I’ve spent 12 years watching companies dump millions into data platforms. I’ve seen the glossy slide decks from the giants like Capgemini and Cognizant, and the agile, tech-focused pitches from shops like STX Next. They all promise the moon. They all promise "AI-ready" architectures. But here is the reality check: most consulting firms are great at pilots and terrible at the 3 a.m. production crash.
When you are vetting a partner for a Lakehouse migration, stop listening to their "success story" slide. Instead, get on a reference call and grill them. If you don't know what to ask, you’re just buying a dream that will turn into a technical debt nightmare. Here is how to conduct a reference call that separates the engineers from the PowerPoint jockeys.
The Lakehouse Consolidation: Why You’re Really Here
Before you talk to their references, you need to be clear on what you’re buying. You aren't just moving files to Databricks or Snowflake. You are consolidating fragmented legacy systems into a unified Lakehouse architecture. The goal is to eliminate the "ETL tax" and get your data scientists and BI analysts working on the same source of truth.
If a firm tries to sell you on a "hybrid, fragmented, polyglot persistence" strategy without a clear path to consolidation, run. You are looking for a firm that treats your platform as a product, not a project.
The "3 a.m. Test": Essential Questions for Reference Calls
When you get the firm’s past clients on the phone, don’t ask if they "did a good job." Of course they’ll say yes. Ask these questions instead.
1. "What breaks at 2 a.m.?"
I ask this every single time I review an architecture. Every platform has a breaking point—maybe it’s a specific connector failure, a credential rotation issue, or a cost-spike on an unoptimized Spark cluster. A good partner knows exactly where the weak points are and has documented the remediation. If the reference says, "Oh, it never breaks," you aren't talking to an engineer; you're talking to a marketing plant.
2. "Show me the production history, not the pilot report."
Many firms treat the "go-live" date as the end of the contract. That is when the real work begins. Ask the reference: "How did the firm handle the first major data quality incident six months post-launch?" If they don't have a story about troubleshooting a production pipeline or an observability failure, they haven't run a real production workload.
3. "How are you handling the semantic layer?"
If they tell you that the semantic layer is "automatically handled by the tool," they are lying. Whether you’re using dbt or native Databricks/Snowflake features, the semantic layer requires manual mapping and strict governance. Ask the reference how they prevent "metric drift" between the finance dashboard and the data science team’s notebooks.
The Governance and Quality Rubric
Too many teams leave governance, lineage, and data quality as a "Phase 2" item. In a Lakehouse, if you don't build it in from the start, you’re just building a swamp with better marketing. Use this table to score the firm's answers during your reference calls.
Area The "Red Flag" Answer The "Pro" Answer Data Governance "We use role-based access control." "We implemented column-level security and automated PII tagging using Unity Catalog/Snowflake tags." Lineage "It's built into the UI." "We enforced dbt lineage tests and documented upstream/downstream dependencies before the first pipeline hit prod." Data Quality "The users tell us when data is wrong." "We have automated unit tests for data (e.g., 'null' counts) that block production promotion if thresholds are exceeded."
Production Readiness vs. "Pilot Success"
A "pilot" is a sandbox. A "production" Lakehouse is a living, breathing ecosystem. When talking to references provided by firms like Cognizant or Capgemini, you need to determine if they are scaling a pilot or if they actually built a durable system. Ask these specific questions:
- Post-Launch Support: "Did the partner leave you with a knowledge transfer document, or did you have to hire them back for every minor change?" Delivery Proof: "Can you describe the process for deploying a new feature to production? Was it automated, or was it a manual 'copy-paste' job?" Cost Governance: "How did the firm help you optimize costs? Did they leave you with a bloated warehouse/cluster configuration, or did they implement auto-scaling and lifecycle policies?"
The "AI-Ready" Myth
I hear this phrase every day. "We are building you an AI-ready architecture." It means absolutely nothing. AI readiness is a function of data quality and lineage, not just the underlying https://www.suffolknewsherald.com/sponsored-content/3-best-data-lakehouse-implementation-companies-2026-comparison-300269c7 cloud storage.
Ask the reference: "How long did it take your data scientists to get access to production-grade, cleaned data after the platform was delivered?" If the answer is "they’re still struggling to find the right tables," the firm failed. A real Lakehouse makes data discoverable through a well-maintained catalog. If the firm didn't build the catalog, they didn't build a Lakehouse—they built a data silo with a fancy name.
Final Checklist Before You Sign
Before you commit to a long-term engagement, make sure your potential partner checks these boxes during the reference call:
Accountability: The reference confirms the firm stayed through the first major production incidents. Technical Debt: The reference can point to specific areas where the firm warned them about trade-offs (e.g., performance vs. cost). Operational Maturity: The firm didn't just deliver code; they delivered a CI/CD pipeline, monitoring, and an alerting strategy. Staff Augmentation vs. Outcome: The firm focused on training the internal team, not making them permanently dependent on the consultancy.Don’t be afraid to be the difficult person on the reference call. The firm is asking for your budget; you are entitled to know if they can keep your data platform from falling apart at 2 a.m. If they can’t answer the hard questions, keep looking.