Enterprise Data Agents: Architecture, Governance, and the Data Readiness Problem
What enterprise data agents are, how their architecture and governance work, and why data readiness decides whether they deploy.
A question that should take a minute takes three days to route: to the analyst who owns the table, then to the platform team that owns the pipeline, then back to the business owner who owns the definition. A dashboard answers yesterday's question because the pipeline that feeds it runs overnight. And an agent demo that returned the right number in a pilot returns a different number in production, on the same schema, without anyone changing the model.
That last failure is the one that matters, because it is not a model problem. Deloitte's 2025 enterprise technology survey found only 11% of surveyed organizations actively using agentic systems in production, with 14% holding deployment-ready solutions and 38% still piloting (Deloitte, Tech Trends 2026, published 2025-12-10, survey fielded June-July 2025). Gartner's cancellation forecast projects that over 40% of agentic AI projects will be canceled by the end of 2027 on escalating costs, unclear business value or inadequate risk controls (Gartner, 2025-06-25).
Enterprise data agents are the category of AI systems that answer data questions and execute data work by operating directly on an organization's schemas, metadata, semantic definitions and permissions rather than treating the data platform as an external tool. The definition matters now because the two figures above describe a gap between demonstration and deployment that the architecture of enterprise data agents either closes or widens.
This guide works through seven questions: what the category is and where its boundary sits, what the architecture contains, what the platform must already be true of, how governance attaches identity and audit, how to evaluate accuracy before deployment, what a workload costs at multi-terabyte scale, and what changes for APAC multi-cloud and cross-border operations. It draws on vendor documentation, academic benchmarks, standards text and analyst survey data, and it states where those sources stop.
What enterprise data agents are, and how they differ from chatbots, BI copilots and generic AI agents
What is a data agent? It is a vertical agent whose data capability is native rather than attached. Schema, metadata, lineage, semantic definitions and permissions are its core context, and its typical actions are generating and executing SQL and diagnosing data problems. A general agent instead relies on tool calling, treating the data platform as an external system it reaches into.
Microsoft's documentation draws the boundary concretely. A Fabric data agent is a generally available feature for conversational Q&A over data in OneLake that binds up to five data sources in any combination, routing relational sources through NL2SQL (natural language to SQL), semantic models through NL2DAX (natural language to DAX) and KQL databases through NL2KQL (natural language to KQL) (what a Fabric data agent is, page dated 2026-05-11). Databricks Genie grounds answers in the organization's data and governs them through Unity Catalog, with quality tuned by adding metrics, business rules and verified answers (how Databricks Genie is governed, current docs). Databricks Genie Agents are domain-specific agents configured with a curated semantic knowledge store of definitions that the product page states improves response accuracy (Databricks Genie Agents, current product page).
Snowflake's Cortex Analyst shows the same shape and its limits. It is a fully managed agentic text-to-SQL system that requires a semantic model and returns what the documentation calls highly accurate text-to-SQL responses, but it cannot access the results of previous SQL queries in a conversation and cannot answer non-SQL business questions such as "What trends do you observe?" (Cortex Analyst, current docs).
Dimension | Chatbot | BI copilot | Generic AI agent | Data agent |
|---|---|---|---|---|
Core context | Conversation history | Report and dataset metadata | Tool descriptions and prompts | Schema, metadata, lineage, semantic model, permissions |
Typical actions | Answer, escalate | Suggest a visual, draft a measure | Call arbitrary tools | Generate and execute SQL, diagnose data problems |
Hardest problem | Intent and tone | Model and report design | Tool selection | Finding the right data, confirming source of truth, metric ambiguity |
Governance requirement | Content policy | Workspace permissions | Tool-level scopes | Row and column permissions, masking, audit, freshness, compliance boundaries |
The trade-off is real and worth stating plainly: a data agent covers a narrower range of tasks than a general agent, and in exchange its answers are more accurate and more auditable. That is the boundary that separates enterprise data agents from the packaging around them, and the analytics front door is where that boundary starts to show up in day-to-day work.
