October 2026 Summaries
48 posts from CData
Filter
Month:
Year:
Post Summaries
Back to Blog
MCP gateways are presented as a central security layer for AI agents using the Model Context Protocol, addressing gaps in MCP’s optional authorization model by enforcing authentication, identity mapping, tool-level access controls, protocol validation, policy enforcement, and audit logging. The discussion highlights risks including prompt and command injection, tool poisoning, supply-chain compromise, credential misuse, and data exfiltration, citing examples involving GitHub, Jira, and MCP SDK vulnerabilities. Recommended practices include short-lived scoped credentials, OAuth 2.1, OIDC, mutual TLS, source-level permissions, artifact signing, version pinning, rate limiting, PII masking, centralized logging, SIEM integration, and tested incident-response procedures. Organizations are advised to use a centralized gateway as a control plane, maintain inventories of approved clients, servers, and tools, automate predeployment validation and runtime monitoring, and align controls with frameworks such as SOC 2, ISO 27001, and GDPR. The text compares gateway approaches and promotes CData Connect AI Gateway as a managed data-layer option that combines source permissions, credential vaulting, policy controls, and audit capabilities.
Oct 30, 2026
2,126 words in the original blog post.
CData describes how its AI gateway connects agents to governed enterprise data across systems such as Salesforce, BigQuery, Stripe, Gong, Tableau, and marketing platforms, reducing manual exports, spreadsheet reconciliation, and delayed reporting. At its Foundations 2026 event on November 5, CData’s demand generation, finance, and business development teams will demonstrate internal applications including automated funnel reporting and ICP analysis, parallelized month-end financial analysis, and AI-based scoring of sales calls and lead intent. The company emphasizes that agents can also take actions such as updating Salesforce records, sending emails, and adjusting account tiers while using each requester’s existing permissions and preserving an audit trail. CData says its gateway is model-neutral, allowing organizations to retain business logic, data definitions, and approved context even if they adopt different AI models or hosts, and the event will cover governance, context management, implementation examples, and research on enterprise AI agents.
Oct 29, 2026
962 words in the original blog post.
CData Sync 26.4 introduces security, scheduling, AI-access, CDC, and loading enhancements aimed at improving control over automated data pipelines. The release adds encrypted local vaults and environment-variable support so jobs can reference secrets by name rather than storing credentials directly, while new multi-rule schedules and recurring ReSync options allow pipelines to adjust frequency and full reloads based on operational needs. A read-only Model Context Protocol server enables AI assistants to inspect Sync resources, logs, and health status under the signed-in user’s permissions, with sensitive values masked. CDC support expands for SQL Server, DB2 for i, and SAP Gateway, including transaction-level consolidation of SQL Server changes so delete-and-reinsert operations are delivered as updates, improved schema alignment and checkpoint handling, and member-specific DB2 for i capture. The version also provides direct-insert and bulk-load paths for eligible initial loads into empty targets, including MySQL and Apache Doris, and is available for existing customers and new trial users.
Oct 29, 2026
679 words in the original blog post.
Data connectors are reusable components that move data between sources and destinations in ETL and ELT pipelines, covering databases, APIs, files, SaaS applications, and reverse ETL targets. Their primary maintenance challenges are source API changes, schema drift, and authentication updates, which can produce either visible failures or silent data-quality problems such as missing fields and inaccurate dashboards. Effective connector management therefore requires monitoring beyond job status, including freshness, row volumes, schema changes, staging environments, contract tests, canary runs, and post-deployment validation. Organizations must weigh building connectors internally against using managed services based on source volatility, engineering capacity, compliance needs, and the stability of source interfaces; a hybrid strategy may use managed connectors for frequently changing third-party SaaS platforms while retaining custom integrations for stable internal systems. CData Sync is presented as a managed option offering pre-built connectors and vendor-managed updates for schema, authentication, and API compatibility.
Oct 29, 2026
1,270 words in the original blog post.
How MCP Gateways Serve Context to Probabilistic Agents: Correct Answers in a Tenth of the Tool Calls
CData Labs’ benchmark argues that AI budgets should be assessed by cost per correct task rather than token consumption, incorporating both model spending and reliability. Across 810 runs involving 27 models from 12 developers completing account-health tasks spanning CRM, data warehouse, and ITSM systems, CData reported that its Connect AI Gateway achieved 83% task accuracy at $0.26 per correct task, compared with 36% accuracy and $5.67 per correct task for an open-source MCP gateway. The study attributes the difference to the AI gateway’s ability to retrieve relevant context on demand and perform cross-system joins, filtering, calculations, and ranking in a single query, whereas the pass-through gateway required models to retrieve paginated data separately and perform computation in their reasoning. CData also reported fewer tool calls, tokens, confidently incorrect read results, and unintended writes through its gateway, particularly enabling lower-cost models to complete tasks more reliably. The company provides the benchmark harness for reproduction and recommends evaluating MCP gateways based on cost per correct task, where cross-system computation occurs, and how business context is delivered.
Oct 29, 2026
2,037 words in the original blog post.
AI data governance establishes policies and controls for determining what enterprise data AI systems and agents can access, what actions they may perform, and how each interaction is linked to an authorized individual. It differs from traditional governance by requiring real-time permission checks, identity-aware access, and detailed audit trails for live AI requests, particularly as regulations such as the EU AI Act and privacy laws increase oversight. Enterprise-ready governance tools should support least-privilege access through existing IAM systems, OAuth or SAML identity passthrough, SIEM-exportable audit logs, workspace isolation, and certifications whose scope covers the relevant product. The overview identifies 11 tools spanning governed connectivity, data cataloging, identity security, lineage, sensitive-data discovery, AI gateways, monitoring, compliance automation, data quality, platform governance, and data masking. It positions CData Connect AI as a connectivity-focused option that applies per-user permissions and centralized controls across AI clients, while emphasizing that organizations should evaluate tools against their existing systems, compliance requirements, and gaps in areas such as lineage, privacy, quality, and monitoring.
