Snowflake VP of AI Baris Gultekin on Bringing AI to Data, Agent Design, Text-2-SQL, RAG & More
Summary
Snowflake’s enterprise-AI thesis is that value begins with bringing inference to governed data, not copying sensitive information into a model vendor’s environment. Customers preserve residency, permissions, and governance while model providers operate as subprocessors without saving state. Baris Gultekin’s practical standard is “high-quality AI agents that create positive ROI for customers quickly”—a capability he says exists today.
The near-term demand catalyst is making the 80% to 90% of enterprise data that is unstructured usable alongside conventional tables. Contracts, research PDFs, images, and tables can now be extracted, classified, searched, and analyzed in natural language; a wealth-management agent, for example, can combine structured stock data with PDF equity research. “The value you get from data is increasing.”
Reasoning models have moved text-to-SQL from fragile demo to broad business-user deployment over the last six months to a year. The hard part remains semantics: “revenue” has one correct answer, yet enterprises may have thousands of tables, thousands of columns, obsolete fields, and definitions embedded in tribal knowledge. Snowflake Intelligence is nevertheless described as Snowflake’s “fastest-growing product,” replacing some analyst requests that previously took a week.
Model economics split sharply by workload scale: frontier models can simplify one-off document work, while specialized models win at hundreds of millions of documents. Labenz observes that Claude 4.5 Opus or Gemini 3 may solve a few messy forms off the shelf, but Snowflake’s extraction model is “multiple orders of magnitude smaller,” faster, and cheaper for repeatable fields. Fine-tuning is not dead; it becomes attractive when proprietary data, throughput, or cost is sufficiently important.
Frontier-model quality matters, but residency restrictions, optimized prompts, and accumulated workflow design make production choices stickier than benchmark tables imply. Snowflake productizes side-by-side evaluations, yet Gultekin sees limited switching because applications can contain “thousands of lines of systems and prompts” tuned to one model. As quality converges, differentiation migrates upward into products such as ChatGPT and Claude Code.
Enterprises want bounded, optimized agents for business users and more open-ended agents for technical builders. Snowflake’s sales assistant serves 5,000 sellers through one governed interface, but returns answers based on each seller’s individual permissions. Multi-agent handoffs are emerging across organizational and platform boundaries, although customers are still mostly trying to “get this one agent right”; MCP is used now, while A2A remains early.
Trust is an entire operating stack rather than a single accuracy score. It spans security, granular access controls, governance, retrieval quality, tool selection, evaluations, monitoring, feedback, and user-facing verification; the same agent can answer two employees differently because each may see different data. Guardrail checks and LLM-as-judge evaluations are available and increasingly used, while Gultekin offered no claim that model-weight security is formally proven or absolutely “airtight.”
Open formats weaken lock-in, forcing Snowflake to compete on execution, while Gultekin expects value to accrue to chips, models, and horizontal applications rather than bespoke business-logic vendors. Snowflake supports Iceberg so customers can retain control of their data, and Gultekin’s call is that “the middle will erode and the sides will continue expanding.” Coding agents accelerate that pressure by turning working prototypes into the new product-development unit. On proposed Q2 drop-in knowledge workers, he does not endorse the timing but sees tightly scoped autonomous roles as a natural progression.
Deep dive
1. Snowflake brings AI to governed enterprise data
Gultekin’s Snowflake 101 starts with data gravity: enterprises consolidate structured data to secure, govern, and analyze it across silos, powering large-scale analytics and BI dashboards. Because customers do not want to replicate sensitive information repeatedly, the platform’s AI posture is to “bring AI to run next to data.”
Asked how “AGI-pilled” Snowflake is, Gultekin rejects grand framing for an operational one. Snowflake serves large enterprises, so the objective is “high-quality AI agents that create positive ROI for customers quickly”—something he says “can happen today, and it is happening today.”
The before-and-after is less about replacing analytics than widening its interface and corpus. Existing structured-data workloads remain, while AI makes document extraction, classification, natural-language querying, and agents dramatically easier to build directly beside the analytical engine.
2. Unstructured data expands the usable enterprise corpus
Gultekin estimates that “80 to 90% of all data is unstructured data.” Much of it previously sat outside databases because organizations lacked economical processing methods; AI now makes those files usable both by extracting structure and by retrieving their contents directly.
