Experiments in Motion: How Anthropic Runs GTM on AI

Part of a series on building enterprise agentic AI. The foundations — the spec, the semantic layer, the identity graph, and memory — define what to build. This is a companion piece: what it looks like when a company already runs on that substrate.

A clean-slate proof of concept, not a migration playbook.

When Claude Opus 4.6 shipped in December 2025, Anthropic’s commercial team came back from winter break to find demand that had, in Head of Industries Eleanor Dorfman’s words, “gone vertical.” They hadn’t hired for it. They hadn’t planned for it. And even if they’d been ready to 3x or 4x or 5x the sales team, you can’t absorb that many people fast enough to keep the customer experience intact.

So in January they rebuilt the sales org around AI. Over the following months, Anthropic published a handful of accounts of how its go-to-market teams now operate — a talk from Dorfman, a profile of one seller, a marketing-ops writeup, a webinar. Read together, they sketch what an AI-native GTM motion looks like in practice.

But I want to be careful about how we read them. These are experiments in motion, not a finished blueprint. Dorfman is candid that forecasting is still a work in progress — her team spends ten minutes at the top of every forecast call debating how they should be forecasting. And, more importantly, Anthropic reached this point without carrying the things that make GTM transformation genuinely hard everywhere else. The tooling is the easy, visible part. The substrate underneath it is the real work. That’s the thread I’ll pull on by the end.

The pattern that recurs

Across every example, the same moves show up.

Thread, don’t replace. Nobody rips out the stack — Salesforce, Gong, Ironclad, Clay, LeanData, Slack, HubSpot, BigQuery, Asana all stay. Claude does the work that used to live in the gaps between them: the connective tissue, not one more tool bolted on.

Encode the best operator. A “Skill” bundles instructions plus MCP connectors, summoned with a slash command. The patterns a top rep runs by instinct become the default for everyone, which collapses the onboarding ramp.

Keep the human on judgment. Claude handles the data hunt; people own validation and the decision. Proposals get validated against policy, numbers get checked against verified sources, forecasts get inspected by managers.

Run it on a cadence. Sunday-night report prep, hourly request routing, 7am briefings. Work that runs on a schedule is work nobody has to remember to do — and repeated corrections get fed back into the Skills, so the system tightens over time.

The same motion, four vantage points

Rebuilding the sales org. Dorfman’s SaaStr AI 2026 talk is the widest lens. Facing four fixed constraints — demand they couldn’t slow, headcount they couldn’t add, a stack they wouldn’t rip out, and support functions that had to scale alongside sales — the team killed the old product-led-vs-sales-led split (54% of new enterprise logos in 2026 now close self-serve, at real ACV), made Slack the single front door for deal desk, legal, and RevOps (“ticket in, ticket out, Claude triages”), and codified their best reps as Skills: morning briefing, call prep, follow-up, competitive intel, asset creation. (SaaStr’s written recap is a good companion.)

One seller’s tools become the team’s. Jared Sires was an AE with no coding background, buried under 600–700 accounts and answering email until 10pm. He used Claude Code to build CLAFTS — a Gmail app that drafts replies in his voice from live product docs — then daily brief and recap Skills, then a full sales plugin shipped through Claude Cowork. Within months, ~80% of the sales org was using it. The full story is worth reading.

Marketing ops. Ian Chan and Annabel Custer compressed days of manual work into hours. Ian’s weekly metrics report went from two days to two hours via a Sunday-night scheduled task and three Skills. Annabel’s event builds run through an hourly dispatcher that routes requests to five specialist Skills — and a separate audit agent that submits a live test registration before marking anything done.

The weekly cadence. In an Anthropic webinar, Travis Bryant and Brittney Tong demo the routine: a daily briefing before the first meeting, a Friday forecast pulled from Salesforce and BigQuery in the format leadership already expects, and an overnight run that scored 4,000 accounts.