Oct 28, 2026
1,979 words in the original blog post.
Model Context Protocol servers expose tools, resources, and prompts that AI agents use to access business systems, while MCP gateways act as centralized reverse proxies that route requests across servers and can manage authentication, authorization, rate limits, policies, monitoring, and audit records. Direct server connections generally minimize latency and operational layers but require each deployment to handle its own security and observability, whereas gateways support multi-server, multi-team, and regulated environments through unified governance, scoped tool access, load balancing, and consolidated reporting, while adding processing overhead and potential single-point-of-failure risks if not deployed redundantly. Organizations should select direct connections for small, independently managed deployments and consider gateway-first or hybrid architectures when shared access, compliance requirements, or server growth require centralized control; hybrid models can place sensitive or shared tools behind a gateway while retaining direct access for selected specialized workflows. CData Connect AI Gateway is presented as a managed platform combining MCP server and gateway functions, with connections to enterprise data sources and third-party servers, centralized SSO, permissions, data-security rules, and request-level auditing.
Oct 27, 2026
1,588 words in the original blog post.
Anaqua, an intellectual property management software provider handling assets worth hundreds of millions of dollars, has adopted a tightly governed approach to AI because its rules-based workflows require accuracy, consistency, and careful limits on write access. CIO Erik Bailey selected CData Connect AI as a centralized control plane for managing AI connectivity, permissions, and context across business systems, ensuring that queries run under each logged-in user’s identity while limiting unauthorized changes and preventing shadow AI or uncontrolled MCP connector use. After beginning with Microsoft Copilot and expanding through an early 2026 Claude pilot, Anaqua made Claude available company-wide and is developing use cases such as Customer 360 views and automated improvement of its AI Skills based on query failures. Bailey will discuss Anaqua’s transition from AI pilots to production, its governance practices, the AI gateway category, and future plans during a November 5, 2026 Foundations 2026 session with CData CMO Will Davis, aimed at CIOs and AI leaders responsible for enterprise-wide AI adoption.
Oct 27, 2026
717 words in the original blog post.
CData Arc’s Q4 2026 release expands AI-assisted integration operations through its Model Context Protocol server, allowing authorized users to run connectors, requeue failed messages, clean up files, access deeper logs, and receive on-demand explanations and recommendations for failed transactions and worker-pool performance, while requiring confirmation for actions and preserving audit records. New connectors include Partner Mailbox for browser-based partner file exchange, Microsoft 365 Email using Microsoft Graph for shared mailboxes and advanced message handling, and Finalize for confirming successful delivery before source-file cleanup. The update also adds batch-group processing, deferred file finalization, external secrets from Azure Key Vault or environment variables, workspace-specific Git repositories, improved automation and database-health visibility, expanded Admin API capabilities, more detailed audit logging, and numerous mapping, security, certificate, file-transfer, and connector enhancements. It removes several classic interfaces, restricts Version Control to administrators, and requires attention to changed v26 driver defaults during upgrades.
Oct 27, 2026
1,972 words in the original blog post.
Model Context Protocol servers connect AI assistants such as ChatGPT and Meta Muse to business systems, enabling them to retrieve information or, when permitted, perform actions in tools such as CRM platforms, support desks, finance systems, data warehouses, and project-management applications. The appropriate server depends on the task: built-in connectors suit common email, file, and calendar tools; vendor-hosted servers such as Salesforce, HubSpot, Atlassian, Stripe, and Dynamics 365 fit work within a single application; Zapier and Workato support cross-application automation; and Snowflake, BigQuery, and Databricks servers provide access to data already centralized in warehouses. The guide emphasizes governance considerations including signed-in-user permissions, read-only restrictions, audit logs, cross-system querying, and setup complexity, noting that these controls generally reside with the server rather than the assistant. It identifies CData Connect AI as a governed gateway for querying hundreds of SaaS, database, ERP, and on-premises sources without copying data, with per-connection access controls and logging across ChatGPT, ChatGPT dots, and Muse, while also noting that simpler built-in, vendor-specific, warehouse, or automation options may be preferable for narrower needs. ChatGPT supports MCP servers as plugins under administrator controls, whereas Muse uses custom connectors that Meta does not review and, as of October 2026, appears to offer limited published enterprise administration features, making server-level permissions especially important.
Oct 26, 2026
3,038 words in the original blog post.
ETL performance benchmarking should rely on controlled, repeatable tests conducted in an organization’s own environment rather than vendor-published results, which often use unrepresentative datasets, configurations, and infrastructure. Although TPC-DI provides a formal model for data-integration workloads, organizations should define their own production-relevant workload, success thresholds, dataset scale, hardware configuration, and correctness checks, then test batch, incremental, and streaming modes separately where applicable. Key measures include throughput, end-to-end and p95 latency, CPU and memory use, error rates, scalability under concurrency, connector and query-pushdown capabilities, and normalized cost per unit of data processed. Fair comparisons require identical transformations, source and target conditions, compute tiers, and validation methods such as hash checks and record reconciliation, while measurements should isolate extraction, transformation, and loading stages to identify bottlenecks. Results should be presented in a common scorecard that balances performance, reliability, connector depth, and total cost of ownership, with testing repeated across increasing data volumes and realistic background activity. The material also promotes CData Sync, citing an internal replication benchmark, but emphasizes that all vendor claims should be independently reproduced.
Oct 26, 2026
2,200 words in the original blog post.
