SaaStr Deploys AI Agent to Automate Marketing Forecasts and Campaign Planning
SaaStr built an internal AI agent called 10K that runs daily forecasting, campaign design, and go-to-market planning without human intervention. The agent pulls data from Salesforce and vendor APIs, synthesizes it into a six-month projection, and distributes updated forecasts to stakeholders via email and Slack every morning.
The codebase spans roughly 14,000 lines of code with over 370 commits. Each day, 10K reviews historical data, current financial records, and deals in flight before producing new projections and posting results to the team.
How It Works
10K integrates multiple operational data sources in a single pipeline. It pulls from CRM records, financial systems, and third-party APIs to build a unified view of pipeline and forecast.
The agent then synthesizes that data into actionable outputs: updated campaign recommendations, revised six-month plans, and forecast summaries. Distribution is automated-results land in stakeholder inboxes and Slack channels without manual formatting or hand-off.
This shifts forecasting from a periodic exercise into a live deliverable. Instead of monthly or quarterly updates, the team receives fresh projections every business day.
What This Means for Practitioners
For marketing and GTM teams, this example shows how AI agents can automate repeatable forecasting tasks while surfacing real operational challenges. Three problems commonly emerge when building similar systems:
- Integrating heterogeneous data sources reliably-Salesforce, finance systems, and vendor APIs often have different schemas, latency, and update cadences.
- Maintaining data lineage and auditability-stakeholders need to understand where forecast numbers come from and why they changed day-to-day.
- Handling stale or incomplete CRM inputs-missing deal data or outdated pipeline records can skew projections without detection.
Automating the pipeline reduces manual toil, but it increases dependency on robust data validation, observability, and access controls.
What's Missing
SaaStr's write-up omits low-level technical choices. There's no detail on which LLM endpoints power 10K, how prompts are structured, or how often the agent retrains on new data. Those details matter for teams trying to replicate the approach.
The public reporting also doesn't address how forecast uncertainty is communicated to humans, which safeguards prevent biased inputs from propagating downstream, or what forecast accuracy looks like over time. Incident logs for integration failures would help practitioners understand failure modes.
Teams evaluating similar projects should ask those questions before committing to an in-production agent. The operational value is clear; the technical reproducibility is not.
Learn more about AI for marketing operations and strategy.
Your membership also unlocks:
