Compliance · 6 min read
Therapist Instagram Automation: Default to Not Replying
On this page
- The rule that comes before every other rule
- Why a keyword block list is not enough here
- The crisis signpost must not depend on the automation
- Never let a template imply somebody is reading
- Public replies, and the thing they reveal
- Takeover and the queue
- What is honestly automatable, and what is not
- The honest ending
Every other post in this series is about which questions an automation should answer. This one starts from the opposite end and assumes the answer is none of them, because a therapist's inbox is the one place on Instagram where the person messaging may be in real trouble at the moment they type.
The rule that comes before every other rule
Never automate a reply to anyone expressing distress, self-harm or crisis. Not a comforting one. Not a signposting one. Not a carefully worded one written by a clinician and approved by a supervisor.
The job of the automation, when such a message arrives, is to get out of the way as fast as possible and put the conversation in front of a human being. That is not a design trade-off. It is the requirement the rest of the build is arranged around.
An automated reply to someone in crisis does not read as efficiency. It reads as: I sent this to a place where nobody was.
Why a keyword block list is not enough here
Every other post in this series tells you to write a good negative keyword list. Here that advice is necessary and insufficient, and it matters that you know why before relying on it.
Distress does not arrive with the vocabulary a filter expects. It arrives as four lowercase words with no punctuation. It arrives as "hi" at 2:40am, followed by nothing, followed by three paragraphs at 3:15. It arrives in Hinglish, in metaphor, as a reply to a story about something else. A blocklist catches the messages that name the thing, and those are not the only ones.
So the posture inverts. Instead of an automation that replies to everything except a list of exceptions, build one that replies to nothing except a very short list of situations you have decided are safe.
- Scope every automation to specific posts. Not the whole account. One psychoeducation post, one workshop announcement, chosen deliberately.
- Make the trigger narrow and explicit. A single word the person types on purpose because you asked them to in the caption — not a broad any-of list that will match a sentence somebody wrote in pain.
- Keep the reply to a resource and a route. A worksheet, a page on your site, the practice's contact route. Nothing that responds to what they said.
- Layer the block list on top anyway. It is the second net, not the first.
- Default everything else to silence. A message that matches nothing gets no automated reply and a person reads it. That is the correct outcome, not a gap in coverage.
The crisis signpost must not depend on the automation
A therapy practice needs a visible, current route for someone in crisis, and it has to work when the automation is off, misconfigured, out of window, or blocked by a check. So it lives where nothing has to fire: the bio, a pinned post, the website. It says plainly what someone should do if they need help now and cannot wait for a reply here.
Publish your own. Crisis lines are national and they change, and a stale number is worse than no number — this post will not print any. Whoever runs the practice picks the services relevant to the country and population it actually serves, writes them down, and sets a calendar reminder to check them twice a year.
Never let a template imply somebody is reading
The most tempting automation on a therapy account is the friendliest one: "Thanks for reaching out, we'll get back to you soon."
At 3am, to someone who has just written the hardest message they have written all year, that is a claim about attention that is not true. It says a person received this. Nobody did, and because it landed instantly it reads as the most attentive response in the thread.
If an automated acknowledgement exists at all, it must be honest about what it is: a machine, no promise about when someone will read, and the crisis route stated. Most practices, having written that out honestly, decide they would rather send nothing. That is a reasonable conclusion.
Public replies, and the thing they reveal
A comment automation posts publicly. Under a post about anxiety, a reply to somebody's real name — however gentle — associates that person with the subject of the post in front of everyone who follows them. Nothing belongs in the public reply field beyond something completely generic, and it is worth asking whether even that needs to exist.
Takeover and the queue
The moment anyone from the practice replies by hand, automation stands down in that thread permanently. That is the boundary between a supervised communication and an unsupervised one, and it should happen early and often.
The review queue holds anything the system is not confident about for a person to approve or reject. Treat that as the default state, not the exception.
What is honestly automatable, and what is not
Safe, because it is administrative and identical for everyone: how to enquire, in person or online, which languages practitioners work in, where the practice is, and where to find a resource you have published.
Not safe, and not close: anything responding to what somebody described, anything reassuring, anything interpreting, anything about medication, and anything stating whether a person should seek help urgently. That is clinical work, and a keyword match is not doing it.
One more that catches people: "currently taking new clients" is a fact with an expiry date. It goes in a temporary automation you switch off, never an evergreen template.
The honest ending
Templated replies are free and unlimited and a credit is spent only when the AI writes a new reply, which on this build is close to never. Free tier is 100 credits with no card, packs from ₹499. Cost is not the deciding factor here and pretending otherwise would be the wrong note.
There is no calendar integration, so nothing can confirm a session. No CRM sync — leads leave as CSV. No WhatsApp. It cannot message anyone who has not messaged first. Official Graph API only: seven days on a comment, twenty-four hours inside a DM thread, restarted only by their message. That window is a platform constraint, not a model of care, and it should never set the pace at which a practice responds to a person.
And it is entirely legitimate to run no automation at all. If nobody has the time to read Activity weekly and re-read the templates quarterly, the correct configuration is off. What the checks do and the clinical line elsewhere in healthcare are both worth reading first.
