Article
→ Insulet replaced legacy ETL with Databricks Lakeflow and achieved a 97% reduction in total cost of ownership (TCO) and 12x faster data processing — the most dramatic ROI metric from Data + AI Summit 2025, and a signal that operational cost destruction, not revenue uplift, is the primary enterprise AI value story. (Databricks blog, 2025)
→ Agentic AI workflows are no longer experimental: 7-Eleven's compound AI marketing system operates live across 13,000+ stores, while Mastercard deploys human-in-the-loop GenAI agents for customer onboarding — confirming that multi-agent AI systems have crossed from pilot to production at Fortune 500 scale. (Databricks blog, 2025)
→ Governance is the scaling bottleneck, not model capability: 40% of technology executives in a 2024 survey of 1,100 leaders said their AI governance programs were insufficient, and 53% named data privacy and security as their top concern — with Gartner projecting a 50% adoption lift by 2026 for organizations that operationalize AI transparency. (Economist Impact / Databricks AI Governance Framework, 2024; Gartner, cited 2024)
→ The competitive advantage in enterprise AI has migrated from model selection to data infrastructure quality: organizations that first unify their data foundations — through platforms such as Databricks' Lakehouse, Unity Catalog, and Lakeflow — consistently report the largest downstream AI performance gains, including Accenture clients achieving 75% ETL time reduction and 14x faster ML training. (Accenture, 2025)
Why This Matters Now
The enterprise AI conversation has reached an inflection point that most organizations are misreading. The dominant narrative — that competitive advantage flows from access to the most capable large language model (LLM) — has been quietly invalidated by the evidence emerging from 100+ production deployments presented at Data + AI Summit 2025. What separates the organizations achieving measurable, scaled AI returns from those still locked in pilot cycles is not the sophistication of their models. It is the maturity of the data infrastructure underneath them.
Consider the magnitude of the shift underway. Databricks now serves more than 5,000 global organizations, and the use cases its customers shared at the 2025 summit span every major sector — retail, energy, financial services, healthcare, and media (Databricks blog, 2025). These are not proof-of-concept deployments. They are production systems processing terabytes of data daily, automating decision-making across tens of thousands of locations, and replacing entire categories of enterprise software. The velocity of this shift has outpaced the governance and infrastructure readiness of most enterprises attempting to follow.
This piece synthesizes what the evidence from those deployments actually reveals: which data intelligence use cases for AI are generating the largest returns, how leading organizations have structured their platforms to enable agentic AI workflows at scale, and — critically — what the organizations falling behind are getting wrong about the sequence of investments required to succeed.
The Evidence: What the Data Shows from 100+ Data Intelligence Use Cases
The most reliable source for production-grade data intelligence use cases AI practitioners can benchmark against is the Data + AI Summit 2025, where Databricks' customers presented verified outcomes with named metrics. The findings challenge several prevailing assumptions about where enterprise AI value is created.
The Cost Destruction Story
The headline metric from the summit is not a revenue figure. It is Insulet's 97% reduction in TCO after replacing third-party ETL (Extract, Transform, Load) tooling with Databricks Lakeflow. The same deployment delivered 12x faster real-time data processing and an 83% reduction in SQL queries (Databricks blog, 2025). For a medical device manufacturer operating in a regulated environment — Insulet makes the Omnipod insulin delivery system — the implications extend well beyond IT cost savings: faster, cleaner data pipelines directly affect the speed and reliability of clinical and operational analytics.
This pattern of cost destruction rather than incremental revenue growth appears consistently across the summit data. When Accenture clients deploy the Databricks Lakehouse Platform with Unity Catalog, they report up to 75% reduction in ETL processing time and 14x faster ML training cycles (Accenture, 2025). These are not marginal efficiency gains. They represent the elimination of entire infrastructure layers and the vendor contracts that sustain them.
🔴 Important
The primary ROI driver for enterprise Databricks deployments is operational cost elimination — redundant ETL tooling, excess query volume, and legacy batch processing infrastructure. Leaders who frame the business case primarily around revenue uplift risk underinvesting in the foundational work that unlocks the largest returns.
Agentic AI in Production: The Scale Evidence
The assertion that agentic AI workflows remain a future-state aspiration is contradicted by the deployment evidence. 7-Eleven has built a multipurpose agentic marketing assistant on Databricks that actively tracks store performance insights across its network of 13,000+ stores, using compound AI systems that combine multiple models, retrieval mechanisms, and decision logic (Databricks blog, 2025). This is not a chatbot. It is a multi-agent AI system operating in real time across a distributed retail operation.
