How Anthropic's sales team actually uses Cowork

01 · The setup

From "AI you talk to" to "AI that does stuff for you"

Chatbots have only existed for about four years, and the first version — the one where you ask Claude to "edit my email" or "summarize this memo" — is fundamentally Q&A. Helpful, but in the loop on every keystroke. Cowork is the moment that flips.

Internally, the lineage went: chat → Claude Code → Cowork. Claude Code proved the engine worked — developers were setting off long-running, autonomous tasks. But sales reps don't live in a terminal. (Travis: "There's something in your little sales lizard brain that's like, I should not be talking to my computer in this way.") Cowork is that same engine wrapped in a GUI for "us mere mortals, or normies."

"I don't just ask Claude questions. I'm delegating tasks. Cowork takes that same execution power that's been transforming coding and opens it up to the rest of us non-technical folks." — Brittney Tong

The teachable example Brittney returns to all session: "Help me prep for my customer calls tomorrow." Cowork looks at her calendar, pulls the stakeholders from Salesforce, scans Slack and recent emails, searches LinkedIn, queries the data warehouse for usage trends — then hands her an agenda, expected objections, sample questions, and next steps. What used to take hours happens while she's making coffee.

02 · The framework

What changes with Cowork

A day on rails

Brittney's actual day:

1. Before her deskMorning digest skill runs automatically. Drops into Slack. Account health, team updates, calendar prep — all there.
2. Mid-dayA skill scans her Slack for unanswered threads. Surfaces what she missed during back-to-backs.
3. End of dayAnother skill sweeps open emails, action items, anything still unfinished. Tags them onto tomorrow's to-do list.

03 · Under the hood

The four moves Cowork makes on every task

Why Cowork output is better than dropping the same prompt into a chatbot: it doesn't just respond. It runs a small reasoning loop you can interrupt at any step.

1. UnderstandBefore doing anything, Claude asks clarifying questions. Which calls? How much detail? What to skip?
2. PlanIt writes out the plan you can read: "I'll pull your calendar, search Salesforce for stakeholders, check Slack threads, query usage in BigQuery, then draft the agenda." This is the cheap interrupt point — fix it now if it's wrong.
3. ExecuteNow it runs. You're not babysitting — you're in another tab answering customer Slacks.
4. VerifyIt cross-references the data, flags low-confidence claims, surfaces contradictions. Then notifies you it's done.

"One of the things I appreciate about Claude is it's not afraid to call you out if you gave some incomplete task. It's a very thoughtful intern — but an intern that's willing to push back." — Travis Bryant

04 · Mobile

Dispatch — Cowork goes wherever you go

The pattern Brittney describes: she walks out of a customer meeting, a thought hits her — "I need a follow-up email drafted and prep for tomorrow's call." She fires it from her phone. By the time she sits back down at her laptop, the output is sitting there.

05 · The stack

Connect everything, then write skills for what you do three times

The non-negotiable: Cowork is only as good as the systems it can talk to. Anthropic's sales stack runs on MCP connectors — Salesforce, BigQuery (the data warehouse), Slack, Gmail, Calendar, Gong, Zoominfo / Apollo / Clay for enrichment. Most of these have point-and-click connectors. Custom internal systems can be wrapped with a connector your IT team builds.

The Rule of Three for skills

Travis's heuristic for when to codify a workflow:

"Anytime you're doing something for the third time, that's the little light bulb of: maybe I could just tell Claude how to do this and put it on rails." — Travis Bryant

And here's the cheat-code part: you don't write a skill from scratch. Have a normal conversation with Claude where you get to a useful output. At the end, say Claude, make a skill from what we just did. It writes the skill file into your directory. Next time, it knows the routine.

Demo 01 · IC morning

Brittney's daily digest — the 6am Slack message that runs her day

Every weekday before Brittney sits down at her desk, a scheduled skill has already swept the last 24 hours of Slack, Gmail, Calendar, Salesforce, and BigQuery and dropped a structured digest into Slack.

Section 1 — Account health, color-coded

Every account in her book is tagged red / yellow / green with the data signals and the action she should take.

Section 4 — Today's prep, with a self-generated prompt

This is the loop-closing trick. The digest doesn't just list today's meetings — it generates a Claude prompt for each one. If Brittney isn't ready for the 10:30 with Beacon, she copies the auto-generated prompt, pastes it into Cowork, and gets full call prep without writing a single character.

Demo 02 · Thursdays

The forecast updater — 2 hours → 10 minutes

Every Thursday, Brittney has to update her Salesforce pipeline before manager reviews on Friday. The old way: open each opportunity, dig through Slack and emails for what changed, type next steps in the format her managers want.

The new way: she runs her /salesforce-pipeline-update skill. It pulls every open opportunity she owns, walks her through them one by one, and proposes the change for each field — Stage, Amount, Close Date, Forecast Category, Next Steps. She approves, edits, or rejects. At the end it pushes everything to Salesforce in one batch.

Where Cowork actually shines: catching what you missed

On one opportunity, Cowork proposes pushing the close date because it noticed a Slack message from the CTO's EA that he's going on PTO. Brittney didn't have to remember that — it surfaced organically from the data pull.

