment information. Stripe’s native MCP Server maps individual REST API endpoints to dedicated tools for actions such as listing customers, creating refunds, managing subscriptions, and searching documentation, providing a direct API-first model. CData’s MCP Server for Stripe instead presents live Stripe API data as relational tables, views, and procedures accessible through SQL, enabling filtering, joins, aggregation, and data modification through a smaller set of generic tools. While the native server may require agents to make multiple API calls, paginate results, and combine JSON responses client-side for complex analysis, CData can execute a single SQL query across customers, subscriptions, invoices, and line items to return aggregated results. The SQL-based approach is positioned as useful for AI-driven analytics, refund workflows, automated onboarding, cross-platform financial reporting, and revenue forecasting, while both options allow AI agents to access and act on Stripe data.

“Vibe Querying with MCP” is a content series demonstrating how business users can use conversational AI to explore real-time sales and marketing data through CData MCP Servers connected to tools such as Salesforce, Google Analytics, and Claude. The first episode examines relationships between website activity and closed business deals, using natural-language questions to retrieve recent deal totals, isolate sales for a specific product line, compare new customers with expansions by industry, and connect Salesforce opportunity data with web-traffic behavior. The demonstration highlights cross-object Salesforce analysis and the ability to associate leads and contacts with Google Analytics metrics, producing findings about traffic channels, visitor engagement, conversion behavior, and differences between technology and professional-services prospects. It concludes by identifying the most visited URLs among high-value technology customers, presenting MCP-based querying as a faster alternative to traditional cross-platform reporting that often requires specialized data, marketing-operations, and technical expertise.

CData MCP Servers are Windows-based local services that wrap CData drivers to give AI clients such as Claude Desktop standardized Model Context Protocol access to enterprise data sources. They provide consistent tools for discovering schemas, tables, columns, procedures, and parameters, then querying data through SQL or executing source-specific stored procedures without custom integrations for each system. The servers dynamically retrieve metadata according to active user permissions, while optionally using caching to improve responsiveness, and translate diverse systems such as CRMs and ERPs into a common tabular SQL model. Performance features including query pushdown, parallel paging, and streaming support interactive workloads with large or complex datasets. Security is based on in-place data access, authenticated user permissions, centralized credential configuration, and support for OAuth SSO, Kerberos, and basic authentication, alongside SOC 2 Type II, ISO 27001, and GDPR compliance. CData initially offers the servers as a free beta for Windows and provides logging for administration and diagnostics, as well as an open-source read-only MCP server built on its JDBC drivers for developers to adapt.

MCP (Model Context Protocol) servers are presented as standardized integration layers that connect AI applications with enterprise data, tools, and business workflows through secure, context-aware interfaces, commonly using JSON-RPC 2.0. Unlike traditional application-specific APIs, they aim to reduce custom integration work while supporting real-time data access, semantic metadata, centralized management, multi-AI compatibility, and governance features such as OAuth, role-based access controls, encryption, auditability, and network protections. An MCP ecosystem includes a host AI application, a client that translates requests, and a server that exposes approved enterprise resources; deployments can use frameworks such as FastMCP, Quarkus, Docker, or Cloudflare Workers depending on language, scale, latency, and hosting requirements. Recommended deployment practices include defining discovery, invocation, and status endpoints, choosing on-premises, cloud, containerized, or edge hosting, and maintaining centralized logging, health checks, metrics, alerts, and vulnerability monitoring. Common uses include AI-driven workflow automation, secure agent access to business tools, cloud API integration, and shared enterprise AI workflows, while CData Connect AI is promoted as a managed platform for connecting MCP servers to more than 350 data sources.

Model Context Protocol (MCP) is an open-source specification introduced by Anthropic that standardizes secure connections between AI applications, particularly large language models, and enterprise data sources, tools, and systems. Rather than improving a model’s inherent reasoning or creating business context itself, MCP provides a consistent runtime interface through which models can access authorized, relevant information such as customer records, policies, business rules, metadata, and system status. This can reduce the need for custom integrations, make context reusable across applications, and give data teams clearer control over what information is exposed and when. MCP can help make AI outputs more grounded in current business operations and support uses such as automation, reporting, and decision support, but it remains an early-stage standard and does not replace sound data architecture, governance, curated data, human oversight, or effective model development.