His contract example carries the first workflow: across thousands or hundreds of thousands of documents, an enterprise can ask which contracts mention a particular provision, count them, or identify which contracts in a category expire soon. Pipelines once requiring dedicated classifiers and extractors become straightforward analytical operations.
The richer use cases combine both data types. A wealth-management agent still needs structured market information about stocks, but it also needs equity research trapped in PDFs; useful real-world agents must retrieve and reason across both rather than declaring either SQL or RAG sufficient.
3. Reasoning models made text-to-SQL deployable
Text-to-SQL historically failed because business questions tolerate little error. If someone asks, “What’s my revenue?” Gultekin notes that “there’s only one answer,” yet revenue and profit definitions vary, their physical representations change, and real systems contain thousands of tables with hundreds or thousands of columns.
The material inflection came during roughly the last six months to a year: stronger reasoning models plus easier ingestion of business semantics produced “pretty substantial gains in quality.” Gultekin now describes text-to-SQL as high-quality enough to deploy broadly beyond trained analysts.
Snowflake Intelligence packages that capability as an agent platform for business users and, according to Gultekin, is Snowflake’s fastest-growing product. Employees can ask questions directly instead of requesting analysis from someone familiar with the schema and waiting perhaps a week for a response.
Nathan Labenz’s original caveat survives the optimism: models need the organizational knowledge that distinguishes the live column from the obsolete one. The model’s SQL syntax is not the decisive bottleneck; discovering and maintaining the enterprise’s intended meaning is.
4. The semantic layer is becoming portable infrastructure
Snowflake uses reasoning models to help construct semantic models from table and column names, underlying data, metadata, historical queries, and connectors to BI dashboards. Those existing artifacts encode definitions and usage patterns that would otherwise remain scattered among analysts.
Its Open Semantic Interchange initiative aims to let enterprises create a semantic model once and reuse it across participating platforms. Snowflake is working with Tableau, Omni, other BI systems, and technology providers, but Gultekin stresses that the exchange format remains in active, early development.
Labenz’s pushback—worth keeping—is that faster software creation may turn former complements into frenemies: Tableau or Snowflake could expand into each other’s territory, and whichever platform has the deepest customer hooks might “colonize” adjacent categories.
Gultekin welcomes the pressure: “The silos are coming down,” differentiation is shrinking, and open standards reduce lock-in. His rebuttal is that “the pie is growing” and no company can do everything; specialization persists, but speed of execution and product quality matter more once walls fall.
5. RAG quality is rising as its plumbing disappears
Snowflake’s RAG stack inherits technology from Neeva, the web-scale search company where Gultekin worked before its acquisition. He identifies embedding quality, chunking, hybrid search, and reranking as the core quality determinants rather than treating vector retrieval as a single interchangeable component.
Complex documents remain a practical frontier: PDFs mix tables, images, multiple columns, and irregular layouts. Builders must decide what to extract and whether text or multimodal embeddings fit, but Gultekin says more of that model and pipeline selection can now be automated.
That simplification unlocks what Snowflake calls “analytical agentic document analytics.” Given quarterly reports spread across multiple PDFs over 10 years, an agent can extract each period’s figures and calculate an average—combining retrieval, document understanding, and analytics rather than merely returning passages.
The direction is therefore dual: RAG is becoming both higher quality and easier to assemble, while agents are moving beyond answering document questions into repeated extraction and cross-document computation.
6. Model size should follow workload economics
Labenz observes that Claude 4.5 Opus or Gemini 3 can now understand a handful of scanned vehicle-sale forms without the specialized embedding, table-extraction, and parsing stack that might have been necessary 18 months earlier. The engineering burden falls even if inference cost rises.
Gultekin’s qualification is scale: “If you are going to process…hundreds of millions of documents, you’re not going to use Claude.” The constraints are “cost and throughput”—both how much inference costs and how long a frontier service takes to complete an enormous batch.
Snowflake therefore built and fine-tuned a document-extraction model “multiple orders of magnitude smaller” than frontier LLMs. When the job is a stable set of fields, specialization makes processing substantially faster and cheaper; powerful generality is unused capacity.
