Can Claude Replace Your Data Team? What Executives Learn After Six Months of Doing It Themselves

Over the past six months, we’ve watched many of our customers venture into data warehousing on their own. The pattern is almost always the same: a founder or an operations leader connects Claude to Snowflake, starts asking questions, and begins building reports without a data team.

I want to be clear about one thing up front: that isn’t wasted time. Teams that do this learn more about their own data than most leadership teams ever do. They start defining the basics, like what counts as a project or a customer. They discover on their own that pointing Claude at a clean, modeled table is faster and cheaper than letting it stitch raw tables together every time.

They also discover the bill. Without a solid data foundation, Claude burns through credits and takes far longer to answer. And there’s a risk most of them haven’t priced in yet, which my colleague Ovais Siddiqui sums up in one line: Claude will always try to give you an answer.

That’s what this article is about.

Why executives are using Claude instead of hiring a data team

I haven’t found a survey that measures exactly this behavior: executives doing data engineering with Claude instead of hiring. But the evidence around it all points the same way.

• Senior leaders are the heaviest users of unapproved AI. In a TrustedTech survey reported in May 2026, 62% of UK senior decision-makers said they use shadow AI tools, versus 31% of employees below decision-maker level.
• Executives are building their own software. The New Stack profiled CEOs who now vibe-code their own tools with Claude and Cursor. OutSystems CEO Woodson Martin summed up the mindset: “I was tired of explaining it to somebody who was supposed to build it for me.”
• The same reflex is hitting other professions. Ravical’s 2025 survey of 500 UK accountancy decision-makers found that 64% of businesses now turn to ChatGPT for advice before their accountant, and 54% of firms have had to correct advice their clients got from it.
• Hiring decisions are moving before the results do. Thomas Davenport and Laks Srinivasan surveyed 1,006 global executives in December 2025 and found that AI-related layoffs are happening almost entirely in anticipation of AI’s impact, not its performance. And in July 2026, Gartner reported that 22% of CHROs say at least one business leader in their organization has stopped hiring for entry-level roles because of AI automation.

The logic is understandable. Claude is available tonight. A data engineer takes months to hire. A consultant needs a scope and a budget. Why not try it yourself?

What Claude gets right: text-to-SQL accuracy in 2026

We use Claude every day at Data-Sleek: to write and review code, to automate infrastructure changes, to draft documentation, and, yes, to help research this article. It’s a remarkable tool, and it’s improving fast.

dbt Labs first ran a text-to-SQL benchmark in 2023: 11 realistic business questions against an insurance dataset. On raw, normalized tables, GPT-4 wrote correct SQL 32.7% of the time. In the April 2026 update, Claude Sonnet 4.6 reached 64.5%. That is real progress.

So the question isn’t “Claude or no Claude.” The question is: Claude on top of what?

5 risks of connecting Claude to Snowflake without a data foundation

1. Claude always gives an answer, even when your raw data can’t support it

Look at the rest of that dbt benchmark. Same model, same 11 questions, different foundations: 64.5% correct when Claude wrote SQL on raw tables, 90% on modeled tables, and 98% with a semantic layer on top of the modeled tables.

The model didn’t change. The foundation did.

Even more important is how each approach fails. When the semantic layer can’t answer a question, it returns an error. When text-to-SQL can’t, in dbt’s words, it “will cheerfully give you a wrong number.”

A wrong number that looks wrong is harmless. A wrong number that looks right gets copied into a board deck, a lender update, or a bid. As Ovais puts it: the AI connects two things that shouldn’t be connected, the result looks good, someone presents it, and that’s how decisions get made.

2. AI sycophancy: why Claude agrees with you (what I call “AI empathy”)

AI models are trained to be helpful, and helpful easily slides into agreeable. Researchers call it sycophancy. A Stanford study published in Science in March 2026 tested 11 leading models, Claude included, and found they validated users’ actions 49% more often than humans did. Users rated the flattering answers higher and were more willing to come back.

That study looked at personal advice, not spreadsheets. But the mechanism doesn’t care about the topic. Ask “why did our margins improve this quarter?” and the model will find you reasons. It rarely stops to ask whether they actually did.

Now add the org chart. When the CEO is the one holding the AI’s answer, very few people downstream will challenge it. A Microsoft Research and Carnegie Mellon study of 319 knowledge workers found that “confidence in AI is associated with reduced critical thinking effort.” The more we trust the tool, the less we check it.

3. Context limits: Claude can’t remember business definitions nobody wrote down

I used to say AI simply can’t hold the full context of a business. That’s no longer quite true: today’s frontier models, Claude included, can work with up to a million tokens of context, and they have memory features.

But two problems remain. First, more context isn’t the same as better use of it. Chroma’s 2025 “context rot” research tested 18 models and found that performance degrades as input grows, even on simple tasks. In their words, “models do not use their context uniformly.” Long working sessions get summarized and compressed, and details fall out.

Second, and this is the real issue, most of your business context isn’t written anywhere. What counts as a closed project? Which vendor record survived the last system migration? Is revenue recognized at contract, at invoice, or at completion? Those answers live in people’s heads. Claude can only remember what someone wrote down.

That is what a semantic layer is: your company’s definitions, written down once, in one place, for every human and every AI that asks a question.

4. AI slop: too many data models, reports and dashboards

I’ll confess this one first. Over the past year I’ve produced more with AI than ever before, and I’ve lost track of a good part of it. It’s so easy to start the next thing that you stop finishing the last one.