What a data agent architecture actually consists of
A data agent architecture is a layered stack in which the agent sits on top of context, semantic, identity and observability layers, and every layer is a named component rather than a prompt. Microsoft's agent data architecture guidance names Fabric OneLake as the context layer, Foundry IQ as the default governed retrieval layer, Microsoft Entra ID with managed identity, OAuth identity passthrough and RBAC for identity, and requires authentication on every MCP tool call (Microsoft's agent data architecture guidance, page last updated 2026-03-10). These are one vendor's layer names, not an industry standard, and the same guidance instructs leaders to document each data domain and classify its retrieval method as search, API or both.
The semantic layer is the load-bearing part. Snowflake semantic views are schema-level database objects that store metrics, dimensions, facts and relationships as metadata, giving consistent business definitions across applications (Snowflake semantic views, current docs). The documentation's own example is a gross revenue concept stored in a column named amt_ttl_pre_dsc, where the semantic view defines net revenue as SUM(gross_revenue * (1 - discount)) so a question about net revenue by region aggregates at the right level. Without that object, the page states, dozens of inconsistent calculations can exist across reports and applications with incorrect aggregation methods.
That is why the query path has two branches. When semantic coverage is high, the agent queries the semantic view directly; when coverage is thin, it borrows the view's definitions to supply context and assembles SQL against physical tables. The second path is where accuracy degrades, and it is the path most enterprise schemas start on.
The "so what" is that permissions and definitions attach at different layers, and a deployment that treats the semantic layer as an application-side convenience rather than a database object inherits its own metadata sync, connection management and caching as separate systems to govern.
What must be true of the data platform before an agent can be trusted
Data readiness, not model quality, is the binding constraint, and the benchmark evidence is unambiguous about the size of the gap. The Spider 2.0 benchmark reports that the best models solve only 17.1% of its tasks (o1-preview) and 10.1% (GPT-4o), against 86.6% for GPT-4o on Spider 1.0 (current leaderboard, benchmark released 2024). Its 632 problems run against enterprise databases with over 1,000 columns, some over 3,000, and queries that often exceed 100 lines.
The BEAVER enterprise text-to-SQL benchmark, built from three anonymized private enterprise warehouses spanning 19 domains, reports a best end-to-end execution accuracy of 11.4% (ReFoRCE with Claude-4.5-Sonnet), with state-of-the-art agentic frameworks on GPT-5.2 reaching 10.8% (arXiv v3). The two benchmarks measure different task sets and are not comparable to each other, only within themselves.
Deloitte's findings on data searchability and reusability locate the constraint upstream of the model: in a survey published 2025-12-10, 48% of organizations cited searchability of data and 47% cited reusability of data as challenges to their AI automation strategy. An AI-ready data platform, in this context, is one where data's location, meaning, ownership and access rules are already recorded as metadata the agent can read, so the agent spends its capacity on the question rather than on rediscovering the schema.
Key Takeaway: Data readiness, not model quality, is the binding constraint on enterprise data agents. A stronger model does not resolve a schema the agent cannot interpret.
The most useful evidence on what semantics can and cannot fix comes from BEAVER's oracle-hint experiment. Given all subtask annotations, the best method reaches 25.9%, with ReFoRCE at 25.9, few-shot at 23.1, DIN-SQL at 13.9 and DAIL-SQL at 11.2. Under schema-linking annotations only, ReFoRCE reaches 18.9% against 9.5% end-to-end. The enterprise semantic layer addresses roughly the schema-linking portion of that gap: table retrieval, join keys and column mapping. It leaves query decomposition and domain knowledge largely untouched. No source measures the semantic-layer effect on a private enterprise schema, so no percentage is claimed here.
