Reasoning Models Just Made the Other 80–90% of Enterprise Data Usable

Baris Gultekin came to Snowflake with current CEO Sridhar Ramaswamy when the company acquired Neeva, an AI-powered search engine, in May 2023. He now runs Snowflake's AI products from the awkward middle of two worlds: frontier model capability on one side, and the security, governance, and reliability bar of large enterprises on the other. Host Nathan Labenz opens there, asking for a Snowflake 101 and then how "AGI-pilled" the company actually is.

Gultekin's answer is deliberately unromantic. Snowflake is a data platform that now calls itself an AI data cloud: customers land data there to secure it, govern it, and run large-scale analytics and BI. Because data has gravity, those customers do not want to replicate it into yet another system. They want AI to run next to the data. Being AGI-pilled, for him, means shipping high-quality agents that create positive ROI quickly — something he says is already happening.

Before generative AI, most Snowflake workloads were structured: transactions, customer tables, financials. The platform's value was breaking silos so analysis could run across sources. What changed is unstructured data. Contracts can be queried for a clause, an expiration window, or a category count; classification and extraction pipelines that used to be custom engineering are now close to built-in; and a natural-language interface can open the warehouse to people who will never write SQL.

Labenz cites a pattern he has heard from vector-database founders: most of what is now being indexed never lived in a database at all. Gultekin agrees, and he puts a number on it. "80% to 90% of all data is unstructured data," he says, and because it was expensive to process, it was treated as the least usable part of the estate. That has flipped. Companies can extract structure and run analytics later, or they can talk to the files directly through retrieval and agents.

The two patterns usually run together. Gultekin's example is a wealth-management agent that still has to look up prices in a structured store and also read equities research sitting in PDFs. Combining both is what real use cases require. It is also why he keeps repeating the slogan: bring AI to the data, rather than sending sensitive records out to a model.

Labenz asks where text-to-SQL actually stands. Gultekin is blunt that it used to be very hard. Ask "what's my revenue" and there is only one acceptable answer, so the margin of error is tiny and quality expectations are brutal. Models also need a semantic map: what revenue means, how profit is defined, and how those ideas are modeled across thousands of tables and hundreds of thousands of columns.

Tribal knowledge makes it worse. A column may be deprecated and never deleted; the people who know that are not in the schema. The turn, Gultekin says, came in the last six months to a year, as reasoning models got substantially better and as it became easier to land semantics on the platform. Quality gains were large enough that Snowflake Intelligence, the company's agent product for this job, can now be deployed broadly — to business users, not analysts. It is Snowflake's fastest-growing product, because people who used to wait a week for someone who knew the warehouse can now ask in natural language.

Building that semantic model is no longer a giant interview project by default. Reasoning models can draft it from table and column names, the data underneath, connectors into BI dashboards, and historical queries. Snowflake has also launched Open Semantic Interchange with Tableau, Omni, and other BI vendors, still early, so a semantic model built in one platform can be reused in another. Gultekin likes the direction because lock-in falls and customers keep the definitions they already paid to encode. He also rejects a fixed-pie view of the ecosystem: more is possible, companies will still focus near their expertise, and speed of product execution becomes the differentiator.

Snowflake's RAG stack is downstream of Neeva, the web-scale search engine it acquired. Gultekin puts embedding quality first, then chunking, hybrid search, and reranking. PDFs remain messy — images, tables, multi-column layouts — and extraction accuracy flows straight into retrieval. Teams still have to choose multimodal versus text embeddings, but more of that specialist work can now be automated, which is what unlocks new uses.

He describes what Snowflake calls analytical, agentic document analytics. Imagine thousands of quarterly-result PDFs over ten years, a question like "what's the average revenue," and a system that extracts the numbers across files and aggregates them. That was nearly impossible before. It is now in reach, which is his summary of RAG as a category: higher quality, simpler to build, and newly able to support agentic analysis rather than single-document lookup.

