Read the Gate, Then Read the Second Page

The decision

If you are about to connect a marketing agent to ads, email, or a store, the first question is not which model it uses. It is what the page says happens before something publishes, sends, or spends.

Between September 28 and September 29, 2026, four public pages described that gate. They are not the same product, and two of them disagree with a second page about the same product. This note is a reading of those pages, fetched September 30, 2026. Nothing here was connected, run, or spent.

What to write down before you connect anything

For each tool, copy four lines off the page, in the vendor's words:

What you hand it (a goal, a brief, a URL, a plain-English rule).
The sentence that says what cannot happen without you.
Who can use it, if the page says.
The price, or the fact that no price is there.

If a second page exists, copy those four lines again. If they conflict, keep both. Do not average them.

What the pages actually say

Tool	Input the page names	Gate, in the page's words	Access	Price on the page fetched
Muse for Small Business	A goal, plus connectors you turn on	"nothing publishes, sends, or spends without your approval"	Muse, per the same post, is a personal agent in the US and Canada. The business post does not say that changed.	Not stated
Klaviyo inside Muse	Connect in the Muse app, authorize, then ask	Welcome series: drafts copy, sets flow logic and triggers, "hands you something finished to approve." Opening also says you can schedule a send in the same conversation.	If you use both	Klaviyo "start for free" is Klaviyo's line, not a Muse price
Pencil Scribble	A brief	Review the plan before it starts. Edit the work after.	Public launch, New York, September 28, 2026	No Scribble price. Pencil annual plans: Core $11/mo, Growth $44/mo, Pro custom. Monthly figures did not render.
Multiply Ad Spend Recovery Agent	Plain English, turned into a schedule you can edit	"Any action that can affect spend requires human approval before it is applied."	Launch release, September 28, 2026, 15:49 ET. Homepage did not load.	Not stated
Thrad Autopilot	URL or product feed	"Nothing goes live until you submit it."	Product page: self-serve, no minimum. ExchangeWire, same week: selected customers beginning September 29, 2026.	Media on CPM or CPC. Managed service from $10k/month.

Sources for each row are at the bottom. Vendor savings claims are not in the table on purpose.

Where the pages slip

Schedule is not the same word as approve. Meta's newsroom line is clean: nothing publishes, sends, or spends without approval. Klaviyo, a named launch partner, tells you to connect inside Muse and then describes both an approval handoff and scheduling a send. If you use that connector, put Meta's sentence in the first instruction anyway. Then check, in the product, whether "schedule" waits for the approval the newsroom promised. The blog post does not show you that screen.

A brief is not a media buy. Scribble's launch post is specific about the jobs: resize and adapt, remix video, make versions, apply Brand Library rules. The link from creative to media buying is "towards the end of the year." If a draft asks Scribble to buy media, it is asking for a feature the post says is not current.

Spend approval is narrower than "human in the loop." Multiply's release holds actions that can affect spend. Ask Dot can also propose responsive-search-ad copy. The release does not say every copy edit is held the same way. Block negatives from the Actions Feed if that is the job. Do not assume the pause button and the copy button share one gate.

Access has to survive a second page. Thrad's product page, fetched September 30, does not say "selected customers" and does not date the launch September 29. ExchangeWire's press box, dated September 29 and fetched the same day as the product page, does. Self-serve with no minimum, and selected-customer access, cannot both be the whole story until one page catches up. Quote both, or quote neither.

Numbers that are not yours to repeat

Leave these out of any note you would stand behind, unless you label the vendor:

Pencil's 60–80% time reduction on resizing and product swaps. Testing claim, in the launch post.
Multiply's 30% of budget, and $126,000 average wasted spend. Multiply's figures, from prospect data, in the release.
Multiply's "up to 700%." Early customers report. It sits in the company description, not in a method section.
Klaviyo's 40% conversion drop and the code that expires at 11:47 p.m. The page presents that as a prompt scenario. A block above the essay is labeled "Generated by Klaviyo AI." Neither is a case study.

What to do next

Pick one tool you already have an account for. Copy the four lines above into the brief, in the vendor's words, with the URL.
If the tool is Muse, add the approval sentence as the first instruction, then connect only the accounts you mean it to see. The newsroom list is Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe, Zoom, plus Facebook and Instagram business accounts. The full list, Meta says, is in the Muse app settings.
If the tool is Scribble, keep the job inside resize, adapt, remix, and versions. Do not add media buying.
If the tool is Multiply, use the Actions Feed for negatives, and do not let a spend change apply on a schedule you have not opened.
If the tool is Thrad, do not start from the press box alone. Open the product page, write down which access line you were actually offered, and treat $10k as the managed-service floor, not a self-serve minimum.

What this is not

This is not a test. No account was connected. Adobe's interactive editors in ChatGPT were reported the same day by Agile Brand Guide. Adobe's own announcement page was not opened, so Adobe is not in the table. Usermaven's Maven AI 2.0 writeup did not load. A model directory row for Claude Sonnet 5.5 was not added: Anthropic's news index, Sonnet page, and models overview fetched the same morning list Sonnet 5 at $2 / $10 per million tokens, and do not list Sonnet 5.5.