Fine-tuning is not dead in this account. Start with frontier models and retrieval, then consider a custom model when an enterprise has abundant proprietary data, a task the base model has not seen, or binding throughput and cost requirements; Cursor’s autocomplete model is Gultekin’s scale-driven example.
7. Enterprises retain model choice, including over Chinese weights
Snowflake offers OpenAI, Anthropic, Gemini, Meta, Mistral, DeepSeek, and other proprietary or open models. In products where the model is embedded, customers may not choose directly; on the broader platform, they select among the models their organization has approved.
Whether customers will use Chinese model weights appears idiosyncratic rather than industry-specific. Even within technology, some accept them and others “absolutely” will not; Gultekin argues the market is competitive enough that declining one ecosystem does not eliminate all capable alternatives.
No single model “dominates it all,” in either proprietary or open source. That abundance gives Snowflake room to broker choice, but actual deployment remains bounded by each customer’s internal policies and AI-governance board.
8. Security architecture determines where inference can run
Snowflake began this architecture roughly two and a half years earlier after hearing that customers would not move data outside its security boundary. Keeping inference beside data preserves existing governance, avoids replication, and shrinks the number of attack vectors.
Through agreements with model and cloud providers, inference runs within the relevant protected environment while the providers remain subprocessors. Gultekin says “there is no state that’s saved in any of these models,” with legal and technical arrangements supporting residency; cloud providers matter because they supply both the physical infrastructure and inference stack.
Asked directly about an xAI or Grok relationship, Gultekin did not answer the partner-specific question and instead explained the deployment boundary. Labenz then pressed on why distributing frontier weights across clouds has not produced visible leaks, including the possible role of trusted execution environments.
Gultekin’s honest limit: security is not his expertise, and he offered no formal-methods or cryptographic proof. He said access to weights is restricted by design and the execution environment can do little beyond inference, while retaining the essential hedge: in security, one can never promise something is completely “airtight.”
9. Production model choices are stickier than benchmarks suggest
Gultekin says “the differences between these models are not large,” although each keeps improving. The first filter may be residency rather than quality: an enterprise committed to Amazon’s boundary, for example, may reject models reachable only through Azure or directly through a provider.
Snowflake productizes model comparison rather than treating it mainly as consulting. Customers can choose models, run side-by-side evaluations, and assess fit; residency, quality, latency, and cost all enter the decision. Reasoning can change an apparent price or speed advantage because the comparison shifts once reasoning is added.
Actual production switching remains limited. Prompts and surrounding systems become optimized for one model, and an application may contain “thousands of lines of systems and prompts”; unless the quality gain is large, migration costs outweigh small benchmark deltas. Gultekin likens the inertia to Google retaining users after Bing became good enough.
Labenz asks whether Snowflake should “commoditize your complements” and minimize those switching costs. Gultekin declines that objective: choice matters, but Snowflake starts with data and focuses on retrieval quality so each approved model receives the best possible context. As models converge, he expects application products and habits to become the next differentiation layer.
10. Agent autonomy should match the user
Labenz places agents between linear workflows and Claude Code’s “choose your own adventure,” warning businesses not to assume an open-ended coding-agent pattern is ideal everywhere. Gultekin agrees that persona decides: builders benefit from flexibility, while business users need a bounded interface optimized for trusted answers.
Snowflake Intelligence resembles ChatGPT over company data. Its internal sales assistant went to 5,000 sellers and answers questions about renewals and books of business; a coding agent integrated into the platform serves data engineers and analysts who need freer data exploration, pipeline construction, and code generation.
Even Claude Code uses skills and instructions, so “one agent” still contains modular structure. Large enterprises also have different teams and platforms: a Salesforce-based sales agent may need to coordinate with another organization’s agent, making handoffs as much an ownership requirement as a performance technique.
That multi-agent world remains early because customers are still focused on getting one agent right. MCP sometimes bridges systems by exposing another agent-like capability as a tool; Snowflake does not yet support A2A, although customers increasingly request some form of agent-to-agent communication and open interoperability.
11. Enterprise trust is a full-stack property
Gultekin calls trust a “full stack” spanning security, governance, answer quality, evaluations, and ongoing monitoring. Running beside governed data means an agent inherits granular access controls rather than reconstructing them separately in every application.
