--- title: "Your Marketing Tools Now Take Instructions From AI Assistants. What to Set Up First." description: "Six tools a marketing team already uses now document a way to let an AI assistant work inside them. What each connection can read and change, what to set up, and who approves before anything goes live." date: 2026-10-15 author: "David Wippel" tags: ["ai-ops", "ai"] url: https://www.essentialcode.eu/blog/connect-ai-assistants-to-your-marketing-tools/ --- # Your Marketing Tools Now Take Instructions From AI Assistants. What to Set Up First. > Six tools a marketing team already uses now document a way to let an AI assistant work inside them. What each connection can read and change, what to set up, and who approves before anything goes live. **TL;DR.** Your CMS, CRM, analytics, design and work-tracking tools have each published a way for an AI assistant to work inside them. The connections differ in what they allow: one can only read, one asks you to pick an access level, one requires confirmation before a deletion. What decides whether any of it helps is whether the team shares one setup or buys a subscription each. Google Analytics publishes a connection that lets an AI assistant run reports for you, and it can only read. DatoCMS publishes one that asks you to choose whether the assistant may publish content. Both shipped, both are documented by the vendor, and the difference between them is the decision your team has to make about every tool on your stack. ## What is actually new here? Three words come with this. - **AI assistant.** The chat tool your team already uses to draft and summarise. ChatGPT, Claude, Copilot and the rest. - **Connector.** A published way for that assistant to reach inside another tool, so it can look something up or change something there instead of only producing text for a person to copy. All six connections below use a shared standard called MCP, which is the plumbing underneath. - **Permissions.** What the connection is allowed to touch. Three things set it: the rights of the account you connect, the access level you choose during setup, and what the connection is built to do at all. The third one is where these connections differ from each other. Each vendor below documents its own access levels, confirmation prompts and admin controls, so the question for your team is what you are willing to let an assistant do inside each tool. ## The tools that already document a connection Six, read from each vendor's own current documentation. | Tool | What it can look at | What it can change | What you have to set up | | ---------------- | -------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Google Analytics | Your traffic and conversion reports | Nothing. This connection can only read | An account that already has access to those reports. Google labels the server experimental | | DatoCMS | Website content, images, and the way your content is structured | What the access level you pick allows: reading only, reading and editing including publishing, or everything your own account can do | The access level, chosen when you connect. It asks you to confirm each editing action, and the limits on what it can run are in an initial beta phase | | Storyblok | The content in the space you connect it to | Creates, updates and deletes content in that space | Confirmation before any deletion, which the vendor requires | | HubSpot | Customer records, the history of contact with them, and your campaigns and pages | Reading the records above | An admin of the account connects first, before anyone else there can. Custom sensitive data properties are excluded | | Figma | Your design files: the components, colours and layout inside them | Creates and changes content on the canvas | The remote connection rather than the desktop one, a Full seat to write outside your own drafts, and an assistant Figma supports. Writing to the canvas is a beta feature | | Asana | The tasks and projects your own account can see | Creates and manages tasks and projects | Sign-in with your own Asana account. On the two Enterprise plans, an admin can allow or block these connections | What differs between them is how much of the decision each vendor hands you. One rules out writing altogether. Two ask you to choose an access level when you connect, and one of those asks again before each edit. The others take their limits from the account you sign in with and from what the connection itself supports. If you want the setup detail for a CMS, we have written it up for [DatoCMS](/blog/control-datocms-from-claude-code/) and for [Storyblok](/blog/control-storyblok-from-claude-code/). ## One shared setup for the team Teams buy subscriptions for everyone and let each person figure it out alone. We recommend one shared setup for the team instead: a single connection at a known access level, with one written brief of your brand and buyers that everyone's work reads from. More on that in [why shared team context beats a copilot per person](/blog/what-we-do-not-let-an-ai-agent-do-in-a-clients-cms/) and [which part of an AI setup still belongs to you](/blog/which-part-of-your-ai-setup-belongs-to-you/). ## What to set up, in order 1. **Write the brief before you connect anything.** Your brand rules, who your buyers are, what a good page looks like, what the tone is. A connected assistant with no brief produces confident work in somebody else's voice, and this step needs no engineer and no budget. 2. **Start with one tool.** Pick the tool where your own team loses time to manual steps today. One working connection teaches you more than five half-finished ones. 3. **Choose the access level deliberately.** Read-only first. Reading tells you whether the assistant understands how your content is put together, and while the access really is limited to reading it cannot change what is there. 4. **Decide who approves, by name.** One person who reads what the assistant produced and either publishes it or sends it back, with the authority to send it back. 5. **Try it on a test copy first.** Storyblok's documentation asks for exactly this: "Verify any new automation in an environment", meaning a separate copy of the site that mirrors the live one. 6. **Then widen it.** Another tool, or write access on the first one, once the approval step has caught a few things and you know what it catches. ## What can go wrong Each of these is documented by the vendor. - **The same request can produce different results.** Storyblok states it plainly: "Because the LLM mediates every call, results are non-deterministic and unsuited for repeatable, high-volume work." Their own recommendation for bulk and scheduled work is their command-line tool instead. Treat a connection as the way to make a single change or produce a draft, and keep scheduled jobs on a script. - **A connection can be wired to run unattended.** Storyblok's documentation tells you not to: "Don't hide the prompt or run the server inside autonomous automation." The prompt asking you to confirm a deletion only protects you while a person is there to read it. - **The connecting account sets the ceiling.** DatoCMS states the boundary, "The MCP cannot perform actions that your role is unable to". Storyblok's own advice is to "Limit the connection to only the spaces and permissions the AI client needs", which is the step to take with any of them. - **Sensitive records need checking separately.** HubSpot excludes custom sensitive data properties, including personal health information, from its server. For every tool, the records an assistant can reach are worth a separate look before you connect it. - **Write access can be a seat change.** Figma's server needs a Full seat to write outside your drafts, so the person connecting it may not have the access the plan assumes. ## Who approves Somebody reviews the work after each significant step and before it goes live. That review is part of how the setup runs, and it is the one requirement none of the vendor documentation can configure for you. In practice it is a named person and a defined gate. The assistant drafts the page, builds the campaign structure, or pulls the report. A person on your team reads it, checks it against the brief, and publishes it or sends it back. For anything with legal exposure, pricing, or a claim about your product, that review is where those get caught. Write the brief, connect one tool read-only, name the person who approves, and only then decide whether write access earns its risk. Essential Code sets this up as one of the five choices you can pick for a [Sprint](/services/fractional-marketing-engineer/): the connection from your team's assistant into your CMS, the written brief behind it, and a person checking each change before it goes live. ### Sources Vendor documentation, each checked 15 October 2026. - [Google Analytics MCP server](https://github.com/googleanalytics/google-analytics-mcp) - [DatoCMS MCP server](https://www.datocms.com/docs/mcp-server) - [Storyblok MCP server](https://www.storyblok.com/docs/tooling/mcp-server) - [HubSpot MCP server](https://developers.hubspot.com/ai-tools/mcp) - [Figma MCP server](https://help.figma.com/hc/en-us/articles/32132100833559-Guide-to-the-Figma-MCP-server) - [Asana MCP server](https://developers.asana.com/docs/using-asanas-mcp-server)