AI GTM: how to run go-to-market from Claude, and which jobs have to leave the chat

The pages that rank for "AI GTM" define the term, list platforms and sketch a strategy. None of them shows how to actually run the motion from the assistant most teams already pay for. This page does that, stage by stage, with prompts you can paste. It then says which jobs a chat cannot hold, because something has to happen while nobody is typing.

It is written by Tulina, an agentic RevOps workflow builder, and the section that uses Tulina is marked. Everything above it works without us. Anthropic's product facts below were read on its own documentation on 31 October 2026, and the figures about our own agents were counted that morning.

What is AI GTM?

AI GTM is the use of AI models and agents to do go-to-market work: finding and researching accounts, building and cleaning lists, writing outreach, logging calls in the CRM, following up, and reporting on the pipeline. It is not a product category. It is a way of running the same motion with fewer people moving data between tools.

The difference from the automation most teams already have is judgement. A rule in a Zap or a workflow fires on a field value: stage changed, form submitted, date passed. An AI step reads something with no fixed shape (a job post, a call transcript, a reply, a company's own page) and decides what it means: is this a buyer, is this reply a yes, which deal does this call belong to. Rules still do the moving. The model does the reading and the deciding that used to need a person.

"Agentic GTM" is the same idea one step further: instead of a person asking a model one question at a time, an agent runs the whole job. It calls the tools in order, checks its own work against the CRM, and stops at the point where a person has to decide.

What does running GTM from Claude look like, stage by stage?

Running GTM from Claude means each stage of the motion becomes a conversation in which Claude reads and writes your tools through connectors, instead of you copying data between tabs. Anthropic's help page on connectors says they "let Claude access your apps and services, retrieve your data, and take actions within connected services". Claude inherits each person's permissions from the connected service, so it can reach what you can reach and nothing more.

Every row is a job Claude can do in one sitting, while you watch. That is the strength and the limit of running GTM from a chat, and the rest of this page is about both. For the rep's share of it, call by call, Claude for sales walks through a seller's day, the skills Anthropic ships for it, and which of them need a schedule.

How do you run GTM from Claude yourself?

You can run most of the motion from Claude with nothing but its own connectors and a few habits. The steps below assume a paid Claude plan and work in Claude or Cowork.

1. Connect the tools the job touches, and only those. Add the CRM, your email, your call recorder and your sequencer from the connectors directory. Every connector you add is one more system a prompt can write to, under your name.
2. Write the context down once. Keep your ideal customer profile, your offer, the roles you sell to and two openers you liked in one document Claude can read, so every conversation starts from them instead of from a paragraph you retype.
3. Make Claude show its reading before it spends anything. For a list, ask it to print the filters it read from your sentence and wait for your go. A wrong headcount band caught here costs five seconds; caught after the search, it costs the credits.
4. Check the CRM before anything is enriched. Ask Claude to drop customers, open deals and anyone in a colleague's sequence first, matched by email or company domain, then look up emails only for who is left.
5. Ask for a source on every claim. An opener that cites a job post with its date can be checked in a second. One that says "I noticed your impressive growth" cannot, and the buyer can tell.
6. Stage, never send. Have Claude add people to a paused campaign or save drafts in your inbox. The send stays a person's click.
7. Save the prompt that worked. The second time you run a job, it should be the same prompt with a new date, not a new conversation from memory.

A prompt for the list-building stage, which works with any company database and CRM you have connected:

Build a target list from this sentence, and do not send anything:
"[B2B software companies in France, 50 to 500 employees, that posted a
sales development job in the last 30 days]."

1. Print the filters you read from the sentence and wait for my go.
2. Search for matching companies. Give me the count first. Score each
   one on what the search returned and keep the top two tiers.
3. For each company kept, list two or three people who own [sales
   development or revenue operations]. Do not reveal emails yet.
4. Check every person against my CRM by email or company domain. Drop
   customers, open deals and anyone already in a sequence.
5. Find a work email for each person left.
6. Write one opener per person that cites one dated fact and its URL.
7. Give me a table: company, person, title, email, opener, source.