Microsoft Copilot agents can access live enterprise data outside Microsoft Graph through Model Context Protocol (MCP) tools, but secure deployment requires carrying each user’s identity and permissions through Copilot Studio, CData Connect AI, and the underlying source systems. CData Connect AI is presented as a centralized MCP gateway that connects agents to systems such as Salesforce, SAP, Workday, NetSuite, and Snowflake while consolidating authentication, query processing, permissions, monitoring, and audit logging that would otherwise be managed separately for each source. The recommended configuration uses Microsoft Entra ID, end-user credentials in Copilot Studio, per-user authentication in Connect AI, and source-level role-based access controls to prevent agents from exposing data beyond a user’s existing authorization. Organizations also need appropriate Copilot Studio licensing and credits, generative orchestration, and optionally Microsoft Agent 365 for agent registries and MCP server approvals. Common causes of generic responses include missing MCP tools, disabled orchestration, unsigned-in users, insufficient source permissions, or inadequate source metadata, while activity maps and gateway logs can help diagnose access issues.
Oct 23, 2026
2,096 words in the original blog post.
CData Connect AI Gateway’s context engine is presented as a system for helping AI agents interpret enterprise questions accurately by combining source-system metadata, business semantic definitions, and unstructured organizational knowledge into a shared, editable context graph. It aims to resolve ambiguities such as metric definitions, fiscal calendars, source-specific terminology, joins, and reconciliation rules before an agent queries data, while also selecting an appropriate model and limiting retrieved data and tool calls. The engine can build context from CData connectors, existing semantic models and catalogs, documents, prompts, user feedback, and monitored conversations, with human review and access controls intended to govern updates. CData says this approach enables consistent definitions across AI tools and models, reduces token use, and improves accuracy; in its cited benchmark of 378 prompts, it reported 98.5% accuracy for Connect AI compared with 65–75% for other MCP approaches, alongside substantial reductions in tokens and tool calls. The context engine is available through an early-access program, with a live demonstration planned for the company’s Foundations 2026 event on November 5.
Oct 22, 2026
2,000 words in the original blog post.
CData Connect AI is presented as a managed Model Context Protocol gateway that enables Databricks Agent Bricks agents to access live SAP ERP data without ETL replication, while enforcing existing SAP permissions at request time. Agent Bricks is Databricks’ platform for developing and governing enterprise AI agents, and its support for external MCP servers allows Connect AI to be registered through Databricks Marketplace and governed as a Unity Catalog MCP service. Deployment requires a Unity Catalog-enabled Databricks workspace, selected preview and beta capabilities, SAP OAuth 2.0 support, and OData services for the SAP modules involved. The setup involves authenticating SAP in Connect AI, creating an OAuth application, connecting the MCP server in Databricks, and configuring the agent to use SAP tools, with access typically based on the designated Connect AI user’s source-system privileges. Unity Catalog and Connect AI provide complementary audit logging, while potential applications include finance reconciliation using FI data and supply-chain or procurement workflows using SD and MM records, including governed write-back actions.
Oct 22, 2026
1,442 words in the original blog post.
Business context, rather than raw data access or model size alone, is presented as a key factor in enabling AI agents to accurately use enterprise data across systems such as CRM, cloud warehouses, and ITSM platforms. In a CData benchmark of 27 models completing read and write tasks, agents supplied with searchable plain-language context entries succeeded in 82% of 405 runs, while raw-access agents often spent their turns exploring schemas and hit a 40-turn limit without completing tasks. Context entries describing data objects, business metrics, eligibility rules, and cross-system joins reduced median tool calls from 40 to 8, tokens from 489,000 to 89,000, median costs from $0.27 to $0.09 per run, and wall-clock time from 176 to 94 seconds. Read tasks achieved 87% success with context, while write-task accuracy also improved substantially, though incorrect writes remained possible, reinforcing the need for controls such as scoped permissions, approvals, and review queues. The suggested approach is to identify common agent failures, document relevant tables and business rules in concise machine-readable entries, provide guidance for cross-system joins, and measure results before and after implementation.
Oct 22, 2026
1,902 words in the original blog post.
CData Drivers 2026.1 expands SQL-based connectivity with five new drivers for Claude Compliance, Linear, JD Edwards EnterpriseOne, Talkdesk, and Sage X3, while promoting Anaplan and SAP BW from beta to full release. The update rewrites the Jira, Stripe, Amazon S3, and Google Cloud Storage drivers with revised data models, performance and metadata improvements, and expanded capabilities such as multi-schema storage access and API-side query processing. It also introduces date-based HubSpot schemas ahead of the platform’s planned 2027 legacy API deprecations, adding migration guidance for users of older schemas. The release adds 25 API Profiles for connecting to REST APIs without dedicated drivers, expands create, update, and delete support for selected profiles, and includes source-specific enhancements across products such as SAP Gateway, SharePoint, GitHub, Splunk, QuickBooks Online, Office 365, Jira, and Zoho CRM.
Oct 21, 2026
1,149 words in the original blog post.
An AI gateway control plane is presented as a centralized governance layer that defines policy, identity, routing, authorization, and audit rules for AI agents accessing models and live enterprise systems, while a separate data plane processes and enforces those rules on individual requests. It supports governance through identity passthrough via OAuth or SAML, least-privilege tool and workspace access, centralized routing, and audit logs that record request metadata without necessarily retaining returned source data. The text explains that MCP can standardize agent connections to external tools and data sources, with OAuth 2.1-based authorization enabling consistent control over accessible tools, resources, and actions. It also contrasts building these capabilities in-house, which requires sustained work on authorization, connectors, logging, and specification changes, with purchasing a managed platform such as CData Connect AI Gateway, which is positioned as providing governed connectivity across many enterprise systems.
Oct 21, 2026
1,884 words in the original blog post.
Small language models (SLMs) are compact, task-focused transformer models designed to provide faster, lower-cost, and more controllable AI capabilities than large language models in enterprise settings. Rather than relying on broad memorized knowledge, they use APIs, databases, and enterprise systems to retrieve current information and perform specialized tasks such as customer-support summarization, sales quote generation, financial analysis, and HR policy assistance. Their smaller size enables low-latency deployment on local, edge, or restricted infrastructure while simplifying auditing, governance, and alignment with business rules. Because SLMs depend on reliable access to live, secure data, CData positions its Model Context Protocol Servers as a standardized interface that connects models to more than 270 enterprise systems, enforcing source-level permissions and producing structured, deterministic results without copying or reindexing data.