The join-key problem is where mechanism and failure mode converge. BEAVER reports that methods perform worst on join key detection because enterprise schemas often lack explicit join relationships, forcing methods to infer them during generation, while cryptic table and column names obscure which attributes are joinable. That is the clearest case of a platform mechanism targeting a measured research failure, and it is why what makes data AI-ready is a metadata question before it is a modelling one. Where the metadata stops short, the business context gap is what an agent cannot close on its own, and no AI-ready data platform can be certified from the model side alone.
How we apply this at Singdata, the publisher of this guide. Our Data Agent is a vertical agent for the data domain, where data capability is native rather than attached, so it works from schema, metadata, lineage, semantic definitions and permissions instead of treating the data platform as an external tool. It does not make a platform data-ready on its own: readiness, source-of-truth confirmation and metric definitions remain work your data team does, with any tool, ours included.
How you govern an enterprise data agent: identity, permissions, audit and human approval
Data agent governance attaches an identity to the agent, inherits the data platform's row and column restrictions, and writes an audit record for every query, with human approval reserved for actions that change state. The NIST AI Risk Management Framework's four core functions, GOVERN, MAP, MEASURE and MANAGE, provide the structure, with GOVERN cross-cutting and addressing the full product lifecycle including third-party software and data (the NIST AI Risk Management Framework's four core functions, NIST AI 100-1, January 2023). The framework is voluntary guidance, not a rule.
Identity is the first control. Microsoft's guidance requires authentication on every MCP tool call and supports key-based access, Microsoft Entra Agent Identity, project managed identity and OAuth identity passthrough, with RBAC applied on both the Foundry project and the target service and identity passthrough preferred when user-level permissions must persist. The Fabric data agent uses the requesting user's credentials and permissions to enforce least-privilege access, evaluates requests against tenant and workspace policy settings before executing any action, and strictly enforces read-only access to all data sources.
The second control is the prompt-injection surface, which is larger for a data agent than for a chatbot because the agent reads documents and holds database credentials. OWASP lists prompt injection as LLM01:2025 and defines indirect prompt injection as the model accepting input from external sources such as websites or files that alters its behaviour, noting that RAG and fine-tuning do not fully mitigate it and that severity depends on "the agency with which the model is architected" (OWASP's prompt-injection entry, LLM01, 2025 edition). OWASP's named countermeasures include enforcing privilege control and least-privilege access, giving the application its own API tokens rather than exposing functions to the model, and handling privileged functions in code.
The third control is the audit record, and the regulatory floor depends on classification rather than on the technology. Under the EU AI Act, high-risk systems must technically allow automatic logging of events over their lifetime, and deployers must be enabled to understand system limits, monitor operation and detect anomalies, with oversight measures commensurate with risk, level of autonomy and context of use (the AI Act's record-keeping requirement, Article 12/14 obligations applying from 2 December 2027 for Annex III high-risk systems). Article 12 does not state a log-retention duration, and a general-purpose analytics data agent is not automatically high-risk under Annex III, so the classification condition has to be assessed rather than assumed.
For procurement, ISO/IEC 42001:2023 specifies requirements for establishing, implementing, maintaining and continually improving an AI management system and is subject to independent third-party audit (independently audited against ISO/IEC 42001, standard published December 2023). It certifies the management system, not the agent's answers, and a vendor's own certification does not cover the customer's implementation.
semantic layer for AI trust
How do you evaluate whether a data agent is accurate enough to deploy
Accuracy alone is the wrong acceptance test for enterprise data agents. The CLEAR framework reports that frontier models score 68 to 74% single-run success on a 300-task enterprise agent suite, and states a production requirement of pass@8 of at least 80% for mission-critical applications (the CLEAR enterprise agent evaluation framework, arXiv 2511.14136, 2025). That 80% figure is the authors' stated requirement, not a measured outcome, and the distinction matters when you write it into an acceptance criterion.
The reason to test repeated runs is that single-run scores hide brittleness. CLEAR documents agent performance falling from 60% at pass@1 to 25% at pass@8, which means a system that looks strong on one attempt can degrade sharply when the same task is retried. An evaluation plan built on a single successful run will not surface that.
