Lindy's AI teammate waits for approval before anything ships

The setup

Lindy markets itself on its own homepage, as retrieved on 16 September 2026, as an AI teammate that lives inside a company's Slack, connecting to tools such as Gmail, Notion and HubSpot and to more than 1,000 other integrations, including any MCP server. Setup, per the site's own FAQ, is meant to take minutes: an admin connects Slack, and from there anyone in an approved channel can mention Lindy to assign it a task, with no separate installation for each teammate. The company's own documentation describes the same product as an AI employee built to be shared by a whole team rather than configured individually by each person who uses it.

What the documents show

Both the marketing site and the documentation describe the same review step in near-identical language. The homepage FAQ states plainly that Lindy will not do things without asking: anything with outside impact waits for a named approver, such as sending an email, updating a ticket, posting to another channel, or publishing a doc. The documentation frames the same behavior as a workflow — write actions always wait for approval, so nothing ships without the user — while read-only lookups from sources the workspace has already approved happen without a pause. Both are vendor descriptions of Lindy's own product, not an independent test of whether the approval step holds up under real use.

The friction

The scope of outside impact is defined by Lindy, not configurable in the FAQ's own examples, and the documentation is explicit that this applies specifically to write actions rather than everything the agent does — a distinction a team has to understand before assuming every Lindy action pauses for review. The homepage also discloses a shared-credit billing model, where every seat draws from one workspace pool and Lindy pauses with a clear message if the pool runs out mid-task, a limitation with an operational cost distinct from the approval mechanism itself.

What changed in the work

The documents support a narrower description than a hands-off AI employee: a team can hand Lindy a standing task, such as watching a support queue or drafting follow-ups, while any step with an external effect still waits for a specific named person, not just anyone available. That is an editorial reading of the setup cost as documented — the approval step means the team retains a review workload proportional to how many write actions Lindy is asked to take, which the vendor's own materials do not attempt to quantify.

• Which of this workflow's steps count as a write action under Lindy's own definition, and which are read-only?
• Who is the named approver for each channel Lindy has access to, and are they actually available when a request lands?
• What happens to a paused task if the shared credit pool runs out before the approver responds?

A vendor that documents its own approval gate in the same terms on its marketing page and its technical docs is at least being consistent about where the human checkpoint sits, even if neither document says how often that checkpoint gets rubber-stamped in practice.