Labenz has watched scanned car-sale forms go from specialist embedders and table extractors to Claude Opus or Gemini 3 solving the document off the shelf — higher inference cost, far less engineering time. Gultekin splits the cases. A handful of documents can go to a frontier model. Hundreds of millions cannot, because cost and throughput both break. Snowflake therefore trained its own document-extraction model, multiple orders of magnitude smaller than an LLM, to pull specific fields quickly and cheaply.

That is the middle of a Pareto frontier. Frontier models sit on the expensive, general left; platform-scale specialists sit in the center, amortized across customers; and a far-right custom model for one enterprise only pays when the task is unseen, the data is large, and cost or throughput is strict. For most customers, he still recommends a well-tuned RAG and text-to-SQL stack on a frontier model as the go-to.

Asked whether fine-tuning is dead, Gultekin says no, and he does not even call it a decline. The starting point has changed. Cursor's custom autocomplete model makes sense at Cursor's scale. The rule he uses is start with a large language model, get the product working, and only then specialize if proprietary data, an unseen task, or hard cost and throughput constraints appear.

On the platform, customers can pick OpenAI, Anthropic, Gemini, Meta, Mistral, and DeepSeek. Some Snowflake products hide the model; others expose the choice. When customers want custom weights, some will fine-tune Chinese open models and others will not touch them. Gultekin says that split is not by industry so much as internal policy and culture.

Models like Qwen are "incredibly powerful," and refusing them leaves something on the table, but alternatives are plentiful. The space is competitive enough that no single model dominates, proprietary or open. Optionality matters more than picking a permanent winner, which is why Snowflake's catalog is wide even when a given customer will only approve a subset.

Two and a half years ago, customers were already clear they would not move data out of Snowflake's security boundary. Inference had to come to the data so existing security and governance could be reused and the attack surface would not grow with another copy. Through partnerships, Snowflake runs inference inside that boundary. The model providers act as sub-processors, no state is saved, and regulated industries can use frontier models without the data leaving.

The weights still belong to the labs; cloud providers run the inference. What Snowflake adds is contractual and technical guarantees around data residency and statelessness. Gultekin is not a security researcher, but from working with those providers he does not see casual human access as the main residual risk. The execution environment is set up to run inference and little else, so access to weights is limited by design.

On commoditization, he says the gaps between models are not large, everyone is improving, and that competition helps platforms and customers. Choice still has real constraints: an AWS-centric shop may not want data leaving for Azure-hosted OpenAI. Quality is the second filter — customers run side-by-side evals in the product, not as a consulting engagement — and reasoning can flip a cheap-fast model into a different cost equation. Claude remains the coding specialist he calls out.

Switching is rarer than outsiders expect, because prompts get optimized for one model and a production app with thousands of lines of system logic has a real switching cost. He compares it to Bing catching Google on quality while habits held. Whoever wins the first eval therefore has stickiness, though he thinks differentiation is moving up the stack to products like ChatGPT and Claude Code. Snowflake's own priority is retrieval quality, not making customers thrash models for its own sake.

Labenz places agents on a spectrum from Claude Code-style "choose your own adventure" to a linear chain of prompts, and he warns enterprises not to start at the adventure end. Gultekin agrees, with a persona split. Coding assistants are powerful for developers and builders. A product manager asking "what was usage last week" wants a structured, already-optimized agent with access to the right data, not a Claude Code interface.

Snowflake Intelligence is built for that business-user side. Internally the company deployed a sales assistant to 5,000 sellers: a ChatGPT-like surface on company data, answering questions about upcoming renewals and the book of business, with agents highly optimized for those jobs. Trust is the product. Separately, data engineers and analysts get a more flexible coding agent for pipelines and analysis, not tuned to a single workflow.

Multi-agent coordination is emerging, not yet the main event. Large companies already build sales agents on Salesforce and data agents on Snowflake, and they want those agents to talk. Customers ask for open protocols so they are not locked in; MCP and A2A are on that list. Snowflake does not yet support A2A. The current workaround is to expose other agents as MCP tools and let an orchestrator call them.