Mastercard's deployment is equally instructive for what it reveals about the maturity of human-in-the-loop design. The company uses a GenAI onboarding assistant to automate customer support interactions, supplemented by domain-specific AI agents that escalate to human reviewers when confidence thresholds or regulatory sensitivity require it (Databricks blog, 2025). This architecture — autonomous agents with structured human oversight — represents the operational standard for responsible LLM production deployment in regulated industries.
NOV (formerly National Oilwell Varco) presents the industrial-scale data argument: the energy services company processes more than 3 terabytes of real-time data daily through Databricks for predictive operations (Databricks blog, 2025). At that volume, retrieval augmented generation (RAG) systems and vector databases for enterprise use are not optional enhancements — they are the infrastructure required to make any LLM interaction with operational data coherent and reliable.
Fan-Facing AI: The Media Laboratory
Fox Sports' Cleatus AI deployment provides a different category of evidence — one where AI performance can be measured against consumer behavior at scale. Cleatus AI is a natural-language sports assistant built on Databricks that enables fans to query complex sports data conversationally. The result: a 2x higher query success rate compared to the preceding search experience (Databricks blog, 2025). Doubling the rate at which user queries return useful responses is a substantial improvement in a consumer context where abandonment is immediate.
The significance of this use case extends beyond sports media. Fan-facing AI deployments function as high-velocity testing environments for natural language interfaces: the query volumes are high, the user tolerance for failure is low, and the feedback loops are immediate. The patterns that emerge from these deployments — around query disambiguation, retrieval precision, and response calibration — are directly applicable to enterprise knowledge management, customer service, and sales enablement systems.
The Governance Data Gap
Against these performance metrics sits a troubling governance reality. An Economist Impact survey of 1,100 technology executives, cited in Databricks' AI Governance Framework, found that 40% of organizations consider their AI governance programs insufficient, and 53% of enterprise architects identify data privacy and security breaches as their primary concern (Economist Impact, cited in Databricks AI Governance Framework, 2024). Gartner's projection amplifies the stakes: organizations that operationalize AI transparency, trust, and security measures will achieve a 50% increase in AI adoption, goal attainment, and user acceptance by 2026 (Gartner, cited in Databricks blog, 2024). Databricks, citing Gartner directly, identifies AI trust, risk, and security management as the number-one top strategy trend in 2024 that will factor into business and technology decisions (Databricks AI Governance Framework blog, 2024).
The table below synthesizes the key performance outcomes across documented enterprise deployments.
| Organisation | Industry | Use Case | Key Metric | Source |
|---|---|---|---|---|
| Insulet | Medical Devices | ETL modernisation via Lakeflow | 97% TCO reduction; 12x faster processing; 83% fewer SQL queries | Databricks blog, 2025 |
| 7-Eleven | Retail | Agentic marketing assistant (13,000+ stores) | Compound AI system in production | Databricks blog, 2025 |
| Fox Sports | Media | Cleatus AI natural-language assistant | 2x query success rate | Databricks blog, 2025 |
| Mastercard | Financial Services | GenAI customer onboarding + human-in-the-loop agents | Automated support with regulatory oversight | Databricks blog, 2025 |
| NOV | Energy | Real-time predictive operations | 3+ TB daily data processing | Databricks blog, 2025 |
| Accenture clients | Cross-sector | Lakehouse Platform + Unity Catalog | 75% ETL time reduction; 14x ML training speed | Accenture, 2025 |
| Databricks (internal) | Technology | AI sentiment analysis on support tickets | 40% reduction in SLA misses | SupportLogic, undated |
| BBVA | Banking | AI-driven digital transformation | ~50M digital customers; 70% digital sales | Accenture, undated |
How Leading Organisations Are Responding
Accenture: Infrastructure Specialisation as Competitive Moat
Accenture has earned the Databricks Consulting and System Integrator Global Partner of the Year designation for seven consecutive years, from 2019 through 2025, and is the first consulting firm to earn all 10 official Databricks Brickbuilder Specializations (Accenture, 2025). This is not credentialling for its own sake. The specialisation portfolio — spanning data engineering, ML operations, security, and governance — reflects the firm's argument that sustainable enterprise AI returns require end-to-end platform mastery, not selective capability deployment.
What Accenture's client outcomes reveal is that the highest-ROI interventions cluster around data infrastructure modernisation rather than AI application development. The 75% ETL processing time reduction and 14x ML training acceleration achieved by Accenture clients come from foundational Lakehouse and Unity Catalog deployments — not from frontier model fine-tuning or bespoke AI application development (Accenture, 2025). This sequencing — infrastructure first, AI applications second — is the structural insight that distinguishes high-performing implementations from organisations experiencing pilot-to-production failure.