The same sales agent can serve two employees while returning different customers because each user only receives data they may view. The same principle applies to an HR or manager query about salary: underlying agent, different authorization context, and no expansion of the user’s existing privileges.
Reliability then depends on retrieval and trajectory: did the system fetch the right context, select the right tool, and follow the correct route? Snowflake evaluates and monitors those components, gathers feedback, and can show that an answer was verified by an owner—trust rendered directly in the interface.
Companies generally begin with internal agents, where the bar is lower than for customer-facing automation. Snowflake also offers checks for hate speech, violence, and other violations, while policy instructions improve as models follow directions better. Practical interpretability today mostly means observing model and agent behavior through evaluations and monitoring—not mechanistic inspection.
12. Coding agents are changing product development before whole jobs
Gultekin believes coding agents have crossed a capability threshold, but rejects the leisure forecast: greater capacity leads people to “do a lot more,” especially in an industry where everyone is sprinting. Better reasoning and tool use should transfer beyond coding, though he avoids a precise 2026 capability prediction.
He says product-management workflows can invert after 20 years in the field. Instead of designing and hardening a conventional UI first, a team can write a skill in a day, put it before customers, gather feedback, and only then solidify the experience. Labenz’s summary: “Working prototype is the coin of the realm.”
Asked to quantify Snowflake’s gain through features shipped or cycle velocity, Gultekin emphasizes the changed philosophy rather than supplying a metric. The impact is behavioral: deployment, testing, and even the definition of the initial product surface have changed.
On Q2 drop-in knowledge workers, his answer is deliberately scoped. Autonomous entities are a “natural progression,” but he does not think a colleague that can do “anything and everything” is ready; clearly bounded roles might work, provided agents receive the relevant company data and action tools. He does not say whether Q2 is the timing.
13. Open formats transfer the moat from lock-in to execution
AI makes data more accessible and increases the number of economically useful analyses. Gultekin cites S&P analyzing earnings calls to detect whether CEOs answer analyst questions directly, indirectly, or had already answered them in prepared remarks—and using that signal as alpha for stock selection.
Labenz’s bear case starts with Intercom: agents compiled roughly 100 documentation pages into a 600-page Google Doc, condensed it, wrote an export, and analyzed the resulting data from a few prompts. His conclusion is that APIs and AI may make leaving established systems dramatically easier.
Gultekin embraces that direction. Snowflake supports Iceberg and open formats so customers can control storage and use Snowflake as the processing engine; absent lock-in, products win on performance and cost. Asked whether dissatisfaction measured over a year could shrink to a quarter, he says competition is “a great incentive” to deliver faster.
His value-accrual call is that “the middle will erode and the sides will continue expanding.” Chips such as NVIDIA, model providers, and differentiated applications such as Cursor retain value; bespoke vendors encoding narrow business logic are more exposed because agents can generate that logic quickly. He agrees with Labenz’s reading that, over time, horizontal applications may be relatively stronger than dedicated vertical software.
14. Enterprise intelligence remains retrieval-led—for now
Labenz asks why companies do not create deeply pretrained “GE GPT,” “3M GPT,” or “Pfizer GPT” systems from their accumulated private history, noting that he was told by Databricks CEO Ali Ghodsi that Databricks had killed MosaicML’s continued-pretraining product. Gultekin answers with the web-search analogy: intelligence can decide when to retrieve current knowledge instead of embedding all knowledge in weights.
Retrieval through text-to-SQL and RAG is cheaper, easier to update, tunable, and automatically benefits as frontier models improve. Gultekin expects most enterprises to keep knowledge outside model weights; task-specific small models remain the exception when proprietary data, low latency, high throughput, or cost warrants specialization.
Labenz’s pushback is personal: he can search his own Gmail or Drive better than a smarter model because he recognizes the definitive document from lived familiarity. Gultekin concedes that “intuition is not really understood”; training on private data might yield consistency and require less steering, but would not eliminate the need for fresh retrieved information.
Gultekin watches Snowflake’s internal quality and latency benchmarks more than ARC-AGI or other public scores, while tracking tabular foundation models and rapid forecasting. His closing call is urgency with foundations: AI adoption is now “a race,” and enterprises need to develop intuition through native use, break down data silos, make tools AI-ready, and build products easy enough to deploy at scale.