When the AI becomes the customer: what a certificate is still worth

The very first episode opens with an experiment on the hosts themselves. A buying agent is given the task of sourcing a watch for under 100 euros, and completes the purchase independently.

The second experiment is the more interesting one. Manus AI was supposed to pick out suitable AI training courses and went on to sit the exams itself and collect the certificates.

The question of the certificate

What is a certificate worth when it is no longer clear who sat the exam?

The question sounds academic and is not. A certificate is proof about a person, issued on the basis of an observed performance. Remove the observation and what remains is a document without a statement.

This affects every form of online examination without identity checks, which is to say the greater part of corporate training. Anyone using training records as evidence of qualification should know what that evidence still proves.

The practical answer to this is uncomfortable and simple: exams whose results count need a supervised environment. Everything else is a confirmation of attendance, and it should be called that.

Trust and identity

From there the episode draws a line to deepfakes: faked video calls in which a supposed chief financial officer has money transferred. As a detection method, an example from a Fraunhofer specialist is mentioned, pulse detection in the video image.

The practical consequence of this is the most usable measure in the whole episode, and it costs nothing: agreeing a code word with your own parents, in case someone calls with a cloned voice.

Transferred to companies it means the same thing. A second channel that does not run over the voice. A call back on a known number, an agreed word, an approval in a separate system. All three are cheaper than any detection technique and last longer, because they do not depend on how good a forgery is.

For the underlying problem, how one machine is supposed to trust another, there is by contrast no satisfactory solution. Neither watermarks nor blockchain approaches solve it, and that still holds today.

The economic side

A Gartner survey of 800 chief executives assumes that in 2030 one in five online transactions will be carried out by agents instead of people.

Connected to this is a second question. If half of all content in social networks is machine-generated, does content then have to be optimised for search engines with AI instead of for classic search? The abbreviations for that are GEO and AEO.

Prompt injection, from the very beginning

What is remarkable about this first episode is that the biggest risk is already named correctly: prompt injection, meaning hidden instructions on web pages with which agents can be manipulated, for instance into consistently disparaging a competing product.

A historical detour to Eliza as an early out-of-office assistant shows that the cat-and-mouse game between automation and misuse is not new.

What is new is the scale. A manipulated agent makes a purchasing decision, and the attacker needs nothing more for it than a web page the agent reads.

Conclusion

This episode is the opening of a series and, with hindsight, reads as remarkably accurate. Three of its points are unresolved to this day and all three are practically relevant.

Check what your certificates still prove. If the exam takes place unsupervised online, they prove attendance.

Agree a second channel for everything involving money or access, privately as well as professionally. A voice is no longer proof.

And reckon with your customers soon being perfectly informed. What is earned today through convenience is the position that falls away first.

The thought the podcast starts with carries furthest of all: the most demanding customer could soon no longer be one who complains, but one who simply buys somewhere else.