Oct 21, 2026
792 words in the original blog post.
Model Context Protocol provides a standard for AI assistants to access enterprise systems but leaves identity, authorization, and auditing largely to implementers, creating governance challenges when agents can retrieve data or trigger actions. An enterprise MCP gateway addresses these gaps by serving as a centralized control layer between agents and MCP servers, handling authentication, identity propagation, role-based access, routing, rate limits, policy enforcement, and unified audit logs across multiple systems. The text argues that gateways reduce the complexity and security risks of direct agent-to-tool integrations, prevent credential sprawl and token passthrough, enforce least-privilege access, and preserve each user’s identity at runtime so downstream systems apply their existing permissions and record accurate activity. It highlights requirements such as per-client consent, token audience validation, HTTPS, SIEM-exportable logging, workspace isolation, and integration with OAuth, SAML, and identity-provider lifecycle management. CData Connect AI Gateway is presented as an example platform offering these capabilities, including private cloud connectors that keep on-premises systems behind firewalls through outbound-only connections while supporting managed, private, and hybrid deployment models.
Oct 20, 2026
2,159 words in the original blog post.
Enterprise AI gateways are presented as centralized layers for routing, securing, governing, and auditing AI models, agents, tools, and data access, but the text argues that conventional model-focused gateways often do not solve the separate challenge of reaching live enterprise data in CRM, ERP, and warehouse systems. It positions CData Connect AI Gateway as a combined model and data-access platform, using managed Model Context Protocol (MCP) connections to provide scoped access to hundreds of sources, identity-based permissions, audit logging, SIEM export, and support for multiple AI clients such as ChatGPT, Claude, Copilot, and Gemini. The discussion distinguishes data-layer AI gateways from traditional API gateways, LLM routers, iPaaS products, and native AI-platform connectors by emphasizing source-level permissions, cross-system federation, and semantic context. Citing reports from MIT, Anthropic, and Gartner, it contends that weak integration, inaccessible data, cost concerns, and inadequate controls contribute to failed AI initiatives, while noting that data access addresses only part of these risks. It recommends evaluating gateways on MCP governance, OAuth or SAML identity passthrough, exportable audit trails, vendor neutrality, and secure cross-system data access, while describing CData’s model-routing capabilities as early access as of October 2026.
Oct 19, 2026
1,763 words in the original blog post.
Multi-source AI agent systems connect specialized agents to live data across enterprise platforms such as CRM, ERP, helpdesk, databases, and SaaS applications, reducing the complexity of maintaining many point-to-point integrations through a governed gateway and common standards such as the Model Context Protocol (MCP). The proposed architecture separates planner agents, which interpret context and create structured plans, from executor agents that perform approved, deterministic actions, while an orchestration layer manages state, retries, task routing, tracing, and escalation. Security and compliance are designed into each request through per-user or role-based identity, least-privilege access, data minimization, immutable audit logs, kill switches, and alignment with frameworks including SOC 2, GDPR, and ISO/IEC 27001. MCP supports agent access to tools and data, while Agent2Agent (A2A) enables agents from different teams or vendors to delegate work to one another. The text also recommends risk-tiered human oversight, continuous monitoring of performance and costs, and staged deployments beginning with measurable, low-risk workflows and shadow-mode testing before broader write access. It presents CData Connect AI as a gateway for accessing hundreds of live enterprise sources and cites its reported testing and operational advantages.
Oct 16, 2026
1,911 words in the original blog post.
CData Connect AI can connect ChatGPT dots, OpenAI’s always-on personal agents, to hundreds of business applications including Salesforce, Jira, Google Analytics, NetSuite, Xero, QuickBooks, and on-premises systems through a single ChatGPT plugin. The process involves connecting and authenticating data sources in Connect AI, installing and signing into its ChatGPT plugin, and prompting a dot to use live, read-only data for questions that span multiple systems. Suggested uses include identifying aging support tickets, reviewing sales pipelines, combining invoice totals from separate accounting platforms, assessing deal risk using Salesforce and Jira information, and scheduling recurring checks that can deliver results in ChatGPT, Slack, or Microsoft Teams. Connect AI is presented as providing source-side filtering, centralized permissions, audit logs, and access controls, while users are advised to limit permissions through dedicated read-only accounts where necessary, reauthenticate expired connections, and explicitly name the plugin and data source in prompts.
Oct 15, 2026
3,041 words in the original blog post.
CData promotes its Foundations 2026 event on November 5, where it will examine how AI models perform on live enterprise data rather than relying solely on public leaderboards. The company argues that benchmark rankings can be distorted by training-data leakage and broad average scores that do not reflect business-specific, agentic tasks or access to enterprise systems. Its open CData Labs program publishes reproducible benchmark harnesses, runners, and raw results for testing models and MCP implementations through CData Connect AI, focusing on accuracy, token efficiency, and behavior under different gateway configurations. Previously published results reported 98.5% accuracy for Connect AI on enterprise questions, compared with 65% to 75% for other MCP implementations, and found that configured context and guardrails reduced token use from 183,000 to 4,500 tokens for one query. A forthcoming session led by Jerod Johnson and Amit Naik will discuss tests of 22 models across live CRM, warehouse, and ITSM data, including costs, write-task safety, token sources, routing, and criteria for selecting models once accuracy and safety are standardized.
Oct 15, 2026
897 words in the original blog post.
