# Building GTM AI Agents: Lessons from Deploying to 6,000 Users Sait Izmit, Snowflake | AI Engineer World's Fair 2026 | 20:39 Source: https://www.youtube.com/watch?v=DrTdD-ttjCY Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/building-gtm-ai-agents-lessons-from-deploying-to-6-000-users Published: 2026-08-26 Tags: agents, evals, team-adoption, workflows ## TL;DR - Sait Izmit argues that a go-to-market agent should answer a smaller set of questions accurately before expanding its coverage. - The difficult part of an enterprise agent rollout is activating users after launch, since low trial rates cannot be fixed by improving the product. - The agent must keep moving from data questions to workflow automation and team-built tools, while the engineering team accepts regular re-architecting. ## Summary Sait Izmit describes Snowflake's internal go-to-market assistant, which launched in September and now serves about 6,000 users. His approach starts with quality. Before connecting more data, his team wrote 150 questions from the sales process and used them to expose gaps. The first test reached 50 percent accuracy, leading the team to focus on answering fewer questions well. The assistant expanded after launch, eventually reaching 15 semantic views, 85 tables, 3,000 columns, MCP connections, and nearly 20 skills. Izmit spends much of his time on activation through meetings, demos, dashboards, and sales leadership support. He also warns that the initial excitement around talking to company data fades once it becomes normal. Teams then expect workflow automation and tools they can build themselves. His advice is to launch with the current stack, use logs to find gaps, and expect the architecture to change as the product grows. ## Key ideas ### Go-to-market teams need one place to work across siloed data [01:21](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=81s) Izmit frames the problem around Snowflake's sales organization, where almost half of the workforce is responsible for revenue generation. Reps may use 15 different tools because each contains a different data point. They combine the information in spreadsheets while tracking up to 1,000 assigned accounts, including news, product consumption, support tickets, and earnings results. Izmit says no person can keep up with that amount of information repeatedly throughout the day. An agent can bring the data together, reduce dependence on dashboards and analysts, and support automation. The expected business effects include more account coverage, better win rates, shorter deal cycles, and incremental revenue. ### Trust depends on the first few answers [03:24](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=204s) Izmit says AI projects often fail because users lose trust quickly in a free-form chatbot. Users who like the first five answers return, while winning back someone who had a poor first experience takes ten times more effort. His team calls this principle 'quality is P minus one.' That standard shaped the launch plan. The agent should answer 50 questions at 95 percent accuracy rather than answer 100 questions at 70 percent. A good first experience gives users a reason to ask for more. Izmit treats quality as the condition for expanding coverage, rather than connecting every possible source before the product has earned confidence. ### Testing real sales questions exposes the gaps before launch [04:16](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=256s) Before trying the existing agent, Izmit opened a spreadsheet and wrote down 150 questions taken from the sales process. Engineers objected that much of the required data was not connected, but Izmit said those were still the questions sellers would ask. The first test scored 50 percent accuracy. That result gave the team a concrete basis for narrowing the initial scope and improving answers. The agent launched with a small data set, and 60 percent of its data was added during the six or seven months after launch. It later grew to 15 semantic views, 85 tables, 3,000 columns, five or six MCP connections, and close to 20 skills. ### A phased rollout tests quality, product fit, and retention separately [05:34](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=334s) Snowflake uses a controlled rollout with three stages. A pilot starts with AI-native employees who can give feedback and help smooth the rough edges. The next stage is a 10 percent beta with 600 people. That group tests whether the product is an MVP for daily workflows, while its requests reveal which missing data connections matter most. The team also measures whether weekly active users return, and Izmit says they exited the beta with retention above 70 percent. General availability comes only after the team has confidence in accuracy, coverage, and repeat use. The staged process protects the first user experience while expanding the product. ### Activation after launch requires sustained sales involvement [07:14](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=434s) Izmit says many AI products fail after general availability because only a small share of the organization has tried them. Two weeks after launch, management may see low usage even though only 20 percent of employees have spent five minutes with the product. Izmit separates that problem from product quality. If people try the tool and do not return, the product has a problem. If they never try it, the team has an activation problem. He spends 60 to 70 percent of his time in sales meetings, giving demos, building adoption dashboards, identifying teams that have adopted the product, and getting sales leaders to encourage trial. He says this work materially affected the scale of adoption. ### The product has to keep moving after talking to data becomes normal [08:56](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=536s) The first version of the product creates a wow effect because sellers can talk to company data instead of waiting for dashboards or analysts. After four to six months, that capability becomes a habit and users start asking why the product cannot do more. Izmit calls this the collapsing wow factor. His roadmap moves from talking to data, to automating workflows through integrations, to giving teams the ability to build skills, dashboards, applications, automations, and alerts. Examples include monitoring inboxes and Slack channels, drafting customer responses into Gmail for review, and automating outreach. If the team stops after the first phase, users can switch to another product quickly. ### The initial architecture should be treated as temporary [11:49](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=709s) Izmit criticizes large enterprises that spend months searching for a perfect architecture instead of launching and learning. Snowflake's first version had nine pages of agent instructions, Cortex Analyst tools, semantic views, a Cortex Search service, and instructions managed in a Google Doc. After launch, the team added CI/CD, evaluation infrastructure, unit tests, routing tests, skills, MCP connections, progressive disclosure, user memory, task scheduling, and interfaces beyond chat. The current architecture shares only about 20 percent with the original plan. Izmit estimates that 30 to 40 percent of ongoing work involves re-architecting as technology changes. Teams need enough flexibility to pivot rather than over-invest in the current design. ### Logs create a feedback loop for product and sales enablement [14:18](https://www.youtube.com/watch?v=DrTdD-ttjCY&t=858s) Izmit recommends investing in logs because they show what users ask and where the agent fails. Snowflake uses language models to classify more than a million questions and break them into categories, subcategories, and example questions. The team can see unanswered questions, quality problems, repeated prompts, and negative reactions. That gives product teams a live view of feature gaps. It also helps sales enablement. When a new product launches, the team can identify changing topics and missing battle cards without interviewing 100 sellers each week. They can connect the findings to Confluence, Jira, and Slack, ingest product requirements, generate enablement documents, and feed those materials back into the agent. ## Notable quotes - "User trust is earned extremely hard and is lost overnight." (03:24) - "We don't want to try to answer 100 questions and get them 70% right. We want to answer 50 questions, but get them 95% right." (04:39) - "A lot of these AI initiatives, they don't fail because there's an issue with the technology, they fail actually in activation." (16:51) - "You should be okay to pivot very easily." (13:29) - "I would really, really recommend investing in your logs." (14:18) ## Tools & references mentioned - Snowflake Co-work - Snowflake Intelligence - Cortex Analyst - Cortex Search - Cortex - MCP - Salesforce - Confluence - Jira - Slack - Gmail ## Who should watch - You are building an internal agent for a sales or go-to-market organization and need a rollout plan that tests quality before broad access. - Your product has launched, but adoption is low and you need to distinguish a product problem from an activation problem. - The first version of your agent works, but users now expect workflow automation, team-built tools, and new interfaces.