CData’s third Vibe Querying episode demonstrates how Senior Sales Director Jared Feeney uses the beta Salesforce MCP Server with Claude to analyze live Salesforce data through natural-language questions. Model Context Protocol connects AI clients to external data sources, and CData’s servers extend this access to more than 350 connectors, allowing sales leaders to explore deal portfolios without detailed schema knowledge or manually built reports. In the demonstration, Claude identified promising and at-risk North American opportunities, analyzed blockers using structured and unstructured Salesforce data, produced an individual deal-risk presentation, and created an interactive dashboard of U.S. closed-lost deals by product line. The post argues that these workflows can reduce report-building, research, and presentation preparation from hours or days to minutes while supporting forecasting, coaching, executive communication, and more timely decision-making. It also notes that insight quality depends on accurate Salesforce data, that the system can incorporate notes and emails beyond standard reporting fields, and that MCP servers can potentially write updates back to Salesforce.

AI systems that independently access enterprise sources such as CRM platforms, email, code repositories, and data warehouses require auditable records showing who queried what data, which systems were reached, what was returned, and when. The material argues that query-level logging at the data connectivity layer, rather than relying on model outputs or application logs, can support EU AI Act, SOC 2, ISO 27001, and GDPR-related compliance needs, particularly for high-risk AI systems. It recommends using OAuth and SAML passthrough authentication to link actions to real end users, workspace isolation to enforce least privilege, and SIEM exports combined with documented review and incident-response processes. It also emphasizes maintaining inventories of AI tools and approved data systems, retaining logs appropriately, capturing failed as well as successful requests, and monitoring for shadow AI, credential sprawl, and stale permissions. CData Connect AI is presented as a governed Model Context Protocol connectivity platform designed to centralize these controls by executing, attributing, logging, and exporting AI data-access events.

An anecdote about an AI-generated backpacking plan that understated route difficulty illustrates the risks of relying on models fed stale, incomplete, or disconnected information. The piece argues that as software companies add AI capabilities, trustworthy outputs depend less on the model itself than on access to accurate, current data from systems of record. It presents CData as an embedded data-connectivity provider used by major enterprise software companies, offering hundreds of maintained connectors to databases, SaaS platforms, ERP and CRM systems, and data warehouses. According to the company, outsourcing connectivity can reduce development and maintenance burdens, accelerate time to market, prevent data degradation caused by changing APIs, and enable AI features to provide more reliable answers.

MCP core maintainers’ August 22, 2026 roadmap identifies five future priorities with no stated release dates: agentic messaging, HTTP-native transport unification, agent identity and enterprise security, improved protocol primitives, and SDK developer experience. The analysis characterizes agent identity as the most urgent area because planned adoption of DPoP, workload identity federation, and token exchange would strengthen machine-to-machine authentication and reduce reliance on reusable bearer tokens, while noting that these measures address only the agent-to-server boundary. Enterprises must still independently manage credentials for downstream systems, enforce source-level user permissions, and maintain query audit logging. Transport work would extend Streamable HTTP to local environments and further replace deprecated SSE and stdio-based patterns, while improved primitives would standardize tool-result behavior and support progressive discovery for large tool catalogs. Agentic messaging is expected to introduce standardized server-initiated events, webhooks, and channels, whereas SDK improvements focus on documentation, ergonomics, and conformance testing. The roadmap follows features already shipped in the July 28, 2026 release, including stateless operation, discovery, caching, tasks, multi-round-trip requests, and enterprise authorization integration.

ChatGPT Enterprise connectors let users access external systems within existing permissions, with native integrations available for selected applications and custom Model Context Protocol connectors extending access to organization-specific tools and data. Secure deployment depends on governance measures including least-privilege access, per-user OAuth or SAML identity passthrough, source-system RBAC enforcement, narrowly scoped credentials, administrative controls over developer mode and write actions, and audit logging routed to SIEM platforms through tools such as the Compliance API. The guide argues that native connectors may be insufficient for data held in systems such as Salesforce, SAP, NetSuite, and ServiceNow, and presents CData Connect AI as a governed MCP layer that can provide centralized access across hundreds of sources while preserving user-level permissions and logging query activity. It also highlights risks including broad connector scopes, shared accounts, prompt injection, hallucinations, shadow IT, and permission creep, recommending testing, semantic data scoping, human approval for sensitive actions, and regular access reviews before scaling connected AI workflows.

