Building GTM AI Agents: Lessons from Deploying to 6,000 Users — Sait Izmit, Snowflake

Company / GTM context

• Presenter leads internal AI tools for Snowflake’s sales team and explains lessons from deploying Snowflake’s internal go-to-market (GTM) assistant at scale.
• Strategy framing: With ~half of a ~10,000-person workforce in sales, the core business problem is revenue productivity, constrained by:
  • Fragmented customer/account data
  • High analyst dependency

Why build an AI GTM assistant (business problem → outcomes)

Common customer pain (enterprise / Fortune 500)

• Data is siloed across SaaS systems (first- and third-party data).
• Reps use ~15 tools because each tool has different datapoints; reps then stitch insights manually in spreadsheets.
• Each rep manages ~1,000 accounts/customers and must track rapidly changing signals (e.g., news, consumption, support tickets, earnings updates) dozens of times per day—which is humanly infeasible.

Claimed business value from AI

• Data democratization
  • Fewer dashboards
  • Less reliance on analysts
• Automation
  • Frees inbox time
  • Consolidates tools
• Higher seller effectiveness
  • Better win rates
  • Shorter deal cycles
  • Incremental revenue
• Tooling/productivity goal
  • Increase rep capacity (“cover more of your book”)

Core GTM/AI execution principles (the “playbook”)

Quality over coverage (explicit rule)

Avoid “answer lots, be wrong often.” Start with fewer, higher-confidence interactions.

• Start with ~50 questions, targeting ~95% accuracy
• Don’t do 100 questions at 70%—it harms first impressions and trust
• Team mantra: “Quality is P minus one”
  • Quality acts as a trust multiplier for adoption

Controlled rollout with staged launches

Face launch → Pilot

• Goal: prove accuracy/quality
• Use AI-native top users for feedback to fix early issues

Better launch

• Target 10% better performance at around ~600 people
• Validate MVP adequacy and measure:
  • Question volume/demand
  • Retention rate
• Target: >70% weekly active user retention

GA (general availability)

• After activation issues are addressed, activation becomes less of a bottleneck (see activation workstream below)

Change management / activation as a first-class workstream

• A key failure mode is not technical quality, but low activation.
• Presenter’s claim: at launch time, only ~20% of the org actually tried the product, so usage stays low even if quality is good.

Operational approach:

• Spend 60–70% of time in sales meetings on:
  • Demos
  • Dashboard usage enablement
• Secure sponsorship from sales leaders
• Address underperformance behaviorally:
  • “Shame” managers whose teams adopt poorly to drive adoption
• After activation:
  • Shift focus from “getting started” to driving depth
  • Increase usage breadth and advanced workflows

Prevent “collapsing wow factor” (keep raising the bar)

• Over months, users can shift from “rockstars” (initial impressiveness) to friction:
  • “I can’t do this anymore—another AI product does.”
• Required evolution of capabilities (staged value ladder):
  1. Talk to your data (reduce dashboards / analyst dependency)
  2. Add integrations (MCP connections) → automate workflows
     • Examples: monitor inbox/Slack, track customer questions, draft Gmail replies, review & send; automate outreach
  3. Empower teams
     • Team-specific dashboards, automations/alerts
     • Reduce IT backlogs and vendor bottlenecks
  4. Hyper-personalization
     • Personalization for reps and customers using living context (customers/contacts)

Build + launch + learn (avoid “perfect architecture” waits)

• Enterprises often chase “perfect architecture” via framework testing and maturity exploration.
• Corrective rule: build, launch, learn—architectural “boxes” evolve after launch.

Expect ongoing re-architecture with new tech

• Re-architecting is normal.
• Example cited: ~30–40% of sprint work later focused on re-architecture and adapting to new capabilities/limits.

Invest in logs for feedback loops

Logs enable scalable iteration by:

• LLM-classifying questions/topics
• Extracting detailed categories/subcategories and examples
• Detecting:
  • Feature gaps
  • Quality issues
  • Repeated questions
  • Frustration signals

This creates automated sales enablement feedback loops:

• Instead of interviewing ~100 sellers/week, infer knowledge gaps quickly (minutes)
• Automatically generate artifacts like battle cards
• Connect those artifacts to systems (e.g., Confluence/Jira/Slack)
• Feed improvements back into the agent

Concrete metrics & KPIs (reported targets and results)

Scale / usage

• >1 million questions answered since September launch
• ~40,000 questions/week

Staged rollout KPIs

• Pilot accuracy goal: ~95% accuracy on a constrained set (~50 questions)
• Better launch: ~600 users and validate MVP coverage
• Retention: exit threshold >70% weekly active user retention

Product/system complexity (signals maturity at scale)

• 15 semantic views
• ~85 tables
• ~3,000 columns
• 5–6 MCP connections
• ~20 skills

Change management KPI / observation

• Only ~20% of the org tried the product during the early launch window (implies activation gating)

Examples / actionable recommendations distilled

• Start by testing seller-realistic questions even before full instrumentation
  • Example: created a spreadsheet with ~150 sales-process questions to pressure-test early even when “data isn’t in the agent yet.”
• Adopt progressive capability rollout to retain trust
  • Don’t expand coverage until quality is proven; otherwise first impressions degrade permanently.
• Treat sales enablement + activation as engineering-adjacent
  • Adoption ramps through demos, meetings, and leadership sponsorship.
• Operationalize feedback via logging
  • Use logs as a continuous requirements pipeline:
    • automate knowledge/battle-card generation
    • update the agent with faster loops than human interviews
• Architect for iteration rather than stability
  • Expect shifts across: agent instructions → CI/CD/eval infrastructure → skills → MCP → progressive disclosures → memory → scheduling, etc.

Frameworks / concepts explicitly referenced

• Quality-first rule: “Quality is P minus one”
• Staged rollout / launch gates: Pilot → Better launch → GA
• Retention as adoption health metric: >70% weekly active users returning
• Trust dynamics (implicit): first five questions are critical; non-deterministic behavior risks “trust loss overnight”
• Wow-factor lifecycle management: “collapsing wow factor” → keep iterating to raise baseline expectations
• Feedback-loop system: logs → topic classification → gap detection → automated enablement content → agent updates

Presenters / sources

• Sait Izmit (Snowflake) — primary presenter.