"I've developed a reputation as a Salesforce traffic cop of next steps, close date, stage, forecast category. There's just no excuse anymore. You've got an alien brain at your beck and call to make sure next steps reflect reality." — Travis Bryant

Demo 03 · Manager view

Travis's weekly forecast roll-up — and what a skill actually looks like

This is the demo to watch if you want to see a skill. Travis opens the skill file in Cowork. It's not magic — it's a markdown document with a description, a trigger, and step-by-step instructions.

The skill does the transfer-tax work: pulls forecast submits from Salesforce, pulls usage data from BigQuery, confirms the AE roster (hiring is fast), and assembles the weekly report Travis sends up to the Chief Commercial Officer.

The output: a forecast call that's a conversation, not a readout

Travis used to keep a "secret push counter" — every time a close date moved between forecast periods, a Salesforce workflow rule incremented a hidden field, and he'd sort deals by push count. Now Claude just surfaces those deals automatically.

Then he put it on a schedule

Demo 04 · Strategy work

4,000 accounts scored overnight — Claude as a thinking partner, not just an intern

The most ambitious demo. Travis needed to score every account in mid-market — 4,000 of them — on growth potential so reps could prioritize. Old playbook: hire a strategy consultant.

His pivot: treat Claude as a colleague, not a task-runner. He started a back-and-forth conversation about how to score accounts. Claude proposed dimensions Travis hadn't considered:

Build small, then scale

Travis applied the scoring to one territory first (healthcare), validated the numbers, refined the weights (dropped "industry fit" because it wasn't moving the score, bumped agent opportunity to 40%), then expanded to legal, FSI, manufacturing.

"Set it before I went to sleep"

When the model was right, Travis kicked off the full run before bed at a QBR. "I woke up and it had gone through deep research across thousands of accounts to come up with a five-dimension score for every account in the book."

The icing

Then Travis dropped one line:

He'd previously stored Anthropic's brand spec — colors, fonts — as a skill. Cowork loaded it, restyled the dashboard, re-published the live artifact.

06 · Audience Q&A

The eight questions worth replaying

Verbatim audience questions from the live session, with the speakers' answers distilled to the essential takeaway.

Matt asks:

"As an IC running a strategic sales territory, my buyers are EVP / SVP at $50 billion banks who can smell AI-flavored outreach in two seconds — the old em-dash. Where's your team's line between 'Cowork drafts it and I send' versus 'Cowork drafts and I rewrite the voice'? And where have you been burned by that line being in the wrong place?"

Brittney: Dictate into Claude — it learns your cadence and capitalization. End every draft prompt with "stay as close to my voice as possible." Travis: No automated send. Ever. However much it drafts versus I edit, I'm still the person who hits Send. Cowork helps with 90%; the human stays responsible for the last 10%.

Emily asks Travis:

"Do you share skills across the team, or does everyone build their own? What's the tradeoff?"

Travis: The answer is yes — both. People parallel-process and you end up with what he calls "mushrooms in the forest": look under a rock and there are five new skills. That's fine — innovation comes from everywhere. Slack channels emerge for sharing; the best skills get blessed by sales productivity and promoted to official internal skills.

Rahul asks:

"What's the core difference of a skill versus a project?"

Brittney: A skill is a reusable workflow you run again and again (e.g. her weekly forecast). A project is a super-folder with persistent memory — dump all the Google Docs, Slacks, and context for one complicated deal in there, and every new conversation in that project remembers it. Travis: Skills work across any conversation; projects scope shared context to a specific deal or initiative.

Elizabeth asks:

"When should we use Cowork instead of Claude Code?"

Both: Same engine under the hood. Skills, MCP connectors, outputs — all interchangeable. Cowork is just friendlier — desktop GUI instead of terminal. Travis's analogy: "It's more Windows than DOS." Pick whichever surface you're comfortable in.

Anonymous asks:

"What is your degree of confidence in this information? Do you ever go back to the source to review it, or are you just happy with the output Cowork is producing?"

Travis: "Verify, then trust" at the start; later "trust and verify." When he built the account-scoring routine, he sampled outputs heavily — caught queries pulling from the wrong Salesforce table — before putting it on rails. Frequency of quality checks should follow a Fibonacci curve: tight at the beginning, looser as confidence compounds.

Brittney asks Travis:

"How else are you measuring the ROI of Cowork across your org, besides time saved? Any metrics you look at?"

Travis: Revenue is the lagging indicator — he obsesses over leading ones. Specifically: (1) how many accounts an AE can go deep on ("velocity with depth"), (2) count of live customer conversations — he doesn't care about mass emails or Slacks, (3) whether every AE has a quarterly project to build a skill that automates their own workflow.

Anonymous asks:

"As a leader, have you had the 'Claude did that' excuse from reps regarding subpar Salesforce updates? How do you level-set with reps on the expectation that they validate everything Claude does?"

Travis: Yes, some reps are "more AGI-pilled than others." But the message is unambiguous: Claude is not the AE; we stand by the work irrespective of where it came from. His analogy: "Waymo is driving, but we're still responsible if something goes wrong." Human stays in the loop because humans are accountable for the output.

Anonymous asks Brittney:

"Do you recommend always using Opus 4.7 for Cowork tasks?"

Travis (interjecting): "Do you like driving the Ferrari to the grocery store?" Brittney: Opus for the multi-turn complex skills — her daily digest, where Claude has to write its own follow-up prompts. Sonnet for everyday call prep. Haiku for trivial. Use the mix.