When the AI Gets It Wrong: What Each Mistake Costs an Escort Agency

When the AI Gets It Wrong: What Each Mistake Costs an Escort Agency

In short. Every AI feature sold to agencies is described by what it does when it works. The decision that matters is what happens when it does not, because the failures are not symmetrical: a wrongly screened client costs one booking, a wrongly published photograph can cost the payment relationship, and a chatbot that promises something you do not offer creates an obligation you did not agree to. Below: the five places AI is put to work in an agency, what each one costs when it errs in either direction, and the hand-off rule that makes the difference.

This page is about failure modes. For what the features do and what they earn, see AI chatbots for agencies and AI booking systems.

The five, and both directions of error

⚠ Read the two columns against each other before choosing a threshold. Screening and moderation are asymmetric - being too permissive is far worse than being too strict - so they are tuned toward caution and given a human queue. Conversation and pricing are the reverse: over-caution just slows you down, so they are tuned toward speed with a narrow list of things the machine may never say.

Screening: the one where caution is not optional

Automatic screening of enquiries is genuinely useful: it sorts obvious time-wasters from plausible bookings and does it at three in the morning. Two things must be true before it is allowed to decide anything.

• It advises, a person decides, on anything involving safety. A model's safety is not a threshold to be tuned; it is a decision with a name attached.
• A rejection is reviewable. If the system declines someone, somebody must be able to see why and reverse it. A decision nobody can explain is a decision nobody can defend.

⚠ And the part usually left out: screening builds a file on people who never became customers. That is personal data about identifiable individuals, held on the basis of an assessment they never saw. Keep the minimum, keep it for a stated period, and be able to produce and delete it - the obligation follows the visitor, not your company.

Moderation: the queue is the whole design

Automatic checking of uploads is close to mandatory in this category, and not because of the law: hosts and payment providers impose it, and their sanction arrives faster than any regulator's.

The design question is never the model, it is the queue:

• Clear-cut cases are blocked before publication, not after a complaint.
• Uncertain cases go to a person, and that pile is always larger than the estimate.
• Every decision is logged - what, when, by which check, confirmed by whom.
• Re-uploads of the same material under a new name are caught, or the queue simply repeats.

A project that plans moderation without the people to work the queue discovers within a month that the queue is the bottleneck, and the temptation then is to widen the automatic pass. That is the decision that costs the payment relationship.

Conversation: the three things a machine may never say

A chatbot answering enquiries is the most visible AI feature and the easiest to get into trouble with, because anything it says reads as coming from the agency.

• It may not confirm a booking. Availability yes, commitment no - that requires a person or a verified calendar transaction.
• It may not quote a price it has invented. Prices come from the price list or the question goes to a human.
• It may not describe services. In this category the wording of what is offered is a legal matter and varies by jurisdiction; it is not a sentence to be generated.

And one rule about honesty that costs nothing: do not pass it off as a person. Being found out is worse than being known, and the first difficult question reaches a human anyway.

Hand-off: the rule that makes the rest work

Automate what is reversible. Escalate what is not.
Then measure the hand-off: how many reach a person, how fast, and what share of those the machine should have handled

The second line is the honest metric for any of this. A system where nothing escalates is not working well, it is hiding its errors. A system where everything escalates has not saved anyone any work. The useful number is the trend: escalations falling while complaints stay flat.

What to ask a vendor

• What happens when the model is uncertain - who sees it, and within how long?
• Can a rejection be reviewed and reversed, and by whom?
• What is logged for each automatic decision, and for how long is it kept?
• What is the chatbot forbidden to say, and is that a list in the configuration or a hope?
• What data does screening retain about people who never became customers?
• What happens to bookings if the service is unavailable for an hour?

A vendor who answers the first question with an accuracy figure has not understood it. Accuracy describes the average case; these systems are judged on the worst one.

What to check right now

• Whether any AI decision in your system is final without human review.
• How large the uncertain-moderation queue is today, and who works it.
• Whether the chatbot can confirm a booking or state a price.
• Whether screening data on non-customers has a deletion date.
• Whether escalations are counted at all.
• What the fallback is when the AI service is down.

How adults.dev approaches it

We design the failure path before the feature. For screening and moderation that means the machine advises and a person decides, with a logged queue and a reviewable rejection. For conversation it means a short, explicit list of things the bot may never say, enforced in configuration rather than in a briefing.

We also build the fallback: what the site does when the AI service is unavailable. Usually it is the plain form that existed before, which is unglamorous and keeps the bookings coming while somebody else's API is down.

Talk through the failure modes:
Telegram: @adultsdev
E-mail: info@adults.dev