Running it
No dispute process protects either of you there
The commission saving is real, and it is not the trade being offered.
Guides on Running it: The admin that turns income into a trade
Sooner or later a buyer suggests dealing directly, usually framed as splitting the commission between you. The saving is real and the offer is still a bad one, because the commission is not a fee for introduction - it is the price of everything that happens when a job goes wrong.
Off the platform, both of you are unprotected, and you are the one holding the risk.
What the commission was actually buying
Strip out the discovery and the traffic, which the buyer's argument correctly notes you have already received, and four things remain.
A payment held by a third party rather than by a stranger's goodwill. A dispute process with a record it can read. A messaging thread that timestamps what was agreed. A public record of completed work that makes the next buyer say yes.
The fourth is the one earners undervalue. An off-platform job produces no review, no completed-order count and no ranking signal, so it is invisible to the mechanism that brings you the next five bookings. What a platform commission actually buys works through the rest of that ledger, and the summary is that the cut looks like a tax right up until you try to replace each line of it yourself.
The risk is asymmetric, and it points at you
Consider what each side loses in the two failure modes.
If the buyer does not pay, you have delivered work and have no process to appeal to. Chasing a small amount across borders is not economic, so the practical outcome is that you wrote it off.
If you do not deliver, the buyer has lost their money with no process either. That is genuinely bad for them, which is why cautious buyers do not make this suggestion.
The asymmetry is in who moves first. In this trade the earner almost always delivers before the money is irrevocably theirs, so the party carrying the exposure off-platform is you, every time. Direct arrangements also tend to run on the rails the FTC's scam guidance warns about: it tells consumers never to pay someone who insists on cryptocurrency, a wire transfer, a payment app or a gift card. Cautious buyers have read that advice, which is part of why they do not make this suggestion.
There is a second asymmetry worth stating plainly. The buyer risks one payment. You risk an account, and account loss is not proportionate to the size of the job that caused it.
The account consequence
Nearly every platform prohibits soliciting or accepting off-platform payment for work sourced through it, and detection is not sophisticated because it does not need to be.
The usual trigger is not surveillance, it is the buyer. Someone who feels shortchanged on an unprotected job goes to the platform anyway, and the message thread where the arrangement was proposed is right there. A second common trigger is the buyer mentioning it casually in a review or a support ticket, with no intent to harm anyone at all.
The outcome is typically suspension rather than a warning, because from the platform's side the offence is unambiguous. And suspension takes the reviews, the ranking and the payout balance with it - account risk and why not to bet everything on one platform covers the general exposure, which this behaviour concentrates rather than diversifies.
Weigh that against the arithmetic being proposed. Half a commission on one job is a small number, and even sustained over a year of repeat work from one buyer it does not approach the value of the record you would be putting up as collateral.
Saying no without losing the buyer
The answer is short and does not moralise.
"I keep everything on the platform - the protection runs both ways and it is what I am set up for. Happy to book the same work there whenever you want it."
Most buyers accept this immediately, because most are not testing you; they read a listed price, did the arithmetic, and asked. A buyer who pushes after a clear no has told you something useful, and it is worth remembering that the same person will be the one to complain to the platform later.
Do not delete the thread. If it ever becomes an issue, the message where you declined is the evidence that matters, and what evidence a dispute process actually weighs applies to conduct reviews as much as to money disputes.
The narrow legitimate case
There is one, and it is not the situation above.
A long-standing buyer who wants a different and larger piece of work, of a kind the platform does not list, is a genuine direct client rather than a diverted one. The distinguishing test is whether the platform sourced this particular engagement, and whether its terms treat past introductions as permanently covered - many do, for a stated period.
If it is legitimate, handle it as business rather than as a favour. A written scope, a deposit before you start, staged payment on anything substantial, and an invoice - contracts for larger jobs covers what that paperwork looks like when the amounts justify it. Read the terms first, because the answer is in them and reading afterwards is how honest people get suspended. For the platform-side version, Rate Cock's judges page is where the current rules for judge accounts are set out.
Two adjacent things explain why the pressure exists at all. Buyers have already had the free version of a rating from an automated tool before they ever pay a person, so they arrive with a sharp sense of what the human part costs and are minded to negotiate it - what an automated score actually delivers is the baseline they are comparing against. The same instinct shows up wherever precision is claimed without an intermediary standing behind it, which is why the measurement side documents its conditions so carefully. And from the commissioning side, the platform is mostly reassurance: the buyer's account of how a commission is placed reads like a list of the risks a buyer is trying to avoid, all of which they take back on themselves the moment they leave.
Which is the argument to make, if you want to make one. The protection is theirs too, and pointing that out ends the conversation more reliably than a policy citation.