Accenture explicitly combines its cybersecurity expertise with Databricks' Unity Catalog fine-grained access control as the governance layer for enterprise-grade deployments (Accenture, 2025). Unity Catalog functions as the metadata and permissions management system across the Databricks Lakehouse, enabling column-level security, data lineage tracking, and audit logging — all prerequisites for regulated industry deployments.
7-Eleven: Compound AI Systems at Retail Scale
7-Eleven's deployment offers the clearest available example of what agentic AI workflows look like in a complex, distributed operational environment. The company's marketing assistant is a compound AI system — meaning it orchestrates multiple AI models and data retrieval processes to answer a single query or execute a decision task, rather than relying on a single model to do everything (Databricks blog, 2025). Across a network of 13,000+ stores, this system synthesises performance data that would be operationally impossible to analyse manually at that frequency and granularity.
The architectural significance is in the "agentic" design: the system can initiate actions, query data sources, interpret results, and generate recommendations without requiring human input at each step. This is the practical definition of intelligent process automation at enterprise scale — and it is live, not theoretical.
Mastercard: Human-in-the-Loop as Enterprise Standard
Mastercard's approach to LLM production deployment addresses the risk management challenge that most agentic AI discussions underweight. The company's architecture explicitly incorporates human-in-the-loop feedback mechanisms alongside its automated GenAI agents (Databricks blog, 2025). For a financial services organisation operating under regulatory scrutiny across multiple jurisdictions, this is not a compromise — it is the architecture required to deploy AI in consequential customer interactions.
The Mastercard model suggests an emerging design standard for enterprise AI deployment in regulated industries: full automation is applied where risk is low and queries are routine; human review is triggered by confidence thresholds, regulatory sensitivity flags, or novel query patterns. This graduated autonomy model allows organisations to capture AI efficiency gains while maintaining the oversight structures that regulators and risk functions require.
💡 Tip
Top-performing organisations in regulated industries do not design AI systems for maximum autonomy. They design for calibrated autonomy — defining exactly where human oversight adds necessary risk control, and automating confidently everywhere else.
The Hidden Risk: What Most Teams Get Wrong About Data Intelligence Use Cases
The dominant failure mode in enterprise AI deployment is sequence inversion: organisations invest in AI use-case development before their data foundations are stable enough to support them. Deloitte frames this directly, arguing that establishing a unified data platform is an imperative prerequisite before organisations can successfully pursue Generative AI advancement (Deloitte, 2025). The argument is structural: a GenAI application is only as reliable as the data pipeline feeding it, and a data pipeline is only as trustworthy as the governance framework surrounding it.
In practice, sequence inversion produces a recognisable pattern. An organisation deploys an LLM-powered application — a customer service chatbot, an internal knowledge assistant, a document summarisation tool. Initial results are promising in a controlled environment. Then the system encounters real production conditions: inconsistent data formatting, stale records, missing lineage documentation, access control gaps. The model's outputs degrade. User trust erodes. The project is deprioritised or abandoned, and the organisation concludes that the AI technology was not ready — when the actual failure was in the data infrastructure supporting it.
⚠️ Warning
Most enterprise AI pilots fail not because the model underperformed, but because the data infrastructure underneath it was insufficiently governed, integrated, or maintained. Organisations that treat infrastructure investment as a prerequisite for AI investment — rather than a parallel workstream — consistently report higher production success rates.
The governance gap compounds this risk. The Economist Impact finding that 40% of technology executives view their AI governance programs as insufficient is particularly concerning given the regulatory trajectory in both the EU AI Act and emerging US AI policy frameworks (Economist Impact / Databricks AI Governance Framework, 2024). Organisations that have accumulated AI deployments without corresponding governance structures now face retrospective remediation — a more expensive and disruptive process than building governance into the original architecture.
The second hidden risk is narrow ROI framing. Many organisations evaluate AI use cases through a revenue-uplift lens, overlooking the cost-destruction opportunity that the Data + AI Summit evidence consistently identifies as the primary value driver. Insulet's 97% TCO reduction from ETL modernisation does not appear on a typical AI business case template — yet it represents a more immediate and quantifiable return than most AI application deployments can demonstrate (Databricks blog, 2025).