Trust, for Gultekin, is a full stack: ease of use, security, governance, quality, evals, and monitoring. Because AI runs next to data, agents inherit granular access control, so two salespeople can ask "summarize my book of business" and see only their own accounts. The UI also flags when an answer was verified by an owner. Guardrails for hate speech and policy violations are product features, and as models follow instructions better, those policies can live in the prompt as well.

Gultekin has spent twenty years in product management, and coding assistants are the shock. The old loop was invent a UI, mock it, then schedule engineering. If users already live in a coding assistant, the first artifact can be a skill written in a day, put in front of customers, and only later solidified into a fuller product. Labenz calls the working prototype "the coin of the realm," and Gultekin says the philosophy is changing faster than behavior is.

That same scoping instinct shapes his view of drop-in knowledge workers. Labenz has heard predictions, including from Anthropic-adjacent voices, that Q2 of this year will bring AI colleagues with names, email, Slack, and meeting access. Gultekin thinks the direction is right and the date is uncertain. A colleague that can do anything is not here. A tightly scoped autonomous agent absolutely could be.

Whatever the agent looks like, he insists the bottleneck is company data: retrieving the right context and handing the agent tools that can act. Making data and tools "AI-ready" only goes up in priority. He already sees access being democratized and the value of existing data rising, because people can ask in natural language. One customer he cited analyzed earnings-call transcripts to score how directly CEOs answered analysts, then used that as trading alpha — the kind of use that used to be too expensive to stand up.

Labenz worries the AI boom looks like an earlier platform cycle: a supernova of edge apps, then a black hole as mega-platforms absorb the features. Gultekin's answer is that enterprise data is not being commoditized, and Snowflake sits between models and applications to help customers build on that data. He thinks competition among model labs is "pro customers and consumers," and that the pie is growing fast enough to create new opportunities even as product shapes keep changing.

On whether agents make it trivial to unplug from a data platform, he is more relaxed than the question expects. Snowflake supports open file formats, including Iceberg, so customers can keep data in a place they manage and use Snowflake as the processing engine. He calls that good for customers and for innovation: they will stay where performance and cost are best. The pressure he feels is from the size of the opportunity, not from a newly short fuse on switching.

Asked where value accrues across chips, models, and apps, he says "the middle will erode and the sides will continue expanding." NVIDIA and the model labs already capture a lot; application builders such as Cursor capture more as they ship unique products quickly. The squeezed middle is companies that encode custom business logic for a niche — the extractor you can now vibe-code in a weekend. He is correspondingly more bullish on horizontal platforms than on vertical SaaS whose main job was to freeze industry best practices into software.

That is also why he does not expect a wave of GE GPT-style models trained on a century of internal documents. Once ChatGPT gained web search, the model no longer needed the whole internet in its weights, only the intelligence to retrieve and interpret — and enterprise RAG plus text-to-SQL is the same cheaper, updatable pattern. Custom small models remain the exception for unseen tasks with brutal cost or throughput constraints. Labenz objects that he can search his own Gmail better than a model with the same access; Gultekin treats that leftover gap as human intuition, which current systems do not yet engineer.

For horizon scanning, Gultekin does not watch public leaderboards closely. Snowflake runs internal quality and latency tests on the jobs it actually ships, and it reruns them whenever a new model is about to drop. He also watches tabular data specifically. Text-to-SQL and semantic models remain the core, but tabular foundation models and fast forecasting are new trends he takes seriously.

His closest thing to a contrarian note is that product-building has to change and is not changing fast enough. Teams are trained on an old sequence. Coding agents make it possible to validate with a working artifact first, and he wants that inversion to spread. Advice to enterprise leaders is blunter still: some companies are still cautious, and he thinks they are in a race. Faster adoption compounds into intuition that changes the trajectory of the business.

Much of the hesitation he sees is really unreadiness of data — silos, access, and foundations that are not AI-ready. Invest there. The last design principle he wants remembered is ease. High-quality AI has to be easy to deploy at scale, or it will not be adopted, no matter how capable the models underneath have become.