This is how a lot of teams start, and for a weekly list or a single account brief it is enough.

Which GTM jobs have to leave the chat?

The jobs that have to leave the chat are the ones that must happen when nobody is in it: on a clock, or the moment something arrives. A conversation only runs while someone is in it. A signal sweep that should be done before the team sits down, a positive reply that should reach its owner within the hour, or a pipeline brief due every Monday at 8 all need something that starts without being asked.

Anthropic has two answers to that. A Cowork scheduled task saves a prompt and runs it "at the cadence you choose", as its own session, with your connectors, and it runs "even when your computer is asleep or the Claude Desktop app is closed". It is available on all paid plans. A Claude Code cloud scheduled task, formerly a routine, can also start on an API call or a GitHub event, which is how an inbound event reaches it. Anthropic labels it a research preview. Both run under one person's Claude account, and our guide to running Claude on a schedule compares them in detail.

Most teams end up with both: Claude in the chat for the work that needs a person thinking alongside it, and agents for the work that only needs to be done, every day, the same way.

How do you run GTM from Claude with Tulina?

With Tulina, the job you ran in a Claude chat becomes a saved process that a hosted agent runs on a schedule or a webhook, through connectors to the tools your team already uses, with every step, every tool call and its cost recorded in a run anyone in the team can open. You still start it from Claude: the same conversation that ran the job once can save it and schedule it. A person still decides the sentence, the rubric and the launch: the run stages a paused campaign and never sends.

Tulina is an agentic RevOps workflow builder: AI agents that run sales and GTM processes on a schedule or a webhook, across the tools a team already uses, with every step and its cost visible. Once Apollo, HubSpot and Lemlist are connected, paste this into Claude:

Run my target-list process in this chat once, then schedule it.

1. Fork the library process into my project:
   oto_procedure(op="fork", public_slug="plain-english-prospecting",
   scope="org", project_id=<my project id>).
2. Read the forked process and list every connector it names. Check
   each with oto_connector(op="list", name=<connector>) and tell me
   which still need a key. Do not guess at a tool name.
3. Run it now for this sentence: "[B2B software companies in France,
   50 to 500 employees, that posted a sales development job in the
   last 30 days]". Print the filters and wait for my go. Stop before
   anything is sent: people go into my paused Lemlist campaign only.
4. Show me the report: companies found and kept, people dropped by the
   CRM check, emails found by provider, and what each step cost.
5. When I say so, create a hosted agent that runs the process every
   weekday at 07:42 Europe/Paris: oto_trigger(op="create",
   procedure=<the forked slug>, cron="42 7 * * 1-5",
   tz="Europe/Paris", tools=<only the tools the process uses>,
   model=<the default served model>).

The process behind that prompt, with every step and the tools it uses:

The process in Tulina

Build a target list from a sentence

Logistics companies in France, 250 to 1,000 employees, hiring supply chain roles. Source 40 accounts.

One-sentence brief in chat

See it in a demo

Claude runs the rest autonomously

1. 1

   AI shortlists accounts and their people.

   Apollo.io

   Apollo.io

   Matching companies are scored on what the search already returned, and each one kept gets two or three people fitting the persona described.

2. 2

   AI drops contacts the CRM already has.

   HubSpot

   HubSpot

   Anyone the CRM shows as a customer, an open deal or in a colleague's sequence is dropped, matched by email or else by company domain.

3. 3

   AI gathers signals for what survived.

   TheirStack

   TheirStack

   Each remaining company gets its open roles and technology, recent funding, moves and expansions, and details from its own pages, each with its source URL.

   Serper

   Serper

   Apify

   Apify

4. 4

   AI finds verified emails and numbers.

   Apollo.io

   Apollo.io

   Each contact gets a verified address and direct number, a second source fills what the first missed, and anyone still unfound is marked unreachable, not guessed.