📘 Note
The most underreported Databricks value driver is not AI capability — it is legacy infrastructure elimination. ETL modernisation, query consolidation, and batch-to-streaming migration consistently produce larger and faster ROI than application-layer AI investments, and they create the data foundation that makes those investments viable.
A Counterpoint Worth Considering: Where Infrastructure-First Has Limits
The infrastructure-first thesis is compelling — and the evidence behind it is strong — but intellectual honesty requires acknowledging where it does not tell the whole story.
Model selection still matters in precision-critical domains. The argument that infrastructure quality dominates model selection is well-supported in general enterprise settings, but it weakens significantly in domains where marginal model capability differences translate into consequential outcome differences. In clinical AI, legal document analysis, or financial risk modeling, choosing between a general-purpose LLM and a domain-fine-tuned model is not an afterthought once infrastructure is clean. The quality of the model's reasoning on ambiguous, high-stakes inputs can matter as much as the data pipeline feeding it. Organisations in these verticals should treat model evaluation as a parallel investment to infrastructure modernisation, not a downstream activity.
Governance investments do not always follow infrastructure gains automatically. The maturity model presented in this piece implies a natural progression from platform unification to governance operationalisation. In practice, governance is frequently treated as a compliance exercise layered onto an existing platform rather than an architectural design principle built into it. As Tanvir Kherada, Senior Director of Technical Solutions at Databricks, noted in the context of real-time support analytics: "By the time you wait for CSAT to be the deterministic factor to understand what the customer experience was like, it's too late" (SupportLogic case study). The same logic applies to governance: waiting until data quality failures or regulatory incidents force the issue is always more expensive than designing governance in from the start. The implication is that governance readiness may need to precede — or at minimum run in parallel with — infrastructure buildout, particularly in regulated industries, rather than emerging naturally from it.
Early-stage organisations face a different calculus. The infrastructure-first argument is most applicable to scaling enterprises with existing data sprawl. For early-stage organisations with relatively clean, limited data estates, a different sequencing question applies: at what point does the overhead of enterprise-grade Lakehouse architecture outweigh its benefits? The honest answer is that a Series A startup with two data pipelines and a single product database does not need Unity Catalog and a full Lakehouse migration. The governance and infrastructure investments described here are scaled to the complexity they are solving. Applying them prematurely is its own form of sequencing error.
These caveats do not invalidate the core argument. The evidence from Insulet, 7-Eleven, Mastercard, and Accenture's client portfolio is consistent and credible. But the infrastructure-first thesis is most powerful when applied to organisations with real data complexity — multiple systems, regulated environments, distributed operations — not as a universal prescription irrespective of organisational context.
A Framework for Moving Forward: The Data Intelligence Maturity Model
Based on the evidence across documented deployments, the path to scaled data intelligence use cases follows a consistent sequence. The model below structures that sequence into five stages, each with a clear prerequisite and diagnostic signal.
| Stage | Name | Core Activity | Prerequisite | Diagnostic Signal of Readiness |
|---|---|---|---|---|
| 1 | Foundation | Unify data into Lakehouse architecture | Cloud data infrastructure | Single platform replaces siloed data warehouses and lakes |
| 2 | Governance | Implement Unity Catalog, lineage, access control | Unified data platform | Audit logs, column-level security, and data lineage are operational |
| 3 | Intelligence | Deploy analytics, ML models, and BI workloads | Governed data foundation | ETL processing time reduced; ML training cycles accelerated |
| 4 | Automation | Build RAG systems, LLM integrations, intelligent process automation | Reliable, governed ML infrastructure | Production LLM applications handling real query volumes |
| 5 | Agency | Deploy multi-agent AI systems, compound AI, human-in-the-loop architectures | Proven automation layer with monitoring | Agentic systems operating autonomously within defined risk parameters |
Stage 1 — Foundation: The Deloitte argument is most applicable here. Organizations pursuing AI before completing Stage 1 are building on unstable ground (Deloitte, 2025). Lakeflow migration, as Insulet demonstrated, is itself a high-ROI investment independent of any AI ambition.
Stage 2 — Governance: Unity Catalog is the Databricks mechanism for operationalising governance. Accenture's integration of cybersecurity expertise with Unity Catalog's fine-grained access control represents the current standard for enterprise-grade governance in regulated industries (Accenture, 2025). This stage is not compliance overhead — it is the enabler of Stages 4 and 5. Critically, governance at this stage must be designed into the architecture — not applied as a retrospective layer — if it is to function as a genuine scaling enabler rather than a documentation exercise.