MCP servers may operate with configured credentials and optional authorization, creating risks such as excessive agent permissions, confused-deputy access, shared-token exposure, missing user-level audit trails, and permission drift. The proposed approach is to enforce authentication, least-privilege policies, role-based access control, workspace and tool isolation, and logging through an independent centralized control plane that checks each request before it reaches an MCP server. Runtime identity passthrough using OAuth, SAML, OIDC, and emerging Enterprise-Managed Authorization mechanisms can allow requests to run under the actual user’s current permissions rather than a broad shared service account. Centralized enforcement also supports query-level audit records, SIEM integration, and compliance requirements such as SOC 2, ISO 27001, GDPR, and CCPA. The discussion contrasts the ongoing costs and maintenance demands of building these controls internally with managed platforms such as CData Connect AI, which positions itself as a gateway for governing agent access to enterprise data.
Oct 15, 2026
1,851 words in the original blog post.
CData Connect AI can link Anthropic Claude to live ServiceNow data through the Model Context Protocol, allowing users to ask natural-language questions about incidents, change requests, and configuration records without relying on separate reports or custom integrations. The managed MCP server acts as a governed gateway between Claude and ServiceNow, querying the live instance without data replication or ETL while applying existing permissions and configurable access controls. Setup requires a ServiceNow OAuth application, an appropriately scoped service account, a Connect AI account, and a Claude account, followed by configuring OAuth credentials in Connect AI and adding its MCP endpoint to Claude Desktop, Claude Code, or the Agent SDK. Security features include least-privilege access, optional per-user authentication where supported, query logging, and SIEM audit integration, while troubleshooting involves confirming authentication status and object-level permissions.
Oct 14, 2026
1,330 words in the original blog post.
Model Context Protocol (MCP), introduced by Anthropic in 2024, is an open standard for connecting AI systems to data sources, while enterprise-grade MCP servers add operational capabilities such as authentication, authorization, audit logging, resilience, and centralized governance that the base protocol leaves optional. The discussion distinguishes local servers from remotely hosted servers and contrasts self-hosted implementations, which require teams to maintain security and integrations individually, with managed platforms that centralize these functions. Key enterprise evaluation criteria include identity and access controls, connector quality, governance and deployment options, request latency and token efficiency, and compliance, with emphasis on OAuth 2.1, scoped and revocable tokens, identity passthrough, and least-privilege enforcement to mitigate risks such as tool poisoning, unauthorized access, credential sprawl, and shadow AI. It compares CData Connect AI Gateway as a managed multi-source platform, Microsoft’s Enterprise MCP Server as a narrower read-only option for Entra identity and directory scenarios, and Anthropic’s reference servers as useful primarily for prototyping and learning. The source argues that organizations moving MCP projects into production should prioritize governed access and validate performance against their own workloads rather than selecting a server based only on connector counts.
Oct 13, 2026
1,755 words in the original blog post.
AI cost measurement has evolved from tracking token consumption, through token-based pricing analysis, to evaluating cost per completed task, but the passage argues that cost per task can be misleading when agents produce convincing yet incorrect results. It proposes cost per correct task, which divides total workload expense by only verified correct outcomes, as a more meaningful measure of AI efficiency. A CData Labs benchmark of 22 models found that none could correctly combine live CRM, warehouse, and ITSM data when given raw system access, despite many returning plausible account lists, while curated data-access tools containing predefined business logic enabled models across pricing tiers to produce identical correct answers. The findings suggest that AI economics depend not only on model selection but also on data architecture, since placing repeatable logic in the access layer can improve reliability and allow organizations to use lower-cost models; in the benchmark, equivalent correct results varied in cost by as much as 178 times.
Oct 12, 2026
992 words in the original blog post.
CData introduced its Connect AI Gateway at FabCon Europe as a platform intended to help organizations build context-aware AI agents that can use Power BI semantic models, external enterprise systems, documents, and internal expertise consistently. Built on CData’s connector network, the gateway combines centralized LLM and Model Context Protocol routing, a Context Engine for integrating system metadata, business definitions, semantic models, and company knowledge, security controls tied to each user’s source permissions, and a virtual data layer for live cross-system read and write access. In a demonstration, CData showed how a Power BI-based account health model could be imported or generated from live connections, governed through a restricted MCP toolkit, and used by a Copilot Studio agent to retrieve health scores, investigate support and usage signals, and create Salesforce follow-up tasks. The platform records requests for auditing and is designed to avoid broadly exposing source systems or replicating data, while CData reports that its context approach reduced tool calls and token usage in internal benchmark conditions.
Oct 12, 2026
1,500 words in the original blog post.
Despite rapidly expanding availability of vendor-native Model Context Protocol servers, CData reports that MCP tool calls through its Connect AI platform grew more than 70% between Q2 and Q3 2026, including for sources with production-ready native options. The analysis argues that native servers vary substantially in maturity and deployability: about 24% of traffic targets sources without any first-party server, while available options may be unsupported code repositories, customer-operated software, hosted services with restrictive quotas, or fully supported hosted products. Even for the approximately 22% of traffic directed at sources with robust native servers, enterprises may prefer a centralized connectivity layer because real agent workflows commonly span multiple systems. CData maintains that a unified gateway can centralize governance, permissions, auditing, semantic definitions, and cross-source joins while reducing the number of tool definitions placed in model context, which it says can lower token use and improve tool selection accuracy.
Oct 09, 2026
1,598 words in the original blog post.
Enterprise data replication can connect long-running operational systems such as Oracle, DB2, AS/400, SAP, and SQL Server to cloud analytics platforms without requiring risky system migrations or direct analytical queries against production databases. The approach uses a replication and orchestration layer, often based on change data capture (CDC), to read committed changes from transaction logs and continuously deliver current data to destinations including Snowflake, Databricks, Microsoft Fabric, data lakes, and BI or AI tools. This separation allows operational and analytical environments to evolve independently while supporting near-real-time dashboards, machine learning, forecasting, fraud detection, and intraday reporting with limited source-system impact. The text argues that effective replication depends on broad connector support for legacy and specialty systems, scalable orchestration, automated schema-change handling, SQL transformations, API-based provisioning, and version-controlled configurations. It presents CData Sync as a platform designed to provide these capabilities across on-premises, cloud, and hosted deployments, while emphasizing predictable pricing and support for both legacy and modern data architectures.
