Scalewell / Blog

Claude for marketing: where a paid media team should start

A paid team's first month with Claude, from business seats and one Project per account to a first skill, one read-only connector and one builder.

Five different tools laid out in a row on a desk, all connected by a single glowing thread to one key.

Claude for marketing works best when a paid media team treats it as five tools under one login: Projects for account context, Skills for repeatable methods, connectors for live data, and Cowork or Claude Code for building. Start with business seats, one Project per account and one skill, then add read-only data in week three.

You've probably already got the problem this solves. Three people on your team pay for personal Claude accounts on expenses, one of them pasted last month's Google Ads export into a chat, and now someone from legal or finance wants to know which plan you're on and whether that data trains a model. Nobody has a clean answer, because nobody chose a set-up. Each person just started typing.

This guide is for the marketing leader who has to choose that set-up. It explains which part of Claude does which paid media job, which plan and data terms apply as of November 2026, and a four-week order for switching things on. If you want the list of skills to build, our post on Claude skills for marketing covers that, and this one stays on orientation and the first month.

Claude has five parts, and each one fits a different paid media job

Claude's five parts for a paid team are Projects, Skills, connectors, Cowork and Claude Code, and they differ in what they remember, what they can reach and who should use them. Plenty of teams only use the chat box, which is a bit like buying a car for the cup holders.

PartWhat it doesPaid media usePlans (November 2026)
ProjectsA workspace with its own chat history, files and instructionsOne per account or client: ICP, conversion definitions, naming rules, last quarter's resultsAll plans; Free is capped at 5 Projects
SkillsA folder of instructions and scripts Claude loads when a request matchesThe weekly readout, search term review, pacing check, done the same way every timeAll plans; needs code execution switched on
ConnectorsLinks to outside tools through the Model Context Protocol (MCP)Reading Google Ads, Meta, HubSpot or GA4 data without exportsAll plans; Free gets one custom connector
CoworkClaude working on your files, connectors and browser from the desktop appBuilding a report pack from several exports, filling a launch sheetPro, Max, Ultra, Team, Enterprise
Claude CodeThe same agent in a terminal or editor, for scripts and repeatable jobsBulk ad copy files, scheduled checks, small internal toolsPro and above

The plan details come from Anthropic's pricing page and the help articles on Projects, Skills and Cowork. The order of the table is also the order we'd switch them on, because each row gives Claude more reach into your accounts than the one above it.

A five-rung ladder where each rung gives an AI assistant more reach, from uploaded documents at the bottom to a machine running jobs at the top, with one person standing beside the top two rungs.
From bottom to top, the rungs are Projects (context you upload), Skills (methods you write), connectors (live data, read first), Cowork (acts on files and the browser) and Claude Code (runs scripts and jobs). In month one, the top two rungs belong to one named builder only.

Buy business seats before anyone pastes account data

Claude's consumer plans (Free, Pro and Max) let Anthropic train on your chats unless each person opts out, while Team and Enterprise run under commercial terms with no training on your content by default. That single difference is the reason to move off personal accounts before you scale anything.

Anthropic's 2025 consumer terms update says the training choice applies to Free, Pro and Max, including Claude Code used from those accounts, and that people who allow training have their data kept for five years instead of the 30-day period. The same page says the change does not apply to services under the Commercial Terms, which covers Team, Enterprise and the API. So a campaign export pasted into a personal Pro account sits under different rules from the same file in a Team workspace, and the person who pasted it may not know which way they answered the prompt. Scalewell works with paid media teams so everyone knows which workspace their files belong in and what Claude can take off their week once each part is switched on.

Here's how the plans compare for a paid team, using prices listed on the pricing page in November 2026.

PlanPriceWhat matters for a paid team
Pro$20 a month, or $17 a month billed annuallyClaude Code, Cowork and Claude in Chrome; training is opt-out
Team, standard seat$25 per seat a month, or $20 billed annuallyTeams of 2 to 150, SSO, admin control over connectors, shared Projects, no training by default
Team, premium seat$125 per seat a month, or $100 billed annuallyFive times the usage of a standard seat
Enterprise$20 per seat a month billed annually plus usage at API rates, 20-seat minimumLarger organizations with procurement and security reviews

Imagine a hypothetical team of six: a head of paid, three channel managers, an analyst and a marketing ops lead. On standard Team seats billed annually that is six times $20, or $120 a month, and you can mix seat types, so the analyst who runs long data jobs could take a premium seat while everyone else stays on standard. It also ends the expenses-claim version of AI policy.

