What AI Actually Does in MLM Software, and What It Cannot

Almost every direct selling platform now advertises AI. Most of what is being sold under that label is ordinary software with a new name on it.

There are places where it genuinely helps, and they are narrower and less exciting than the marketing suggests. Here is the line, so you can read a vendor demo and tell which side of it you are being shown.

Start with what is not AI

Commission calculation is not AI, and it should never be.

Working out who earns what from a genealogy tree is arithmetic against a set of rules. It has to be exactly right, it has to be identical every time you run it, and it has to be explainable line by line to a distributor who disputes their payout and to a regulator who asks how it works.

Those requirements rule out anything probabilistic. If a vendor describes AI-powered commission calculation, ask whether the same inputs always produce the same output. The answer has to be yes, which means it is rules, which means it is not AI. That is the correct design and they should be proud of it rather than dressing it up.

Where it genuinely helps: 1. Finding abuse in the network

This is the strongest real use case and it is a graph problem.

Fake accounts created to trigger a rank bonus. Rings of accounts that only ever buy from each other. Sign-ups clustered on one device or one payment card. Positions stacked to game a binary leg. Individually each looks plausible. As a pattern across the whole network they are visible, and a person scrolling through a genealogy tree will never see them.

This is worth real money, because the abuse is being paid out in commissions and because tolerating it is a regulatory problem as well as a financial one.

2. Compliance monitoring of what distributors publish

The one most operators underestimate, and the one we would prioritise first.

Your distributors post income claims and product claims across social media, and in most jurisdictions the company carries responsibility for what its representatives say. Nobody can read all of it manually. Language models are genuinely good at flagging text that looks like an earnings claim or a health claim for a human to review.

The important design detail: it flags, a person decides. Never let it act alone. The value is coverage, turning an impossible manual review into a manageable queue, not replacing the judgement at the end of it.

3. Predicting which distributors are about to go inactive

This is ordinary machine learning and it works, provided you have enough history.

The signals that precede somebody quitting are usually visible weeks ahead: ordering rhythm changes, logins stop, their own recruits stop moving. A model trained on people who already left can spot the pattern in people who have not yet.

The caution is what you do with it. A prediction is only useful if there is a real intervention behind it, and if that intervention is a mass automated email, you will get nothing. The value is in directing limited human attention to the right people.

4. Support that answers the same forty questions

Distributor support is dominated by a small set of repeated questions. How is this bonus calculated. Why is my rank different this month. When does the payout run. Where is my order.

A system that answers from your actual policy documents and your actual data handles a large share of that. Two rules make the difference between useful and dangerous: it must answer from your documents rather than from general knowledge, and anything touching money must route to a person.

What it cannot do

How to test a vendor's claim in one question

Ask: what happens when it is wrong, and who finds out?

A vendor with a real system answers immediately, because they have had to design for it. There is a confidence threshold, a review queue, somebody who checks. A vendor whose AI is a marketing layer will treat the question as unexpected, and that hesitation is your answer.

What we would build first

If you run a direct selling business and want one thing built, we would argue for compliance monitoring ahead of anything customer-facing. It reduces a risk you already carry, it does not touch the payout logic, and it is straightforward to evaluate: you can measure whether it catches what your team catches.