Our read
Start Lindy as an assistant that prepares and asks, not as an autonomous operator. Lindy publishes that lookups run on their own and that anything with outside impact waits for your approval until you tell it to proceed. It does not publish which actions count as sensitive or whether permissions can be set per action, so you have to find that boundary in a trial. Keep customer-facing messages, payments and anything irreversible under approval.
Prepare, approve, send: where Lindy draws the line
Lindy presents itself as an agent that carries out tasks rather than only answering questions: pulling reports, updating CRMs, triaging support queues, preparing decks. It works from company data in connected tools, runs scheduled routines, keeps an editable memory and, in the vendor's words, asks for approval before sensitive actions.
The homepage reduces the mechanism to three steps: Lindy prepares the work, you approve it, then it goes out. Lookups just happen. Anything with outside impact waits for your OK, until you tell Lindy to go ahead. That last clause matters. Autonomy is something you grant, workflow by workflow, rather than a fixed trait of the product.
One gap is worth knowing about early. The security page says approvals are built in and that users and systems only touch what they are explicitly allowed to, but it does not list which actions count as sensitive. That list is something you will build yourself by testing.
Triggers and actions Lindy documents
On the trigger side, the pricing page lists scheduled routines, a Slack-native teammate that responds in threads and mentions, inbox management, replies drafted in your voice, and texting Lindy over iMessage. The vendor's description also mentions recurring reports. In practice, work starts either on a clock or when someone hands it over in Slack or email.
On the action side, Lindy lists 3,000+ integrations plus MCP. The catalogue record names Outlook, Gmail, Google Calendar, Docs, Drive and Sheets, HubSpot, Notion, Salesforce, Slack, Microsoft Teams and Shopify. The integrations page also lists Stripe, PayPal, QuickBooks and DocuSign. It names apps but does not describe read or write scopes per connection, so what an agent may do inside HubSpot or Salesforce is something to confirm when you connect them.
A trigger, permission, action and approval grid to draft (editorial recommendation)
The tiers below are our suggestion, not a Lindy feature. Tier one is read-only work on a schedule: a weekly report from Sheets or HubSpot, a meeting summary, a lookup. Lindy states that lookups just happen, so this is the natural place to let it run unattended.
Tier two is internal writes: updating a CRM field, adding a note in Notion, filing a ticket category. Begin with approval on and relax it one workflow at a time, once you have read enough proposed changes to trust them. Tier three is anything that leaves the building: emails to customers, messages to clients, invoice reminders. The homepage itself shows Lindy chasing overdue invoices by email and phone, which is exactly the kind of action to keep behind your OK at first.
Tier four is irreversible or financial: payments, contracts, deletions. Even if a connector exists, keep a person on the final click. Nothing published says approvals can be required for a named action only, so test whether your setup can enforce that before you rely on it.
Permissions, logging and the open questions
The security page states that you decide what Lindy can access, and that permissions can be set, changed and revoked at any time. It also says data is encrypted in transit and at rest, is never sold, and is not used by Lindy or its providers to train models. The vendor cites a SOC 2 Type II audit by Johanson Group; treat that as a vendor-stated audit, and request the report in your vendor review.
There is one inconsistency to clear up. The security page says every action Lindy takes is logged, while the pricing page lists audit logs as an Enterprise feature, alongside SSO, SCIM and HIPAA with a signed BAA. Ask what a Team workspace admin can actually see, and whether the record shows which approval allowed which action. Granularity of permissions per action is the second open point.
What two customer deployments show
Winnipeg Denture & Implant Centre uses Lindy for consultation notes generated once an appointment finishes, and for referral letters drafted and sent with custom formatting, plus cover when staff are away. The clinic reports saving more than ten hours a week and a 97% task completion accuracy out of the box. Those are the clinic's figures. The case study describes drafting and sending but does not spell out a review step, so it shows the type of work, not the approval settings.
Truemed's support lead built internal applications without engineering time: ticket classification, escalation routing, replies to routine inquiries and compliance-related answers. Truemed reports 36% of support volume automated and a cost per resolution that fell from $1.00 to $0.33. The bounded actions here, classifying and routing, are the easiest to hand over. Both cases are vendor-published and are not a comparison between each other.
Credits as a cost brake, not a behaviour control
Team costs $49.99 a month for 5,000 credits, and every active user in the workspace is billed. Credits form one shared pool, admins can set allocations per seat, extra credits cost $10 per 1,000, and the homepage says overages stay off unless you turn them on. Typical asks are priced in ranges: 2 to 250 credits for an everyday task, 250 to 1,000 for deep work such as triaging a day of support tickets, and more for large builds.
If the pool runs low, Lindy pauses and tells you. That caps spend, but it is not an approval gate: a paused agent does not ask permission, it stops. Watch for scheduled routines silently missing a run. The Free plan gives $50 in credits that last 7 days, which is enough to test write actions on dummy records before connecting live data.
Method and limits
2026-11-22
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.