Anthropic's Sales Team Ships Cowork on Their Own Stack

The Stack Anthropic Ships On

Travis Bryant, Anthropic's Head of US Mid-Market GTM, runs a 4,000-account book on Salesforce (system of record), BigQuery (data warehouse), and Claude Cowork (the AI ops layer on top). The case study published in May 2026 documents three production workflows: daily call prep, Friday forecast rollup, and overnight territory scoring. The team reports 90 minutes a day back from micro-optimizations and 3 hours a week from the forecast alone.

The architecture is worth studying because it is the pattern we see mid-market B2B SaaS and FinTech shops converging on: Salesforce stays the source of truth, the warehouse stays the analytical substrate, and the LLM layer sits above as an augmented ops runtime. Cowork does not replace the CRM. It does not send emails on its own. It does briefing, scoring, forecasting, and assembly work that used to live in spreadsheets and analyst hours. Every customer-facing send still goes through a human.

If you are a VP Sales or RevOps leader running a growth or mid-market team and you are trying to decide where to wire in your first agentic workflow, this is the reference architecture.

The Three Workflows That Matter

The daily call prep workflow pulls account context, pipeline data, recent engagement signals, and competitive intel from Salesforce and BigQuery, then assembles a briefing doc for the rep's first meeting of the day. This used to be the first hour of every sales morning: opening tabs, cross-referencing the CRM, pulling usage data from the warehouse, checking Slack for the last time someone mentioned the account. Cowork runs this overnight as a scheduled skill. The rep opens a brief, not a browser full of tabs.

The Friday forecast rollup is the workflow that ships 3 hours back to the team every week. Leadership expects a specific format: pipeline breakdown by stage, risk flags on deals over a certain size, a narrative summary of movement since last week. Cowork pulls the data from Salesforce, applies the team's scoring definitions (encoded as a skill the team controls), and generates the rollup in the format leadership actually reads. The sales leader reviews, edits if needed, and sends. The assembly work is gone.

The overnight territory scoring workflow ranks the 4,000-account book by propensity to close, engagement velocity, and fit against ideal customer profile. This historically took hundreds of analyst hours per quarter. Cowork runs it as a scheduled skill every night, writing results back into a BigQuery view that feeds into the next morning's call prep. The ranking logic is a skill Travis's team wrote, not a black box. If the scoring definition changes, the team updates the skill.

The Hard Rule: Human Approval on Every Send

The case study is explicit about the gate: every customer-facing send goes through human approval. No autonomous SDR. No AI rep cold-calling prospects. The human still owns the relationship, the call, and the decision to send. Cowork augments the ops layer, not the customer layer.

This is the architectural discipline that separates production agentic stacks from demos. If you are a mid-market sales leader evaluating where to introduce agentic workflows, the rule is not automate everything. The rule is automate the assembly, briefing, and scoring work that does not require human judgment, and keep the human in the loop on every interaction that touches a customer or a forecast that leadership will act on.

The Anthropic team enforces this with connectors and skills they control. Salesforce is still the system of record. BigQuery is still the warehouse. Cowork reads from both, writes scoring results back into BigQuery, and surfaces briefings and rollups to humans who decide what to send. The LLM does not have write access to the CRM. It does not send emails. It assembles, ranks, and prepares.

If you are running a similar stack (Salesforce + Data Cloud or Salesforce + Snowflake or Salesforce + Databricks), the pattern is the same: the LLM layer sits above your system of record and your warehouse, it runs scheduled skills that encode your team's definitions of good scoring and good briefing, and it surfaces work to humans who own the final send. The connector architecture is the control plane.

What This Means for Mid-Market Sales Ops in 2026

If you are a VP Sales or RevOps leader at a mid-market B2B SaaS, FinTech, or HealthTech company, the question is not should we buy Cowork. The question is which weekly sales-ops rhythms are good first targets for AI augmentation. The Anthropic case study gives you three proven patterns: daily call prep, Friday forecast rollup, and overnight territory scoring. Start with one. Wire it in with connectors and skills you control. Keep human approval on every customer-facing send.

The lesson is not replace your CRM with an LLM. The lesson is Salesforce stays the system of record, your warehouse stays the analytical substrate, and the LLM layer sits on top as an augmented ops runtime that handles the briefing, scoring, forecasting, and assembly work that used to eat the first two hours of every sales day. Anthropic's own Mid-Market GTM team shipped this stack and reports 90 minutes a day back and 3 hours a week back on the forecast. That is the signal.

The reps still own the relationships, the calls, and every send. What Cowork replaces is the spreadsheet-and-tabs assembly work. That is the move worth copying.