Researchers now have a name for what this does to teams: workslop, “content that appears polished but lacks real substance.” In a BetterUp Labs and Stanford Social Media Lab study published in HBR, 41% of workers said they had encountered it, and each instance cost nearly two hours of rework.

In data, it looks like this. Ovais is seeing AI-generated changes that add 150 to 200 new data models at a time, each one longer than it needs to be, full of comments and tests that look thorough and aren’t. Or views built on views built on views, something I learned to avoid early in my career as a database administrator, because the complexity hides inside them and nobody can debug it. At 500 models, nobody on your team understands the system anymore. You need an agent to fix what an agent built.

5. AI governance risk: your data and decisions leave your controls

When Claude queries your warehouse from the outside, the results travel to a third-party model provider. That may be fine under your agreements. But it should be a governance decision, not a side effect. And the people most likely to skip that decision are at the top: in Teramind’s June 2026 report, 69% of senior leaders said they prioritize speed over security when using AI tools, versus 37% of frontline staff.

There’s a second control question: what happens when the answer is wrong? Before you rely on AI for decisions, ask two things. If this number is wrong, what does it cost? And who else will see it: vendors, architects, lenders, partners?

AI that runs inside the warehouse on curated semantic views (Snowflake Intelligence is one example) keeps the data under the same governance and access controls as everything else. It’s a narrower tool by design. For leadership questions, narrow is a feature.

Personal AI productivity vs. process automation: where the ROI is

This is the distinction I keep coming back to with clients.

Claude makes individuals faster. That’s real, and it’s worth paying for. But personal productivity rarely shows up on a P&L. MIT’s 2025 GenAI Divide report found that 95% of enterprise generative AI pilots delivered no measurable P&L impact. Gartner’s July 2026 survey found that 95% of organizations had implemented AI in some capacity, yet only one in five had realized significant or transformational value.

Where does the value come from? From changing a process: AP approvals, job costing, change orders, vendor performance, claims triage. The slow, repetitive, cross-team work that eats weeks. And processes don’t run on one person’s chat history. They run on shared data that everyone trusts. Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. If you’re not sure where you stand, an AI readiness assessment is a good place to start.

If the warehouse can’t answer the question, trust evaporates. And once people stop trusting the warehouse, they go back to their spreadsheets.

How to build an AI-ready data foundation: a 6-step playbook

If you’ve been running your own Claude-and-warehouse experiment, you’re not starting from zero. You’re starting with a better understanding than most. Here’s how I’d use it.

1. Start with the vision, not the tables. Where does the company need to be in three years, and which two or three processes are in the way? That decides what to build first.
2. Build one subject area properly. Pick the one with the most business value (project profitability, for example), model it as facts and dimensions, prove the numbers, then replicate the pattern. Don’t build everything at once. Our data warehouse implementation plan walks through the sequence.
3. Write down your definitions. Project, revenue, margin, active customer. Put them in a semantic layer so every person and every AI uses the same ones.
4. Split the audiences. Leadership asks questions through a governed interface on curated data. Builders use Claude for modeling, testing, and automation.
5. Keep a human in the loop on everything that ships. Small changes, reviewed by someone who knows the business. If nobody can explain a model, it doesn’t go to production.
6. Price the risk. For every AI-assisted decision, know what a wrong answer would cost and who would see it.

And yes, use Claude to build the foundation. In the dbt benchmark, the three models that took accuracy from 64.5% to 90% and above were written by an LLM. The difference is that someone with context decided what needed to be built, and checked the result.

The bottom line: Claude is a carpenter, not a foundation

In French we say: on ne construit pas une maison en commençant par le toit. You don’t build a house starting with the roof. Claude is an excellent carpenter. It is not a foundation.

If you’ve spent the last few months doing your data team’s job with Claude, you’ve earned a sharper view of what your company needs. The next step is turning what you learned into a foundation the whole business can trust. That’s the work our data warehouse consulting team does every day: modeled data, a semantic layer, and governance your team can run, so the answers your AI gives are ones you can put in front of a board.

Sources

1. dbt Labs, Semantic Layer vs. Text-to-SQL: 2026 Benchmark Update (Apr 2026)
2. TechCrunch, Stanford study on AI sycophancy published in Science (Mar 2026)
3. The Register, Microsoft Research / Carnegie Mellon study on generative AI and critical thinking (Feb 2025)
4. Chroma, Context Rot (Jul 2025)
5. Harvard Business Review, AI-Generated “Workslop” Is Destroying Productivity (Sep 2025)
6. CMOtech UK, TrustedTech shadow AI research (May 2026)
7. HRD, Teramind Shadow AI Behavior Report (Jun 2026)
8. The New Stack, Meet the executives vibe-coding their own tools (Apr 2026)
9. Ravical, Two thirds of businesses now turn to ChatGPT before their accountant (Oct 2025)
10. Harvard Business Review, Companies Are Laying Off Workers Because of AI’s Potential, Not Its Performance (Jan 2026)
11. Gartner, AI automation is reducing some entry-level hiring (Jul 2026)
12. Gartner, Lack of AI-Ready Data Puts AI Projects at Risk (Feb 2025)
13. Legal.io, summary of MIT NANDA’s The GenAI Divide (Aug 2025)

FAQ: Claude, Snowflake and semantic layers