Evaluation dimension | NIST function | Named failure mode |
|---|---|---|
Task efficacy | MEASURE | Confabulation |
Reliability across repeated runs | MANAGE | Single-run success masking brittleness |
Cost per completed task | MANAGE | Cost misestimation up to 100% |
Latency against SLA | MEASURE | SLA breach under concurrency |
Assurance and policy adherence | GOVERN | Data privacy, indirect prompt injection |
The NIST Generative AI Profile supplies the risk vocabulary to map those dimensions onto, naming confabulation, data privacy, and information integrity among the risks specific to generative systems (the generative AI profile's risk list, NIST AI 600-1, July 2024).
One question the evidence does not settle is whether benchmark results transfer to production. CLEAR reports a 37% gap between lab tests and production deployment, alongside cost misestimation up to 100%. Treat the readiness framework in this guide as a decision aid for a governance-first rollout, not a certification.
What an agent workload costs at multi-terabyte scale, and does it change platform TCO
An agent workload is priced per completed task, not per query, and the distribution is wide enough that the median hides the tail. The cost distribution in that evaluation reports a median task cost of $0.34, a p99 of $2.72 and a most expensive task of $3.87, with only 4% of tasks above $2 (Databricks' 401-task data-agent evaluation, vendor-run, no publication date shown on the page). The same run put 33-40% of tasks above $1 for the general coding agents against 16% for the data-native one, which averaged 8.3 tool calls per task. Accuracy moved with cost: 76.6% at $0.55 mean cost per task for the data-native agent, against 72.1% at $1.09, 55.9% at $0.91 and 56.1% at $1.16 for the general coding agents. All of it is vendor-run, undated, and measured on 401 self-contained tasks distilled from the vendor's own internal usage under an identical 20-minute wall-clock budget, so treat it as an observed envelope rather than an independent audit.
At multi-terabyte scale the cost structure underneath the task matters more than the task count. Lakehouse unit economics decomposes a query into compute time, bytes scanned and per-file request cost, normalising vendors to dollars per terabyte scanned and dollars per hour of effective compute, with compute cited as 60-80% of total expenditure and data transfer as up to 30% of total cloud data platform expense (2026-09-10). That source gives a working self-managed rate of $0.05 per core-hour and $5 per TB scanned on scan-priced engines, and notes the effective rate has to be measured on your own infrastructure rather than read off a list price.
The cost-per-completed-task formula follows: compute time × effective core-hour rate, plus bytes scanned × per-TB rate, plus per-file request cost, plus model tokens, plus retrieval and transfer, all divided by tasks completed successfully. The allocation of those layers to a single task is modelled, not measured, because no source publishes a per-task split of platform spend.
On TCO, the honest answer is that no vendor-neutral study with a stated methodology was found. The available consolidation figures are customer-reported case results, a 35% storage and 50% compute reduction at one automotive customer and a 35% cost reduction at a logistics customer, and they read as case figures rather than a market finding. The lever they point at is the same one that governs agent cost: reduce data platform TCO by consolidating the platform layer an AI-ready data platform runs on, so scanned bytes and core-hours fall before the agent multiplies them.
What changes for APAC, multi-cloud and cross-border data operations
Cross-border operation is where enterprise data agents meet transfer law, and the first question is which regime governs the inference call. Singapore's cross-border transfer rules require that personal data moved outside the jurisdiction receive a standard of protection comparable to the PDPA, and the Personal Data Protection Commission has endorsed the ASEAN Model Contractual Clauses as the regional instrument for that flow, the counterpart to the EU Standard Contractual Clauses (Latham & Watkins guide to Singapore data privacy transfer investigations, 2021; DLA Piper Data Protection Laws of the World, Singapore chapter, last modified 2026-02-27). Both sources are law-firm reporting of the regulator's position rather than the regulator's own text, so treat the obligation as established and the drafting detail as something your counsel confirms.