Authorization is optional in Model Context Protocol (MCP), creating security risks when AI agents access enterprise data through broad permissions, shared credentials, or inadequately scoped tools. The guide argues that conventional access controls designed for human users are insufficient for agents, which can chain actions across systems and expose information through tool calls, APIs, or outputs. It recommends enforcing field-, row-, and table-level permissions at query time through identity passthrough, where source systems apply each requesting user’s existing OAuth or SAML entitlements rather than relying on service accounts or forwarding raw tokens, a practice MCP prohibits. Least-privilege controls should also isolate workspaces, limit available tools and operations, use short-lived scoped credentials, validate token audiences, and authorize every request independently. Detailed per-query audit logs linking user identity, actions, systems accessed, returned data, and timestamps are presented as important for incident response and compliance with ISO 27001, SOC 2, and GDPR. CData Connect AI is described as a platform implementing these practices through identity passthrough, workspace and toolkit boundaries, source-system enforcement, and audit logging.

Governed LLM data access ensures that AI models and agents retrieve enterprise information under each requesting user’s identity, enforce source-system permissions before retrieval, and create auditable records, reducing risks associated with shared service accounts, excessive access, data duplication, and compliance failures. The comparison evaluates 10 tools across identity passthrough, inherited role-based access control, audit logging, workspace isolation, and security certifications, grouping them into managed MCP platforms, AI gateways, integration platforms, and related identity, security, catalog, warehouse, and cloud services. CData Connect AI and Databricks Unity Gateway are presented as options emphasizing user-level permission inheritance and in-place querying, while Kong, MuleSoft, Merge, FastMCP, Okta, Wiz, Collibra, and Amazon Bedrock AgentCore address different aspects such as traffic policy, integration, framework development, identity issuance, monitoring, metadata governance, or cloud-native agent connectivity. The discussion advises organizations to verify OAuth and SSO behavior, source-level authorization, SIEM-ready audit trails, certification scope, and GDPR DPIA requirements, particularly because MCP itself does not independently enforce authorization and self-hosted solutions leave ongoing security and audit responsibilities with internal teams.

AI agents increasingly access enterprise systems and sensitive data, making them non-human identities that require unique credentials, least-privilege role-based access control (RBAC), traceable actions, and regular reviews rather than shared API keys. A recommended implementation process begins by inventorying all models, agents, data sources, and credentials; classifying data by sensitivity and regulatory requirements; defining minimal, time-bound roles; and enforcing permissions through a centralized AI gateway integrated with existing identity systems. Attribute-based access control (ABAC) can supplement roles by considering context such as device, time, location, and task, enabling just-in-time elevation for sensitive operations. Centralized logging, SIEM integration, anomaly detection, immutable retention, and periodic recertification support compliance, incident response, and ongoing governance, while common pitfalls include over-permissioning, incomplete audit trails, lack of tool-level controls, and neglected role reviews. CData Connect AI is presented as a managed gateway that applies identity passthrough, RBAC and ABAC policies, workspace and toolkit restrictions, and query-level auditing across enterprise data sources.

AI gateway performance should be evaluated by business outcomes rather than adoption or usage alone, using pre-deployment baselines and a three-layer framework covering infrastructure, data access, and organizational results. Infrastructure measures such as request volume, latency, token consumption, cost per request, cache performance, and provider availability reveal operational efficiency and potential issues with routing or data retrieval. Data access measures, including governed query volume, permission denials, audit-log completeness, schema exposure, and self-service rates, support compliance while helping diagnose whether failures arise from models, permissions, or underlying data. Business metrics such as time-to-answer, analyst hours recovered, first-contact resolution, cost per resolved inquiry, and attributable revenue connect AI investments to established financial measures, although returns may emerge over several years. A business case for a gateway can compare centralized integration with the costs and risks of custom connectors, account for token savings from schema-aware access, and estimate avoided governance costs. CData presents its Connect AI platform as providing built-in audit, token, access, and permission observability, and cites its own testing and benchmarks to support claims of improved token efficiency and grounded accuracy.

Model Context Protocol enables AI agents to connect with enterprise tools and live data but does not natively provide identity management, credential protection, access controls, or auditing, creating security barriers to enterprise adoption. The discussion identifies credential exposure, inadequate audit trails, permission drift, shadow AI, and prompt injection as central risks, citing research that found many MCP servers use hard-coded secrets and incidents involving malicious packages and destructive agent actions. It presents CData Connect AI as a managed MCP platform that addresses these gaps through identity passthrough from enterprise identity providers, OAuth 2.1 with PKCE, SSO-based authentication, per-request RBAC down-scoping, workspace isolation, and default-deny tool access. The platform also records attributed, query-level logs detailing users, queries, accessed systems, returned data, and timestamps, with SIEM export intended to support compliance and incident response. The proposed enterprise baseline emphasizes short-lived, user-bound authorization rather than shared accounts, centralized governance, and independently validated controls aligned with SOC 2 Type II, ISO/IEC 27001, GDPR, CCPA, and the Enterprise-Managed Authorization extension for MCP.

