Cross industry

Every AI win we found was invisible to the customer

Six operators, seven industries, one sorted list: every AI deployment that worked stayed hidden from the customer. Every failure went straight to them.

Every AI deployment that worked across Artificial Reality's first season was invisible to the customer, and every one that failed was visible to them. Host Galina Fendikevich sorted six operators across seven industries into two lists in the Season 1 recap, and nothing crossed over: translation running inside a broker's inbox, a forty-five-page lease read in three minutes, forecasting no guest ever sees.

Her summary of it:

"AI is winning exactly where nobody can see it and losing exactly where everybody can."

That is not a rhetorical flourish. It is a sorted list.

What is the actual dividing line?

Not customer-facing versus internal. Plenty of internal AI fails quietly and plenty of customer-facing AI works. The line is who absorbs the error.

In every deployment that worked, a competent human sat between the AI's output and any consequence. The broker reads the translated message before sending it. The communications lead edits the draft. The restaurant founder's attorney reviews the lease findings. The hotel's revenue team acts on a forecast it can sanity-check against occupancy it already knows.

In every failure, the output reached the customer directly. The concierge answered the guest. The chat assistant answered the charter client. The valuation went to the business owners, who believed it.

The rule that falls out: deploy AI where a knowledgeable person will see the output before anyone else does. Wait where they will not.

Lease review before the lawyer

A restaurant founder handed a forty-five-page commercial lease to ChatGPT with the instruction to act like a commercial real estate agent and find everything that could hurt her. Three minutes later: eight red flags and a drafted letter to the landlord. Her attorney reviewed it in ten minutes for free instead of billing $1,000.

The structure is exact. The AI did the volume, the licensed human did the last mile. Fendikevich treats that as a repeatable play, and it is the general form of every success in the season.

Context before drafting

Sam Polstein keeps client history, voice and writing samples as a library of markdown files, read by an AI assistant before any writing begins. In his words, by the time the prompt has started, the output is already 90% of the way there. He still edits. The AI never speaks to a client.

Translation inside the inbox

Jennifer Kerum runs translation inside her own email, so she can hold client conversations in Chinese or Spanish in messages she composes and sends. The guest never queries the model directly and experiences only that they were understood. Her comparison: ten years ago you would ask Google Translate for a coffee and it would offer you a cat. The capability improved enormously, and the deployment works because the guest never talks to the model.

Forecasting nobody outside the building sees

Sofiane Ghorbel, who has spent twenty-five years in luxury hospitality, lists where his industry's AI actually works: revenue management, demand forecasting, marketing optimisation and guest sentiment analysis. All back-office. No customer ever sees a forecast.

What the four wins have in common

A competent reviewer sees the output before anyone else does.
That reviewer's expertise matches the kind of error the tool makes.
The AI does the volume work and the judgment stays with the human.
None of them is announced to the customer as AI, because the customer never touches it.

AI tools named in this report

Tool	Named by	Verdict	Used for
ChatGPT	A restaurant founder, Season 1	Worked	Reading a 45-page commercial lease before her attorney did
Google Translate	Jennifer Kerum, yacht charter broker	Named as the old benchmark, not as current use	Her comparison for how far translation has come
No product named	Sam Polstein, PR and deep tech communications	Worked	An AI assistant reading a markdown client-context library before drafting
No product named	Sofiane Ghorbel, luxury hospitality	Worked	Revenue management, demand forecasting, marketing optimisation, guest sentiment analysis

Tools named by operators on the record. Inclusion is reporting, not endorsement.

What to do first

Spend on the back end now and wait on the front end, then revisit when you can put a competent reviewer between the model and your customer. That is a sequencing instruction, not a prohibition.

Start with the lease-review play, because it is the cheapest of the four and needs no procurement: give the document to the model, ask it to act like the professional who would normally read it, and hand the findings to the professional you were going to pay anyway. Three minutes of AI work, ten minutes of review, $1,000 of billing avoided.