Oct 09, 2026
1,909 words in the original blog post.
AI-enabled self-service reporting is presented as a way for business users to create and share dashboards without relying on separate data engineering, modeling, and analyst teams, reducing delays and dependence on spreadsheet exports. CData describes its internal CData Reports platform as a custom Azure-hosted application that integrates with Claude through an MCP server, allowing employees to build, publish, and update reports from a central location. Its Connect AI product provides governed, live access to sources such as Salesforce, Zendesk, Snowflake, Google Analytics, and Stripe, supplying business context and enabling AI tools to query multiple systems through a universal toolset. The approach preserves source-level user permissions, avoids shared credentials, and logs queries for auditing, while aiming to let nontechnical users generate reports in plain language and validate results using their knowledge of the business.
Oct 08, 2026
1,269 words in the original blog post.
AI model data security governs what AI assistants and agents can read, retrieve, and act on when connected to live enterprise systems, distinguishing it from protections for training data, model parameters, or data at rest. The text identifies credential exposure, shadow AI, audit gaps, permission drift, and scope creep as principal risks, citing IBM breach research that associates inadequate AI access controls and unauthorized tools with higher breach costs and greater exposure of personal data. It recommends enforcing least-privilege access at the connectivity layer where AI requests reach source systems, rather than relying on chat interfaces or models to control permissions. Proposed controls include per-session access decisions under NIST zero-trust principles, OAuth or SAML identity passthrough, role-based access control, MCP authorization requirements, field-level restrictions, audit logging, and workspace isolation. CData Connect AI is presented as a gateway that applies existing source-system permissions to each request across interfaces such as MCP, ODBC, JDBC, REST, and OData without replicating data, while attributing activity to individual users for compliance and investigation.
Oct 08, 2026
1,449 words in the original blog post.
CData Connect AI provides a way to connect Claude to current Excel workbooks in OneDrive or SharePoint rather than relying on static file uploads, allowing Claude to query live data while retaining source-system permissions and audit logs. Setup involves adding an Excel Online connection in Connect AI through Azure AD or organization-managed Azure app credentials, then enabling the listed CData connector in Claude’s settings and configuring individual tool permissions for automatic use, approval, or blocking. Workbooks should use clean tabular structures with consistent headers, formats, and uninterrupted ranges to support reliable analysis. Claude can then answer questions about variances, revenue, inventory outliers, and other spreadsheet content, create separate reports or spreadsheets, and, where authorized, insert, update, or delete source rows, although write actions are immediate and should be tested cautiously. The connection can also combine Excel data with systems such as Salesforce, SAP, or NetSuite, while OAuth passthrough, tool-level controls, and query logging support governance by ensuring users access only files they are already permitted to view and by recording how AI tools use that access.
Oct 08, 2026
1,843 words in the original blog post.
AI agent integration architecture defines how autonomous agents access enterprise systems at runtime, including the protocols they use, authentication methods, and enforcement of user permissions, addressing challenges that traditional point-to-point APIs face when agents dynamically select tools and data sources. The discussion identifies three scalable patterns: a governed Model Context Protocol (MCP) layer that provides a shared, access-controlled interface across agents and sources; real-time data federation with query pushdown to retrieve only needed live data without replication; and an agent mesh that adds consistent identity, observability, discovery, and governance for distributed multi-agent environments. It emphasizes OAuth 2.1, PKCE, identity passthrough, user-level CRUD restrictions, and audit logging as production-readiness requirements that prevent agents from accessing or modifying data beyond a requester’s permissions. It also argues that platform-agnostic connectivity layers can reduce maintenance and preserve interoperability as organizations change models or agent frameworks, presenting CData Connect AI Gateway as an example of a managed MCP-based implementation.
Oct 07, 2026
1,939 words in the original blog post.
Multi-system LLM data access architecture determines how AI models securely retrieve current enterprise information from fragmented sources such as CRM, ERP, databases, and SaaS platforms, while controlling data movement, permissions, performance, and auditability. The text compares native Model Context Protocol (MCP) connectors, iPaaS synchronization, unified API or data federation, and centralized warehouse-based retrieval-augmented generation, emphasizing their trade-offs between live access, duplication, governance location, and scalability. It identifies prompt injection, excessive agent permissions, and inefficient cross-system data transfers as key risks, and recommends controls including identity passthrough, least-privilege role-based access control, request-level logging, and query pushdown to limit retrieved data and execute filtering or aggregation near the source. It argues that a centralized governed AI gateway or MCP endpoint can simplify management compared with many point-to-point connectors, while keeping credentials outside the model and requiring external authorization for consequential actions. The text presents CData Connect AI as a product offering these capabilities through prebuilt connectors, OAuth/SAML identity passthrough, RBAC, source-level pushdown, and auditing.
Oct 06, 2026
1,757 words in the original blog post.
As major AI gateway vendors rapidly add the same newly released frontier models, model coverage is increasingly presented as a baseline requirement rather than a meaningful differentiator, though it remains useful for provider failover and pricing flexibility. The comparison should instead focus on how narrowly a gateway scopes an agent’s access, whether permissions are enforced directly in source systems through delegated identities, how precisely tools retrieve and process data, and whether approved business definitions remain portable across models and data sources. Narrow, source-enforced access can improve governance, attribution, and token efficiency, while schema-aware retrieval can return targeted results rather than large record sets that may introduce semantic errors, increase costs, and persist in later conversation context. The piece argues that durable organizational context, including approved definitions and corrections, should be maintained independently of any individual model, and presents CData Connect AI’s Context Engine as a product designed to combine connector knowledge, internal documentation, and expert input for that purpose.
Oct 05, 2026
1,069 words in the original blog post.
