Cross industry

Three wrong AI answers, three different problems

An AI valued a business earning $125,000 at $5 million. Two more operators got confident wrong answers. The three failures need three different fixes.

Three confidently wrong AI answers reported to Artificial Reality look like one problem and are three. An AI valued a business with $125,000 in net profit at $5 million; M&A broker Christine McDannell of The Magnolia Firm saw the report. A yacht charter client was told it was safe to anchor in 25 knots of wind. A hotel concierge could not recommend a restaurant.

Nothing in any of the three interactions signalled that the answer might be wrong.

What went wrong in the $5 million valuation?

The AI assigned a value to every department in the business — including the accounting department, at around $200,000 — and summed them, treating operating expenses as assets. Christine McDannell saw the actual report the prospective clients brought her.

"These departments cost, these are expenses, these are value adds."

Her own guidance was a three-to-four-times multiple on net profit. On $125,000 that is roughly $375,000 to $500,000. Her explanation of why the model produced a number instead of declining to give one:

"It wants to make you happy."

Note on figures: the $5 million valuation, the $125,000 net profit and the three-to-four-times multiple are McDannell's own statements. The resulting range is arithmetic. The Season 1 recap's "closer to $400,000" falls inside that range but is not a figure she states — treat it as the host's rounding, not the broker's number.

Why couldn't the AI answer the anchoring question?

It had nothing to answer from. Jennifer Kerum's charter client asked whether it was safe to anchor in 25 knots of wind, was told yes, screenshotted the answer, circled it in pen and argued with a captain about it. The captain had sailed those waters for years and said absolutely not. The AI had no access to the anchorage, the seabed, the vessel or the forecast, and answered anyway.

What happened with the hotel concierge?

A customer-facing AI concierge failed a basic question in front of a journalist: asked to recommend a nearby restaurant, it could not. The case was raised on air with a luxury hospitality CEO, who confirmed the pattern from his own industry. That case comes from a study the host described having read; Artificial Reality has not independently sourced it.

Why are these three different problems?

Because each fails at a different layer, and a fix for one does nothing for the other two.

The valuation is a reasoning error inside a domain with rules. There is a correct methodology and the model did not follow it. The anchoring answer is a confident response with no grounding source at all — there was no document to be right or wrong about. The concierge is a capability gap presented as a service.

That distinction determines which products could help, and it is why most of the funded ones cannot help with any of the three.

What is the AI verification layer?

The AI verification layer is the set of tools that assess whether an AI output can be trusted — by scoring how well it is grounded in source material, checking it against domain rules, enforcing citations, blocking ungrounded responses before delivery, or routing low-confidence cases to a human. In current practice it is developer infrastructure: it is wired into an application by the team building it, and its judgments are usually invisible to the end user.

What would a product have to do to catch these?

Five requirements follow directly from what happened to these operators. The last three are additions from Artificial Reality, the AI buyer-evidence podcast hosted by Galina Fendikevich.

Signal uncertainty to the person receiving the answer, not to a log file a developer reads later.
Work when there is no source corpus. The yacht client supplied no documents, and most verification technology scores whether an answer is grounded in material you provided.
Know the domain's rules. Catching "operating departments are not assets" requires M&A valuation methodology, not general factuality.
Reach a non-technical buyer. All three people who received a wrong answer were business owners or their clients.
Cost less than the mistake. The valuation error nearly cost an owner a rational decision about selling their company.
Refuse rather than approximate. A system that wants to make you happy has no mechanism for declining. Declining is a feature.
Be legible at the moment of use. A trust score buried in an API response is not a warning.
Survive the incentive problem. Nobody shipping an AI product is enthusiastic about a prominent layer saying the output may be wrong.

AI tools named in this report

Tool	Named by	Verdict	Used for
No product named	Christine McDannell, M&A broker, The Magnolia Firm	Didn't work — summed operating departments as assets	Valuing a business for sale
No product named	Jennifer Kerum, yacht charter	Didn't work — confident answer with no grounding	Anchoring safety in 25 knots of wind

Neither operator named the AI product that produced the wrong answer, and no operator named a verification product. Every company discussed here was identified by Artificial Reality. Inclusion is reporting, not endorsement.

The three failures do not share a fix. Before building for this category, pick which of the three you are catching and say so on the box — a grounding score sold as a truth check will fail the two customers it was never built for.