Behind the Build: How to Create Trustworthy AI for Go-to-Market Teams

When stepping into any C-suite sales meeting, the stakes are high. It’s essential that any seller knows the customer business, is up to date on current goals and projects, and remembers the ins and outs of every recent touchpoint with the customer. Accuracy is critical. Sounds like hours of prep right? Unless … you can turn to AI and trust AI to give you an accurate view of the insights across your product data, CRM data, internal sales knowledge and more to keep you up to speed.

With Snowflake Intelligence now in public preview, we’re making this possible. We’ve spent the past year building a generative AI tool, the GTM AI Assistant, for our go-to-market teams. It’s designed to do more than surface information—it interprets structured and unstructured data, understands business context, and delivers trustworthy, actionable insights at the speed of conversation.

This is not just another internal AI project. It’s an enterprise-grade system that supports our sellers, operators and executives throughout the sales journey, standardizing and democratizing all the information they need to do their jobs efficiently and effectively. The outputs must be accurate. And the design decisions matter.

Here’s what we’ve learned.

Lesson 1: Understand the pain points clearly

We started this process by understanding the seller workflows and identifying the friction points. Take customer meeting preparation as an example. Today, prepping for a customer call is a multitab exercise. Sellers toggle between dashboards, Salesforce, internal docs, support systems and web searches — manually piecing together context.

We realized that with AI, that experience can become conversational, and we used Snowflake Intelligence to make that possible. The GTM AI Assistant allows sellers to tap into the full breadth of our sales data and knowledge repository, drastically reducing the time they spend hunting for, verifying and synthesizing information. Ultimately, it allows them to spend more time on core selling activities that translate to real-world business outcomes.

For example, a seller can ask:

• “What training is most relevant for this use case?”
• “What features has my customer adopted recently, and what value do they bring?”
• “Summarize their Snowflake consumption over the past 30 days.”
• “What areas need attention for Customer ABC?”
• “What similar use cases could be worth mentioning?”
  
  

And the seller receives a concise, contextual answer, including such customer information as:

• Company overview: A brief profile from trusted public sources
• Snowflake footprint: Contract terms, workloads, account team, usage trends and forecasts
• Consumption insights: Trailing 12-month performance with key patterns and anomalies
• Use case adoption: Active and emerging use cases mapped to Snowflake capabilities and industry benchmarks
• Support engagement: Summary of recent tickets, sentiments and open escalations
• Enablement and best practices: Curated training, product guides, talk tracks and case studies tailored to industry, persona and maturity

This isn’t just retrieval — it’s reasoning. The assistant accesses an array of underlying resources, interprets context and surfaces what matters. The result: less prep time, more impact in every conversation and more time to focus on core go-to-market responsibilities.

Lesson 2: The right content beats all the content

We had to think carefully about how to design the GTM AI Assistant. With the wealth of information available in Snowflake, we had to decide what information would drive trustworthy, accurate results. 

One of the first decisions we made when designing our assistant was not to index everything. Just because AI can access every slide deck, email thread or Slack channel doesn’t mean it should.

The quality of an LLM’s responses is only as good as the quality of the context you give it. In the enterprise, that context is often messy, redundant, outdated or contradictory. Indexing thousands of documents across dozens of systems may seem comprehensive, but it often results in confusion rather than clarity.

So we took a different approach.

Instead of crawling every workspace and chat log, we curated a set of trusted, up-to-date content for our GTM AI Assistant to reason over. This includes:

• Official sales training materials
• Certified product and solution decks
• Key Slack channels (with a signal-to-noise threshold)
• Enablement documentation from our product and marketing teams

The result? When a seller asks a question like “What’s the best way to position Snowpark to a data engineer?” the assistant doesn't guess from outdated sales scripts; it pulls directly from vetted materials created for that purpose.

By narrowing the scope to high-quality inputs, we dramatically improve the signal and reduce hallucinations. It’s a lesson we’ve learned again and again: In AI, more context isn't better. Better context is better.

Lesson 3: Structured data is different

Depending on the type of question, “accuracy” of results can have very different interpretations and evaluations. For an open-ended question, there is a range of replies with different phrasing, length or structure that can deliver “accurate” information. 

Now, consider a common question: “What was my customer’s Q1 consumption revenue?”

Here, precision is nonnegotiable. The answer must be accurate down to the penny; there’s only one correct number.

This is where most LLMs face a challenge. They are probabilistic and designed to predict the next word, which can lead to inaccuracies with structured data where precision is essential. Moreover, unlike traditional, deterministic queries, an LLM might not give the same correct answer twice, making consistency as important as accuracy.

This is why the stakes are higher. And this is where governance and precision need to meet generative AI.

This challenge meant we had to think carefully about how to bridge the gap between LLMs’ flexibility and structured data’s precision. We approached this challenge using three key design pillars.

1. Building the right semantic data foundation

Before we could trust AI to answer structured data questions, we had to align on what those answers mean. 

Definitions such as “fiscal year” or “active customers” can be defined differently at different companies. If that ambiguity exists in the business, an LLM will amplify it.

We addressed this by building a shared semantic layer: defining key metrics, entities and logic across sales, marketing and support. These definitions are modeled at the warehouse level in Snowflake Semantic Views and documented for both humans and AI systems to interpret.

2. Verified query repository

We maintain a curated library of SQL queries vetted by our data team for common questions. When a seller asks “What’s the current utilization of my customer?” the assistant doesn’t compose a new query; it maps the request to a known, tested one. When a verified query is used, we display a shield icon with a checkmark. The user sees “This response is based on a verified query.” That small signal builds user trust and allows sellers to differentiate between experimental outputs and enterprise-grade answers.

3. Continuous testing and evaluation

Our team continuously monitors for quality and consistency, running automated tests on approved prompts. When we find failures, we refine the system by expanding our verified queries, improving LLM instructions or adding metadata. We also evaluate real-world performance through a combination of sampling, manual review and LLM judges to ensure the assistant meets user needs.

Looking ahead

We believe GTM workflows will increasingly be served with AI solutions but recognize that those AI solutions need to be grounded in rigor. Our team’s goal is to remove friction, automate the repetitive and make insights flow naturally in every workflow within Snowflake. Already, our GTM AI Assistant is empowering teams across Snowflake to quickly stay up to speed on accounts, get insights into customer data quickly and tap into the right sales resources in an instant. We are serving more than 5,000 user queries a week, saving our teams hundreds of hours — and we’re just getting started. 

We are moving forward by treating data as a first-class product. We have to embed trust directly into the answers. And we have to design systems that know when to guess and when not to.