The argument in the room

It’s worth putting this next to a skeptic. Marketo cofounder Jon Miller recently argued that legacy marketing platforms can’t survive by bolting AI on. They were architected 15–20 years ago around rules-based logic, lead-centric data, and email-first, form-fill-to-handoff workflows. Adding an AI add-on, he says, “paves the cowpath” — it speeds the work but can’t make the platform reason in buying groups. His four marks of genuinely AI-native: headless operation, agents that reason over shared context, decision traces and shared memory, and personalization through intelligent assembly.

Line that up against the Anthropic examples and it’s nearly a checklist. Claude threaded via MCP is the headless layer. Skills pulling live context are the reasoning. Policy-validated proposals, proofread numbers, and “flag the mismatch instead of guessing” are decision traces. Overnight account scoring is intelligent assembly.

The interesting tension: Miller says legacy platforms can’t escape the cowpath, and Anthropic kept its stack. But that’s not really a contradiction — their rails are already modern, and the AI-native work happens in the context and memory layer regardless of stack age. The sharp version of the claim isn’t rip-and-replace or bolt-on. It’s: build the reasoning-and-memory layer either way.

The human stays in the loop — by design

The most underrated detail across all of this is that the human doesn’t disappear; the loop is engineered to keep them in it.

It’s structural, not willpower. Follow-up drafts sit in your inbox, and if you don’t send them, they resurface in tomorrow’s brief. The 24-hour SLA holds because the workflow nags.

Verification gets delegated to a fresh agent. Annabel’s audit agent starts with no prior context, runs a live test, and only then marks the task complete — with a human confirming each result. Machine checks machine; human confirms.

On ambiguity, the system escalates. When a sales reorg broke the reporting reconciliation, Claude flagged the mismatch and asked how to handle it rather than guessing. And the model’s own refusals count too: when Jared fed his voice-matching feature a sequence of increasingly angry emails, Claude mirrored the tone, then refused to keep generating them — which he read as the feature working.

The net effect is that the human role moves up, not out. Forecast calls become “where do AEs need help,” not data-scrubbing. Analysts spend recovered hours on the data layer and on helping others frame their own questions.

What Anthropic didn’t have to carry

Here’s the part that matters most if you don’t work at an AI-native company.

Anthropic is a roughly three-year-old commercial org that built its stack in the AI era. It reached this point without paying three debts that define the effort everywhere else — and for most companies, those three debts are the entire project.

No data debt. There’s no 15-year-old CRM carrying dead custom fields, five definitions of “active account,” and years of duplicate records. When Anthropic says “Claude threads the stack,” the stack is already coherent enough to thread. Miller’s own thesis cuts both ways here: if legacy platforms fail on architecture, then Anthropic’s real advantage is having no legacy architecture to fight. As Jeff Kew noted under Miller’s post, clients keep discovering that “defined data governance across the board is required first.”

No org or process debt. No decades of siloed teams, entrenched tool ownership, or turf to negotiate. A small, technical, dogfooding-by-culture org can move a whole GTM team onto a plugin in a way a large, federated company struggles to. That velocity is a function of size and culture — not the tooling.

No adoption debt. They never had to sell the change internally or train people against resistance. The builder and the buyer were the same people: those who felt the pain built the fix and adopted it themselves. Jared’s tool spread org-wide in 24 hours via a Slack post because people wanted it. Contrast that with Ben Cirillo in Miller’s thread: “getting me to implement AI-powered workflows is easy, me getting everyone else to do it is going to take some time… one of the major switching costs is change management.” Or Robert Rose: customers pay for “outcomes minus switching costs” — and switching cost is exactly what Anthropic didn’t have to pay.

So read these examples as a clean-slate proof of concept. They show what’s possible when data, org, and adoption are already coherent — not how you get there from a messy, siloed, resistant start.

For everyone else, the real roadmap is the prerequisite work that Anthropic’s clean-slate start spared them — the spec work to define it, a clean semantic layer, a reliable identity graph, and durable memory — plus the change management to bring people along. Build that substrate first. The Skills come easy after.

Anthropic’s GTM story is a proof of what’s possible, not a map of how to get there. The tooling is the demo. The substrate — data, identity, memory, and the spec that defines them — is the work. Build that first.