Stage 3 — Intelligence: This is where the Accenture performance benchmarks live. The 14x ML training acceleration and 75% ETL reduction are Stage 3 outcomes — achievable before any generative AI investment is made (Accenture, 2025).
Stage 4 — Automation: RAG systems, vector databases for enterprise search, and LLM integration use cases belong here. Fox Sports' Cleatus AI is a Stage 4 deployment — a governed, production LLM application with measurable success metrics (Databricks blog, 2025). Azure Databricks provides native AI Functions and generative AI engineering paths that support this stage (Microsoft Learn, 2025).
Stage 5 — Agency: 7-Eleven and Mastercard are operating at Stage 5. Multi-agent AI systems and human-in-the-loop architectures require the full stack below them — robust data foundations, governance infrastructure, reliable ML operations — to function at scale without systematic failure (Databricks blog, 2025).
Does Your Organisation Need Lakehouse Architecture? A Decision Framework
Before committing to a full Lakehouse migration, organisations should evaluate their actual data complexity against the investment the architecture requires. The checklist below helps distinguish organisations for whom the infrastructure-first thesis applies most urgently from those where a more selective approach is appropriate.
| Evaluation Criterion | Yes → Lakehouse investment likely justified | No → Assess scope before committing |
|---|---|---|
| Data sprawl | You operate 3+ separate data stores (warehouse, lake, operational DB) that require manual reconciliation | Your data estate is consolidated in 1–2 systems with clean integration |
| ETL overhead | Your team spends more than 20% of engineering capacity on ETL maintenance and pipeline debugging | ETL pipelines are stable, automated, and require minimal ongoing intervention |
| ML training velocity | ML model retraining cycles take more than 24 hours due to data preparation bottlenecks | Training pipelines run in under a few hours without data readiness delays |
| Governance exposure | You operate in a regulated industry (financial services, healthcare, energy) or handle PII at scale | Your data handling obligations are limited and existing controls are demonstrably sufficient |
| AI ambition | You have active plans for production LLM, RAG, or agentic AI deployments within 18 months | AI use cases are exploratory or limited to standalone analytics tools |
| Query inefficiency | Analysts regularly run redundant or overlapping SQL queries across multiple systems | Query consolidation is already in place; a single semantic layer serves most BI needs |
| Audit and lineage gaps | You cannot currently trace a data point from its source system to a model output or dashboard | Full data lineage is documented and auditable today |
| Vendor proliferation | You pay for 3+ separate data tooling vendors (ETL, orchestration, warehousing, cataloguing) | Your tooling stack is consolidated and contract overhead is manageable |
Scoring guidance: If you answered "Yes" to four or more criteria, the infrastructure-first investment case is strong and the ROI benchmarks from Insulet and Accenture's clients are relevant reference points for your business case. If you answered "Yes" to fewer than three, a targeted modernisation of specific pain points — rather than a full Lakehouse migration — may deliver better near-term returns with lower disruption.
📘 Note
This framework is intended as a diagnostic starting point, not a definitive investment decision tool. Organisations in regulated industries should weight the governance exposure criterion heavily regardless of their overall score.
Governance as Architecture: Building the Compliance Foundation That Scales AI
Governance is consistently identified as the primary bottleneck to enterprise AI scaling — yet it is also the investment most frequently deferred, scoped narrowly as compliance overhead, or designed as a retrofit to an existing platform. The evidence from production deployments suggests this sequencing is the source of significant downstream risk and cost.
The Economist Impact survey finding — that 40% of technology executives consider their AI governance programs insufficient — is not a data point about future risk. It describes the present state of most enterprise AI portfolios (Economist Impact / Databricks AI Governance Framework, 2024). For the 53% of enterprise architects who name data privacy and security as their top concern, the gap between concern and program maturity represents a live exposure, particularly as the EU AI Act's compliance requirements take effect and US regulatory frameworks continue to evolve.
Databricks' own framing, citing Gartner, positions AI trust, risk, and security management as the number-one top strategy trend of 2024 — not a functional responsibility for IT and legal teams, but a board-level strategic priority (Databricks AI Governance Framework blog, 2024). The implication is that governance readiness is increasingly a condition of enterprise AI deployment, not a consequence of it.
What Effective Governance Architecture Looks Like in Practice
The Mastercard and Accenture deployments together define the current standard for enterprise-grade AI governance in production environments. Three structural elements appear consistently across both:
1. Unified metadata and access control. Unity Catalog provides column-level security, data lineage tracking, and audit logging across the entire Databricks Lakehouse. Accenture's combination of its cybersecurity expertise with Unity Catalog's fine-grained access control creates a governance layer