
This month, marketing software vendors have been racing to plug their tools into ChatGPT and Claude. Connecting takes one sign-in. The harder job is deciding what each connection may do. The safest starting point is that it can read and draft but cannot publish.
We build an AI marketer, so we have a stake in this.
Connecting is now the easy part
MarTech’s news roundup, published on 12 November, lists a run of vendor launches. They connect marketing tools to AI assistants, and many are dated 11 November. Social9 lets people plan, write and schedule social posts from an AI chat. First Insight passes purchase-intent data to outside assistants. Marcora sends brand and product context into ChatGPT and Claude.
Most of these use MCP, a standard way for an assistant to call another tool. These are vendor announcements, so they show what is being built, not how many teams use it.
Reading is not the same as acting
A connection can do three different jobs. It can read your information, draft something from it, or act, which means scheduling or publishing. Each job carries different risk.
If an assistant reads a stale fact sheet, you see the wrong draft and fix it. If it publishes, the mistake is public before anyone looks. Yet many connect screens ask one yes or no question, so teams grant all three at once.
What careful defaults look like
Social9’s release is useful because it spells its defaults out. According to MarTech Series on 5 November, new connections start in test mode. There, assistant-made posts are labelled and checked, and never sent to a network.
Immediate publishing and post approval are off by default. Account owners choose which brands an assistant can reach and which actions it can take. Each connection keeps a 90-day activity log that includes refused requests.
We have read the announcement, not used the product, so treat it as an example of good questions rather than a verdict. Whatever tool you use, ask the same ones.
Decide who signs off, and on what
Switching publishing off is only half the job. Someone has to own the step that replaces it, and that person needs time and the authority to say no. An approval queue that nobody reads is a delay, not a control.
Social9’s announcement describes one useful habit: prices, dates and numbers that are not in the brand’s fact sheet get flagged for checking. You can copy the idea by hand. Ask reviewers to look hardest at figures, claims and names. A wrong one does the most damage.
Keep the reviewer close to the brand. Choose a reviewer who knows the voice and which claims you can support.
Start small and widen on evidence
The first connection should be boring. One brand, one channel, drafts only, with a named reviewer. Run it for a few weeks and keep a note of what the reviewer changed.
If the same fixes keep appearing, the problem is the reference, not the assistant, so fix the source before widening access. If the drafts need little editing, you have evidence to extend the permission one step. Change one thing at a time, so you know which change caused a problem.
Scope the connection before you connect
You can do this without any particular product. Work through this list before you approve the next connection.
- Name the brands and channels the assistant may touch. Start with one of each.
- Begin read only. Let it see the brand facts and nothing else.
- Allow drafts next, saved where a person can see them before anything is queued.
- Keep scheduling and publishing as a separate permission. Leave it off until drafts have been good for several weeks.
- Ask for an activity log that includes refusals. A refusal means a boundary held, or that the task was scoped wrongly.
- Review every connection monthly, and whenever someone leaves the team.
Reading is only as good as what is read
A tight connection to a weak source just produces wrong answers faster. If the brand facts are out of date, every assistant you connect will repeat them.
So keep one reference for the brand, and give each tool read access to that, not its own copy. Put an owner’s name on it and a date on the last review.
Where we stand at Mavic
In Mavic, each brand has its own workspace with a Brand DNA, which Mavic reads on every generation. Team members hold a role: Owner, Admin, Editor or Viewer. Viewers can only look, while editors can create, schedule and publish.
Nothing goes live without the user’s approval, and we say so plainly because it is the point. Tandem splits one brief across specialist agents, and each one works from the same saved Brand DNA.
The limits matter too. The ban list is a strong nudge, not a hard filter, so a person still reads the work before approving it. We would not tell you any setup removes that step.
Sources
- MarTech, “The latest AI-powered martech news and releases”, published 12 November 2026. Vendor announcements, not adoption data.
- MarTech Series, Social9 MCP announcement, 5 November 2026. Product specifications from the announcement, not independently tested.
- Mavic Tandem.