When an agency screens for fraud on a client's behalf, who decides and who is told

When an agency screens, someone else's money is at stake and someone else's reputation is on the line, including the creator's. That changes how the work should be set up, and none of it is a product question.

No ranking appears here because the decisions that matter happen before any product is chosen. The deciding factor is a client's tolerance, not a feature.

Who decides, and who is told

The decision belongs to a named person at the agency who owns that client relationship — the account lead or talent lead on the account, not the analyst who ran the check and not the tool. It has to be someone with the authority to decline a booking, because declining is the only consequence a finding can really carry.

The client is told in a written decision, sent to the marketing lead who owns the campaign: what was checked, what was seen, what it cannot establish, what is recommended, and who decided. The creator is told in a written notice: what was looked at, that a check is not a claim about them personally, what happens next, and the response window in which they can reply. Where the creator is represented, that notice goes to the manager, not the creator.

What to take away

Why no ranking is offered

The agency version of this purchase depends on three client-specific factors. What their contracts allow you to do about a finding, what data they let you access, and how they weigh a wrong accusation against a missed case vary per client.

A ranked list cannot see any of them.

There is also nothing stable to rank against. No platform publishes confirmed truth, so vendors calibrate against their own history, and the comparison depends entirely on your creators and categories.

The platforms quoted for this work — audience-quality screening such as HypeAuditor, and the wider influencer suites such as CreatorIQ — answer a question next to the client's rather than the client's own question, and what they return is a number, which is exactly why none of them belongs in the client conversation.

Assaying is the closest historical analogy to this work, and the useful part is what an assay office actually produced. Not a verdict about a person, but a stated measurement of a stated sample, on a stated date, by a named party.

Every good screening record has the same four parts: a stated sample, a stated measurement of that sample, a stated date, and a named party who decided. In practice they live in two documents. The written finding that goes to the client carries all four. The written notice that goes to the creator — or to the creator's manager — carries what was looked at, what it does not establish, and how to reply.

Settle the consequence with the client first

Detection is only useful if a finding leads somewhere, and where it leads is a commercial matter that belongs in your agreement with the client and in theirs with the creator.

Question	Who answers it	What goes wrong if it is left open

Four questions decide how a finding ends, and each of them lives in a clause rather than in a product.

Without those clauses, a detector's output is your opinion against theirs.

Where the booking runs through an intermediary, several of these clauses already sit in the platform's terms rather than in yours, and the guide to creator marketplaces sets out which ones and where the rest are quietly left with you.

What never goes to a client

A score, as a headline figure. Once a client has seen one, they will ask for it, and shortly after that the number becomes the deliverable and the judgment behind it stops being read. That is Goodhart's law working exactly as advertised, and it is much easier to prevent than to reverse.

What goes to a client instead is a decision and its basis. Here is what was checked, here is what was not, here is what we saw, here is what we recommend, here is who decided. That format survives scrutiny, survives a change of tool, and does not create an expectation you cannot meet next quarter.

What you tell a creator

Write this once, in advance, in plain language, and keep it. It needs to say what was looked at, that the check is not a claim about them personally, what happens next, and how they can respond.

Two rules make it survivable. Never send an accusation you cannot explain in components, because the reply will ask for them. And never let a machine appear to be the author of the decision; a named person deciding, having looked, is defensible in a way that an automated flag is not.

Where the creator is represented, the conversation runs through a manager rather than the person, which changes the tone and the balance of it, and the relationship it sits inside is described in the guide to talent representation.

Access, and what happens when a client says no

Many products degrade quietly when they cannot get the data they were designed for. They fall back to public-data estimation and keep presenting numbers in the same interface, so one screen shows two very different kinds of claim.

Ask any candidate directly what it does in that case, and ask to see both versions of the same profile. Then decide, per client, what you will do when access is refused: proceed with weaker evidence and say so, or decline to screen at all. Both are defensible. Proceeding with weaker evidence and not saying so is not.

Thin access also produces false alarms. A surprising share of what arrives labeled as fraud turns out to be a reporting problem in the other direction, which belongs with the guide to measurement layers.

There is also a records question here that agencies underestimate. Observations about identifiable people need a retention period and a reason. In the United States, what makes an undisclosed paid partnership a legal problem rather than a scoring problem is the FTC's Endorsement Guides, 16 CFR Part 255, and Section 5 of the FTC Act on deceptive acts and practices. On the data side, guidance from the Information Commissioner's Office — its data-protection guidance on the security principle and on storage limitation — is worth reading in the current version before you build an archive you would have to defend.

The controls that come before any purchase

For most agencies, the cheapest protection is commercial rather than technical; exhaust it before considering a subscription. Use unique links and codes per partner, pay against validated outcomes, and hold back payment until validation completes.

Request analytics from the creator's own account for a stated window, and keep them with the deal. None needs a vendor, and all survive a change of tool.

What a product genuinely adds beyond those, and what a score can and cannot establish, is in the pillar guide to detection and audience quality.

Common questions

A client insists on a standard we cannot verify. What do we do?

Say so in writing, and agree what you will do instead. Accepting a standard you cannot check means taking on a liability you cannot manage.

Should the same threshold apply across our clients?

Only if their tolerance for a missed case and a wrong accusation is identical, which it will not be. Set it per client and record the reasoning.

Do we tell creators that screening happens?

Yes, before they agree to anything, and say what is looked at. It is fairer, it deters the least serious cases, and it makes the difficult conversation possible.

Who at the agency should own this?

The account lead who owns the client relationship — someone senior enough to decline a booking, with the check itself done by whoever will have to explain it. Ownership by a person who cannot say no produces findings that change nothing.