   FullEnrich

   FullEnrich

5. 5

   AI writes each opener from a signal.

   Each opener names one dated finding about that company and cites its URL, and a company with nothing found gets a plainer version, never an invented hook.

6. 6

   AI stages the campaign, updates the CRM.

   Lemlist

   Lemlist

   The sequence sits as a draft with each contact's own opener, the qualified list is in the CRM with its signals as a note, and the launch is left to a person.

   HubSpot

   HubSpot

See the full process

The order of those steps is the part worth copying, with or without Tulina. The CRM check comes before any paid lookup, so a company the team already knows costs nothing further. The opener is written only from a signal the run found and dated. And the run ends at a paused campaign, because a campaign that fills itself is fine and one that sends itself is a decision nobody made. For the version that starts from the signal rather than from a sentence, Outbound from buying signals scores new hiring and technology signals each morning.

What GTM work do we run on agents ourselves?

We run most of our own GTM on agents, which is why this page can say which jobs belong there. On 31 October 2026, Tulina's own workspace had 49 hosted agents switched on: 40 on a schedule and 9 started by a webhook from another tool. Grouped by the job they do:

So 39 of the 49 do a sales, marketing or customer job. The ones that stayed in the chat are the ones on the left of the table above: account briefs, call prep, a list for a single campaign, a draft to rework. Two smaller lessons from the schedules. 29 of the 40 fire at a minute other than :00 (07:42, 08:12). Anthropic's docs give the same advice for its own cloud tasks: a run scheduled exactly on the hour "can start several minutes late", so pick 9:07 rather than 9:00. And no agent sends a cold email on its own: each prospecting agent ends at a paused campaign or a draft for a person.

What AI GTM can't do, and what gets you in trouble

• It cannot fix a vague customer profile. Claude scores against the rubric it is given. A profile written in adjectives returns a confident list of the wrong companies, faster.
• It writes with your name and your permissions. A connector acts as the person who connected it. A prompt that can update the CRM can update the wrong deal, so give a job only the tools it needs.
• Unsourced personalisation is worse than none. A model asked to "personalise" will flatter. Make every opener cite a dated fact, or send the plain version.
• A chat forgets; an agent forgets too. A scheduled run starts from nothing. If it does not write down where it got to, tomorrow's run redoes today's work or skips a day.
• A green status is not a result. Anthropic's own docs say a scheduled run's status describes how the session ended, not whether the work was done. Read the run.
• Sending limits apply to the sender. Email deliverability and LinkedIn's limits do not change because a model wrote the message. Our page on LinkedIn outreach with Claude covers the LinkedIn side.

What are the best AI GTM tools?

The best AI GTM tool is the one that does the job you are missing, connected to the CRM you already have. The tools fall into a handful of jobs: a CRM as the record, a company and people database, a signal source, enrichment, a sequencer, a call recorder, and something that runs the steps between them. Our list of GTM tools sorted by job compares them with prices read on each vendor's page, and sales automation tools covers what rules alone can do. For the agent layer, AI agents for business compares the ways to run one. If you are weighing Clay for the research and enrichment layer, Clay pricing and a Clay alternative is the fair comparison.

What is an AI GTM engineer?

An AI GTM engineer is the person, or increasingly the agent, that builds and runs the systems turning buying signals into outreach: signal, enrich, qualify, write, stage, log. The role grew out of RevOps and growth teams working in Clay, n8n and the CRM. Running GTM from Claude is what the role looks like when the builder's tool is a conversation instead of a canvas. GTM engineering with AI agents explains the role and builds one loop end to end.

Is AI going to replace GTM teams?

No, and the agents we run show where the line sits. Agents take over the work of moving and checking records: finding accounts, matching them against the CRM, logging calls, routing replies, assembling reports. People keep the work that needs a decision someone answers for: who to sell to, what to offer, what to say to a real objection, and whether a message goes out at all.