For China, this research obtained no official source on the PIPL outbound-transfer mechanism. Rather than assert what a China-to-overseas inference call requires, the honest position is that the mechanism is outside what the evidence here supports, and a deployment touching mainland data needs its own regulatory review.
Multi-cloud estates add a second constraint that is architectural rather than legal. NinjaVan's stated requirement is a multi-cloud environment, and the pattern that answers it is consolidation onto a single engine that reads across the estate rather than replicating the semantic layer into each cloud. Synagie's selection turned on data freshness, which sets the practical test: if the agent's answers lag the source systems, the cross-cloud topology is buying reach at the cost of the one property the agent exists to provide.
The AI Act's application timeline for Annex III systems adds a dated obligation for anyone with EU exposure: logging traceability and deployer-side human oversight apply to those high-risk systems from 2 December 2027 (EU AI Act Article 12, Regulation 2024/1689). That is a design deadline, not a policy statement, and it lands on the audit trail the agent produces.
Common misconceptions about enterprise data agents
The most consequential misconception is that a vendor accuracy claim is a benchmark result. Vendor-run evaluations typically test the vendor's own agent against competitors' agents on a task set the vendor selected, and no independent replication of those 401-task figures was found. That does not make the claims false. It makes them unverified, which is a different thing. Read a vendor accuracy claim as a mechanism statement, and ask for the task set, the run count and the harness before treating it as a number.
How to read a vendor accuracy claim: check who ran the evaluation, check what it was run against, and check whether the page carries a date. A claim with no date cannot be compared to anything, including itself six months later.
The second misconception is that a better model fixes accuracy. The benchmark evidence points elsewhere: the constraint sits in schema and semantics, not in model capability. A semantic layer addresses the schema-linking portion of the problem and leaves query decomposition and domain knowledge largely untouched, which is why the oracle-hint ceiling matters more than any leaderboard movement. The model is often the same across a working deployment and a failing one. The definitions are not.
The third is that agentic AI is broadly deployed. Deloitte's survey found production use running well behind pilot activity, with a large share of organizations still developing a strategy or holding none at all, so a reader who assumes everyone else has shipped is reading the market wrong.
The fourth is that the widely quoted Gartner cancellation figure represents many independent findings. It is one press release, repeated across outlets, and it counts as one source. Treat the echo as volume, not corroboration.
The practical test for any enterprise data agents claim is the same one this section keeps returning to: name the evaluator, name the comparison, name the date. Where a claim survives that test, it is worth planning around. Where it does not, the honest answer is that the evidence stops there, and the limits of LLM-generated SQL are a better place to start than a vendor's accuracy percentage.
Conclusion
The evidence points to one conclusion: the constraint on enterprise data agents is the platform underneath them, not the model inside them. Spider 2.0 and BEAVER both measure single-run accuracy on real enterprise schemas in the low teens, CLEAR sets a production bar of pass@8 of at least 80% and documents agents falling from 60% to 25% between single-run and repeated-run success, and Deloitte's survey places data searchability and reusability ahead of model capability as the stated barriers.
The architecture follows from that. A context layer, a semantic layer that holds definitions as database objects, an identity layer that authenticates the agent as itself and inherits row and column restrictions, and an observability layer that writes an audit record per query are the components that make an answer reproducible. Governance attaches at those layers rather than after them, and the evaluation plan has to measure cost, latency, assurance and reliability alongside accuracy, because the cheapest agent with comparable precision is the one that survives a budget review.
The next step is a readiness assessment rather than a pilot. Take one high-value question, trace it to the tables and definitions it depends on, and record what the platform cannot currently answer: which joins are implicit, which metrics have more than one definition, which columns carry restrictions the agent would need to inherit. That inventory is the artifact a security reviewer, a procurement team and an evaluation plan all need, and it is the part of the work that no model upgrade replaces.