Week one: one Project per account holds the context Claude would otherwise guess

A Claude Project gives each ad account or client a self-contained workspace with its own files, chat history and instructions, so Claude stops answering from generic assumptions. Without one, every chat starts from zero and Claude fills the gaps with whatever a typical B2B SaaS company might do.

Put five things in each Project. First, your ICP and the buying committee you sell to. Second, your conversion definitions, including which events count as a lead, an MQL and an opportunity in your CRM. Third, your campaign naming convention, because Claude can't read a campaign called "LI_ABM_T1_Q3_v2" without it. Fourth, last quarter's results by channel with spend and pipeline. Fifth, a short project instruction that tells Claude how you want answers: which metric to lead with, what to flag, what never to recommend.

On paid plans, Anthropic's help article says Projects switch to retrieval mode when the knowledge nears the context limit, which expands capacity by up to 10 times. On Team and Enterprise you can share a Project with "Can view" or "Can edit" permission, so the account context lives in one place instead of six people's private chats.

By Friday of week one you should have one Project per main account and a habit of starting work inside it. It's dull set-up work, and it's the part that decides whether anything later in the month produces answers specific to your accounts.

Week two: turn the most repeated job into one skill

A Claude skill is a folder of written instructions, and optionally scripts, that Claude loads when a request matches its description, so a job like the weekly readout gets done the same way every time. Pick the job your team repeats most and argues about least, which for many paid teams is the Monday performance summary.

Skills need code execution switched on: individuals turn it on under Settings, then Capabilities, and on Team the setting is on by default at the organization level. You upload a skill as a zipped folder under Customize, then Skills, and Team and Enterprise owners can provision a skill for everyone so it appears in each person's list.

You may be tempted by Anthropic's own marketing plugin, which bundles commands such as /campaign-plan, /performance-report and /seo-audit with skills for brand voice and competitive analysis. It's worth reading as an example of how skills are written. Its connector list includes HubSpot, Supermetrics, Ahrefs, Similarweb and Klaviyo, and no direct Google Ads, Meta or LinkedIn ads connector, so its performance report knows nothing about your accounts until you connect a data source or give it your exports. Treat it as a starting draft that you rewrite with your own conversion definitions.

Week three: connect one data source with read-only access

A connector lets Claude read live data from another tool through MCP, which replaces the export, paste and reformat routine many teams run every Monday. Connect one source in week three, read-only, and leave anything that can change spend until your team trusts the output.

On Team plans, only Owners and Primary Owners can add custom connectors, and members then connect individually, which is the control you want. Anthropic's help page also warns that custom connectors reach services it has not verified and that "Malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions." So read what a server can do before anyone clicks "Allow always."

As of November 2026, the main ad platforms offer different routes:

  • Google Ads: Google's open-source Google Ads MCP server reads account data through GAQL queries. Google's developer page now lists it as "Read and partial mutate", and the GitHub README for version 0.0.6, released November 2, 2026, adds write tools for assets and labels, each behind a "Propose & Confirm" step and each one you can switch off. It has no tool for bids, budgets or campaign status.
  • Meta: Meta runs a hosted ads MCP server at mcp.facebook.com/ads with seven tool groups, including reporting and the ability to create and edit campaigns, ad sets and ads. That second group is why read-only habits matter here.
  • LinkedIn: As of November 2026 LinkedIn's own MCP server, which reads and writes ad accounts, is in beta and already in use but not yet publicly documented. Its one documented route is a read-only beta LinkedIn Ads plugin inside ChatGPT. Third-party connectors exist, and our LinkedIn Ads MCP guide compares them.

For a B2B team, we'd usually start with the CRM rather than an ad platform, because pipeline lives there and the ad platforms already have reporting screens. Our post on MCP for marketing sets out which connections a paid team needs and in what order.

Week four: give Cowork or Claude Code to the one person who will build

Cowork and Claude Code let Claude act on files, folders and scripts instead of answering in a chat, and in month one they belong with one named person who likes building things. Rolling them out to the whole team on day one risks a folder of half-finished experiments and nobody who owns them.