Connecting Claude to Salesforce through a managed Model Context Protocol (MCP) gateway is presented as an alternative to direct API integrations that rely on long-lived application credentials, which can obscure individual user activity and create broader access risks. The approach uses delegated OAuth or SAML identity passthrough so requests are evaluated against each user’s Salesforce permissions, including object access, field-level security, and sharing rules, while tokens can be scoped and revoked individually. Salesforce’s move toward external client apps, MCP’s OAuth 2.1 requirements, resource-bound tokens, and mandatory PKCE are cited as supporting more standardized authentication, though organizations remain responsible for access scoping, logging, and endpoint protection. CData Connect AI is described as a managed gateway that centralizes these controls, adds per-user connection permissions, workspace management, audit logging, and Claude organization policies that can restrict tools to read-only use or approved actions. With appropriately limited access, Claude can retrieve live Salesforce information for tasks such as pipeline analysis, opportunity reviews, and account summaries, and may update records or trigger workflows only when explicit write permissions are granted.
Oct 05, 2026
1,687 words in the original blog post.
Philip Schmid’s reflections on early Model Context Protocol (MCP) development highlight recurring challenges involving stdio-based local setups, tool naming conflicts, excessive tool context, and schema compatibility across language models. CData argues that its MCP offering, launched in May 2025, addresses these issues through both local stdio servers and a cloud-hosted, streamable HTTP option in CData Connect AI for more stable, collaborative, remote-first deployments. Its approach standardizes a SQL-based set of data-access tools across local and remote servers, including querying data, executing actions, and discovering catalogs, schemas, tables, columns, and procedures. CData also supports per-agent tool filtering to reduce context overload and derives schemas from underlying data sources rather than optimizing them for individual models, while advising developers to test workflows across GPT, Claude, and open-source models.
Oct 03, 2026
839 words in the original blog post.
A CData Foundations session featuring FinThrive’s Andrew Chabot and Databricks’ Eric Tome described how Databricks and CData can jointly support enterprise AI initiatives by combining scalable data engineering, governance, analytics, and AI capabilities with broad data accessibility. Databricks consolidates, cleans, and prepares data from numerous operational systems, while CData’s virtualized semantic layer presents disparate sources as standardized, governed assets that can be consumed by Databricks and accessed through business tools such as Tableau, Power BI, and Excel without custom exports or additional pipelines. The speakers highlighted that this approach can reduce ETL work, shorten reporting and experimentation cycles, and allow business users to work with live, curated data while engineers retain governance and optimization in the lakehouse. A demonstration showed datasets such as assets, locations, and sensors appearing as native Databricks tables, then being made available for real-time analysis and operational workflows. The partnership was presented as complementary: Databricks provides the compute and AI foundation, while CData accelerates data integration, virtualization, and delivery of insights across the organization.
Oct 03, 2026
983 words in the original blog post.
Vibe Querying is CData’s term for using natural-language conversations with large language models to access and analyze business data without requiring users to know SQL, database schemas, or specialized BI tools. Unlike traditional querying, which offers precise and controlled access but often depends on data teams or administrators, this approach lets users ask questions directly, iteratively refine them, and receive structured results quickly. CData says its Model Context Protocol servers support this model by presenting CRM, ERP, finance, HR, and other enterprise data through a tabular, SQL-native interface that LLMs can interpret, while allowing successful questions to become repeatable data assets. The approach could help teams across sales, finance, product, marketing, support, HR, and leadership self-serve insights, though its usefulness and safety depend on carefully designed permissions and governance to prevent unauthorized access or overly limited results.
Oct 03, 2026
868 words in the original blog post.
Secure enterprise integration with ChatGPT requires organizations to prepare and classify data, identify sensitive information and data flows, test connections in sandbox environments, and apply data loss prevention and monitoring controls before linking AI to systems such as CRMs, ERPs, HR platforms, and finance tools. The material recommends a zero-trust model based on continuous identity verification, multifactor authentication, least-privilege access, network segmentation, and ongoing monitoring, alongside automated detection and redaction of confidential information to reduce leakage risks. It emphasizes securing APIs through authentication methods such as OAuth 2.0 and managed secrets, encrypting data in transit and at rest with standards including TLS 1.3 and AES-256, using role-based access controls, rotating credentials, and complying with regulations such as GDPR, CCPA, and PCI DSS. Detailed audit logging, anomaly detection, regular reviews, and employee training are presented as essential governance measures, while CData Connect AI is promoted as a managed Model Context Protocol platform offering no-code, real-time connections to more than 300 data sources with inherited source security, access controls, encryption, and audit capabilities.
Oct 03, 2026
1,278 words in the original blog post.
A comparison of HubSpot’s native MCP Server, Zapier’s HubSpot MCP Server, and CData’s MCP Server for HubSpot found meaningful differences in setup, tool breadth, speed, and analytical capability when Claude Sonnet 4 answered a CRM sales-analysis question using the same fictional dataset. HubSpot’s local native server required Node.js, npm, manual private-app setup, and configuration-file editing, but delivered the fastest result in about 64 seconds with nine requests; its direct API approach was token-efficient but limited to basic JSON operations without joins, advanced filters, or aggregations. Zapier provided the most guided, OAuth-based remote setup and the broadest toolset, with 53 tools for specialized HubSpot actions, but required eligible paid Claude plans and took nearly four minutes despite only 10 requests, likely due to remote processing and an additional agentic layer. CData offered a wizard-driven local installation with OAuth or token authentication and fewer tools, but its SQL-powered run_query feature enabled joins, filtering, aggregations, and other complex analysis on live HubSpot data; it completed 18 requests in about two minutes, though its flexible database-like approach could consume more tokens while the model explores the data structure. The comparison concludes that HubSpot’s server is suited to simple, fast API access, Zapier is strongest for broad action automation, and CData is best positioned for sophisticated data-intensive analysis.
Oct 02, 2026
1,540 words in the original blog post.
