Playbooks · 5 min read
How to Check an Instagram AI Claim Before You Buy It
On this page
- The four questions
- Claim: AI agents that go and find your ideal customers, then DM them
- Claim: voice-note replies and full vernacular handling
- Claim: AI that predicts which Reel will perform
- Claim: a dashboard showing your DM conversion rate
- Claim: one AI assistant handling every channel at once
- What is actually changing, and is dull
- The one-line version
Every few weeks a new capability gets described as though it already ships. Some of it does. Some of it is a demo that worked once, and some of it is a screenshot.
This post does not add a claim to the pile. It is a method for checking one, followed by five claims currently circulating, run through that method. Where the answer is that we do not do the thing, it says so, because a checking method that exempts the person holding it is not a method.
The four questions
Ask these of any capability, ours included. They are boring on purpose, and between them they separate a feature from a slide.
- Who acts, and as whom? Is this an authorised integration holding a token Meta issued, or software signing into your account with your password? That single answer sorts most of the category, and everything downstream depends on it.
- What permission does it hold? Not what it does — what Meta has granted. Reply to comments on your own posts. Reply inside a message thread. Those are grants. "Reaches your ideal customer" is not a grant.
- What does it refuse, and when? A system that cannot describe a send it will not make has not thought about the day the post goes wide. Ask for the list of refusals and the order they run in.
- What does it record? After a bad week you need to know what went to whom, and for the ones that did not go, why. If there is no answer screen, there is no answer.
A capability that cannot survive question two is not early. It is a different product with a password field.
Claim: AI agents that go and find your ideal customers, then DM them
This is the most common claim and the easiest to check. Question two ends it.
Meta does not issue a permission to message a person who has not messaged you. Not on a plan, not on request, not for enterprises. Whatever is doing the sending in that demo is signing in as a user and driving the app, which is a different risk profile that the buyer inherits and the vendor does not.
Claim: voice-note replies and full vernacular handling
Half real. Text in mixed Hindi and English is genuinely handled better by models than it was, and it matters here because a large share of Indian creators receive exactly that.
But check question three before you get excited about the language part. Matching and refusing are separate from writing. Our keyword matching folds case, punctuation and Latin accents — so Price!! and price are one word. It does not fold Indic vowel signs. दाम and दम stay two different words, and if your audience types both, both go on the list. No model improves that; it is a matching rule, not a writing problem.
Claim: AI that predicts which Reel will perform
We do not do this and have no view on whether anyone does it well. What we can say is what it would need: a large sample of your own content with outcomes attached, a stable algorithm to predict against, and a definition of "perform" that survives the week.
Two of those three are outside any vendor's control. Treat a prediction feature as an opinion generator and price it accordingly.
Claim: a dashboard showing your DM conversion rate
We have no analytics product. No conversion rate, no funnel chart, no per-template performance view, no attribution. When a comparison table says otherwise about us, the table is wrong.
The question worth asking whoever does show you that number is where the denominator came from and what event counted as a conversion. On the official API there is no tracked link and no purchase event coming back from Instagram, so a conversion rate has to be assembled from something. Ask what.
Claim: one AI assistant handling every channel at once
There is no WhatsApp here, no Messenger, no Telegram. Instagram only.
Where the claim is real elsewhere, question four is the one to press. Cross-channel usually means separate connections presented behind one screen, which is fine — but ask whether a conversation on one channel is actually joined to the same person on another, or whether they merely sit next to each other. The join is the hard part and it is the part the demo skips.
What is actually changing, and is dull
Three things worth planning around, none of which will be on a conference slide.
Generated text is becoming ordinary, so register is doing more work. When every account replies in the same smooth, slightly-too-complete sentence, sounding like that stops being a sign of competence. The replies that read as yours are the ones built from things you actually wrote.
Audiences type in more than one script and always did. The tools got better at replying in mixed language faster than the average keyword list got better at catching it. Widening the trigger list is still the cheapest improvement available to most accounts.
Metering shapes are diverging. Contacts, conversations, messages, generated replies — the unit differs and the same business gets a very different bill depending on which one a vendor chose. Ours meters generated replies: templates are free and unlimited, a credit moves only when the AI writes something new, free tier is 100 credits with no card, packs from ₹499. Comparing tools by what they meter and what they refuse is the comparison that survives a year.
The one-line version
Run the four questions. Most of what gets called a trend fails at question two, and the ones that pass are usually much narrower and much more useful than the headline suggested.
For the longer-horizon version of this — the constraints that are structural rather than fashionable — that post is here. For the claims the category makes about itself, including ours, the myths post starts with our own.