Anthropic's own growth team shows what that one person can do. In a January 2026 write-up, Anthropic describes growth marketer Austin Lau using Claude Code to build a Google Ads copy workflow that produces a CSV ready to upload "after undergoing manual review," and a Figma plugin for ad variants. The post says "What used to take 30 minutes per ad now takes 30 seconds." Read that as a vendor describing its own team, and notice the detail that matters for you: a person still reviews the file before it reaches Google Ads. His advice is also the right brief for your builder: "just start with a very, very small and easy thing."

Cowork is the easier entry point for a marketer, since it uses the same agent as Claude Code with no terminal. Anthropic's Cowork guide says it can read and write your local files and "can take real actions on your behalf," and it asks for permission before deleting files. If your builder wants to go further, our post on Claude Code for marketers covers that route without assuming they code.

Usage limits and missing context are the limits you'll hit first

Claude's paid plans have a session limit that resets every five hours and a weekly limit, and Anthropic's usage limits article says "Tools and connectors are token-intensive." So a connector-heavy audit run on Monday morning uses up a session's allowance faster than the same question asked in a plain chat.

Three other limits show up in the first month. Claude only knows what you give it, so an answer about LinkedIn performance without CRM data is an answer about clicks and form fills. Claude can get arithmetic wrong in long tables, which is why good skills move the sums into scripts. And Claude will sound equally sure when it's right and when it's guessing, so the review step stays in place for anything that changes spend or reaches a customer. None of this is unusual for AI tools, but it surprises people who expected a finished analyst.

Teams that should wait before rolling out Claude

Some paid teams will get little from Claude in the first month, and it's better to know that before buying seats.

  • Teams with broken conversion tracking. Claude will read your bad data quickly and confidently. Fix offline conversion imports and CRM matching first.
  • Teams whose agency owns the accounts. If you don't have admin access to the ad platforms or a contract that says who does each job, connectors will stall at the permission screen.
  • Teams without a legal or security sign-off. If your company hasn't approved an AI vendor, a Team plan bought by marketing alone can end up switched off three weeks later.
  • Leaders expecting a pipeline result in 30 days. The first month is for buying back hours and catching errors. Pipeline effects take longer to show, as our 90-day rollout plan explains.

You'll know the first month worked when Monday changes

The first month with Claude has worked when the weekly report comes out of a skill instead of a spreadsheet, the account context lives in shared Projects, and one connector replaces at least one export. Nobody on the team should still be pasting data into a personal account.

Measure it with two numbers you can collect yourself: hours spent on the weekly report before and after, and the number of errors caught in review. If both moved and nobody has gone back to the old way, add the next skill and the next connector. If neither moved, the problem is usually the Project context or the data, and more tools won't fix it.

Where to start this week

Move your team onto business seats, build one Project for your largest account, and write down how your best person does the Monday report so it can become your first skill. If you'd like help setting Claude up around how your paid team already works, that's the kind of work Scalewell does.

Frequently asked questions

Is Claude better than ChatGPT for marketing?

Neither is better for every marketing team. For a paid team the choice usually comes down to which vendor your company has approved, which of your ad platforms and CRM each one can connect to today, and which assistant your team will actually use. Pick one, set it up properly, and judge it on hours saved and errors caught.

Can Claude make changes in Google Ads or LinkedIn Ads?

Claude can only change an ad account through a connector that has write tools, so the question is which tools you switch on. Google's Google Ads MCP server can run strictly read-only if whoever installs it turns off its label and asset tools in the server's configuration file, which is a sensible default for month one. On LinkedIn, the official MCP server can write but is in beta and not yet publicly documented as of November 2026, and LinkedIn's ChatGPT plugin is read-only, so most write access today comes from third-party connectors. Read any connector's tool list before approving it. Keep every change that touches spend behind a named human approver.

Can I use Claude for free for marketing work?

Claude's Free plan includes Skills, connectors and up to five Projects, so you can test the ideas in this guide at no cost. It lacks Claude Code and Cowork, has tighter usage limits, allows one custom connector, and falls under consumer terms, so keep company and client data off it.

Does Claude remember my brand between chats?

Claude has a memory feature on every plan (on Team and Enterprise, owners decide whether it's on), but for a marketing team the dependable way to keep brand and account context is a Project. Instructions and files in a Project apply to every chat inside it, and on Team plans everyone with access sees the same context, which personal memory can't offer.