CData’s 2025 major release of Drivers and Connectors, generally available July 31, 2025, adds an Adobe Target connector, native Apple Silicon support for Model Context Protocol Servers, and promotes Bitbucket, Cvent, Greenhouse, HubDB, Okta, and PingOne drivers from beta to general availability. Major enhancements include Apache Doris access through MySQL drivers, broader Shopify GraphQL support, a universal Adobe Analytics report interface, and improved Salesforce batch and Data Cloud ingestion capabilities. The release also adds Python 3.13 and JRE 17.0.13 support, customer-provided OAuth configurations for Tableau connectors, redesigned documentation, and a rebranded Vault CRM connector with create, update, and delete operations. More than 40 source-level additions improve integrations including MongoDB, GitHub, HubSpot, SharePoint, Stripe, Databricks, and Kafka, while over 360 updates address vendor API changes. It also retires or deprecates several drivers and platforms, including Microsoft Dynamics GP, Bing Search, IBM Cloud SQL, older Python versions, and 32-bit Python driver editions.
Oct 02, 2026
758 words in the original blog post.
CData Virtuality’s Q3 release, covering versions 25.2 and 25.3, expands its enterprise AI, data-access, and security capabilities through enhancements to Talk To Your Data, the Business Data Shop, and centralized authentication management. Talk To Your Data uses a semantic layer, large language models, a vector database, and the Virtual SQL engine to turn natural-language questions into governed queries, and it now allows organizations to upload business definitions as a glossary that continually strengthens shared semantic context. The updated Business Data Shop enables users to discover, access, and directly request curated datasets through a self-service interface, aiming to reduce IT bottlenecks while maintaining centralized governance. A new Authentication Management Component consolidates identity and access management, supporting providers including EntraID, Okta, Auth0, Google, GitHub, Shibboleth, and WSO2, along with SAML, OAuth2, OIDC, OTP, MFA, user federation, and role mapping.
Oct 02, 2026
734 words in the original blog post.
A GigaOm CxO Decision Brief authored by Darrel Kent and commissioned by CData argues that semantic layers can help enterprises balance data accessibility, governance, and AI adoption amid increasing regulatory and business pressures. It describes a semantic layer as a unified, governed delivery layer that applies shared definitions, access controls, lineage, and compliance policies to data used in analytics, reporting, and AI. The brief identifies three primary business outcomes: greater confidence in decisions through consistent and compliant metrics, AI readiness through reliable and explainable data, and organizational agility through secure self-service access that reduces reliance on IT. It also highlights broader organizational effects, including improved collaboration among engineers, data stewards, and business users, with data treated as a reusable governed asset. CData is presented as providing infrastructure for semantic access through data virtualization, replication, transformation, and policy enforcement without requiring organizations to replatform existing systems.
Oct 02, 2026
758 words in the original blog post.
Claude API costs are driven not only by prompts and model output but also by repeated system context, conversation history, tool schemas, retrieved data, and multi-step agent workflows, particularly when enterprise systems are connected through MCP. The discussion recommends routing simple work to lower-cost models such as Haiku, using Sonnet for typical production tasks and reserving Opus for complex reasoning, while reducing repeated-context charges through prompt caching and using batch processing for asynchronous high-volume jobs. It also advocates context management through trimming irrelevant history, summarizing long interactions, and storing large intermediate state externally, while warning that excessive reduction can impair quality and create costly retries. A central architectural recommendation is consolidating MCP integrations behind a governed access layer that limits tool definitions and pushes filtering, joins, and aggregations to source systems before data reaches the model. CData Connect AI is presented as an example of this approach, with published benchmarks claiming substantial token and inference-cost reductions. The text further emphasizes cost attribution, per-team monitoring, access controls, and auditability as ways for organizations to make AI spending more predictable.
Oct 02, 2026
1,448 words in the original blog post.
Manufacturers facing new EDI requirements must bridge the gap between ERP business data and X12 transaction formats, often under tight customer deadlines for acknowledgments and compliance. The core order cycle commonly uses the X12 850 purchase order, which must be translated into an ERP sales order and acknowledged with an 855; the 856 advance ship notice, which combines ERP and warehouse data to describe shipment contents before delivery; and the 810 invoice, which must reconcile with both the original order and actual shipment to avoid payment delays. Errors or late transmissions, particularly for 856 documents, can result in retailer chargebacks and disputes that require reliable audit records. Companies may use a value-added network, develop custom integrations, or adopt a dedicated EDI platform, with custom projects requiring ongoing maintenance as partner specifications and ERP systems change. CData Arc is presented as a platform-based alternative that uses visual workflows, X12 parsing, XML mapping, AS2 or SFTP transport, and prebuilt ERP and database connectors to automate inbound and outbound document flows, reuse integrations across transaction types, and reduce implementation work to trading-partner-specific rules and data mappings.
Oct 02, 2026
2,128 words in the original blog post.
AI data access control governs which enterprise data AI models and agents can access and which actions they may perform, with enforcement placed at the connectivity layer rather than relying solely on written policy or model-level settings. The material identifies credential exposure, excessive data scope, and incomplete auditing as central deployment risks, referencing sensitive information disclosure as a prominent LLM security concern. It recommends prerequisites including data inventories, source-system role and identity mapping, and least-privilege policies, followed by controls such as OAuth or SAML identity passthrough, use-case-specific workspaces, read-only and CRUD restrictions, live data access without replication, narrowly scoped tools, and SIEM-exportable request logs. A proposed implementation involves routing sources through a governed gateway, configuring permission inheritance, testing workspace boundaries, collecting audit evidence, and conducting a limited pilot before production. These practices are presented as supporting GDPR, SOC 2, CCPA, and potentially HIPAA requirements by demonstrating data minimization, purpose limitation, access control, and monitoring, while CData Connect AI is positioned as a platform for applying these controls across multiple AI assistants and data sources.
Oct 01, 2026
1,598 words in the original blog post.