Enterprise AI delivers the greatest value when it can securely access and act on live business data rather than merely generate text, but adoption is often limited by data connectivity, context, and governance challenges. An AI gateway with a shared data layer can support natural-language business intelligence, conversational CRM and ERP copilots, cross-system enterprise search, regulated-industry applications, and automated actions in systems such as Salesforce, NetSuite, ServiceNow, and EHR platforms. Effective deployments require permission enforcement at the user, row, column, and record levels; comprehensive audit logging; schema-aware retrieval to limit irrelevant context and token costs; and real-time read and write access to source systems rather than stale warehouse copies. CData positions its Connect AI platform as an MCP-compliant enterprise data layer that connects AI tools and frameworks, including ChatGPT, Claude, LangChain, and CrewAI, to hundreds of cloud and on-premises sources while centralizing authentication, governance, auditing, and source-specific access controls.

AI agent data governance comprises the policies, technical controls, and monitoring practices that determine which enterprise data autonomous agents may access, what actions they can take, and how those actions are attributed to individual users. Unlike traditional governance focused largely on data access and movement, it must manage runtime behavior because agents can invoke tools, chain tasks, and potentially modify live systems, creating risks such as credential exposure, excessive privileges, incomplete audit trails, and uncontrolled expansion of access. Recommended foundations include source-level ownership, classification, lineage, RBAC, retention policies, and distinct agent identities, while four core controls are OAuth or SAML identity passthrough, permission enforcement under end-user roles, least-privilege workspace isolation, and detailed per-query logging exported to SIEM platforms. The approach aligns with frameworks including NIST AI RMF, SOC 2, ISO/IEC 42001, GDPR, and the EU AI Act, and emphasizes testing access boundaries, resource-bound tokens, token lifecycle management, and behavioral monitoring before moving pilots into production. It also argues that a centralized, platform-neutral connectivity and enforcement layer, such as CData Connect AI, can reduce duplicated governance work across multiple agent platforms while complementing their native security controls.

A context layer is presented as infrastructure that supplies AI applications with live, accurate, permission-filtered data from enterprise systems such as CRM, ERP, HRIS, warehouses, and ticketing platforms, positioned between AI applications and underlying data sources. Unlike retrieval-augmented generation, which is suited to unstructured documents indexed periodically, a context layer is intended for current, structured operational data that may require cross-system joins, semantic interpretation, and transaction-level access controls; the two approaches are described as complementary in enterprise deployments. The proposed evaluation framework centers on source coverage, including custom fields and on-premises connectivity; schema intelligence and semantic resolution for identifying relevant fields, business definitions, and relationships; enforcement of each requester’s source-system permissions with auditable query logs; and measured outcomes such as grounded accuracy, token efficiency, and latency. The piece argues that enforcing permissions at the data source is essential for security and compliance, while noting that organizations should request benchmarks reflecting their own workloads. It concludes by describing CData Connect AI as a product designed around these capabilities and cites company testing and external research to illustrate the market’s growing focus on governed, hybrid retrieval for agentic AI.

AI routing gateways effectively manage model selection, costs, rate limits, identity propagation, auditing, and request-layer security, but they cannot ensure that models receive accurate, current, or authorized enterprise data. The proposed context layer complements routing by connecting AI to live systems of record, applying user-level permissions before data reaches the model, and supplying semantic definitions that explain the business meaning of raw fields and relationships. Without this layer, models may rely on incomplete training data or stale retrieval indexes, resulting in hallucinations, incorrect interpretations, and possible unauthorized disclosures. Enterprise context consists of live operational data, shared semantic definitions, and harder-to-formalize tacit organizational knowledge, with the first two viewed as essential current capabilities. Effective AI infrastructure therefore requires both traffic governance through routing and data governance through a context engine, which can also reduce token costs by returning only relevant, permission-filtered fields and records. Organizations evaluating or building such a layer should consider live source access, delegated authorization, semantic resolution, source coverage, custom and on-premises system support, and detailed query logging.