The cost of an AI hallucination in a quote: a B2B contract worth PLN 200,000 lost

I costed what a single AI hallucination in a quote was worth - PLN 200,000 plus a referral channel that froze for three months. The contract was lost in Q2 2026, and with it part of a reputation the company had built for years on recommendations.

I know this case from a conversation with the owner of a services company who drives to meetings himself and guards client relationships personally. His salesperson received an enquiry from a referred client - defined scope, defined budget, a response deadline under a week. End-of-quarter target pressure was real, so the salesperson reached for the AI tool the team had been using for months.

How the pricing error is born

The model produced a price table in minutes. The discount for a referred client was to be "attractive but within margin" - an ambiguous instruction, because the salesperson himself did not know what threshold he could offer without asking the owner. The model filled the gap with a hard number - 23 percent instead of the permitted 12 - and added an SLA clause guaranteeing a four-hour response, which the company never offered because it has no 24/7 support team. The document looked like a professional offer, so it went to the client unchanged.

The mechanism is simple once you take it apart. A language model works like a very advanced autocomplete - it predicts the next fragment from statistical patterns, so when data is missing it fills the gap with content that fits the pattern. It guesses with full confidence, and that confidence is the most treacherous part.

Mount Sinai research from August 2025 showed two things at once. Leading models hallucinate in 53 to 82 percent of cases, and users lose the ability to verify critically because the tool presents content in an orderly, convincing form. MIT Sloan Management Review stresses that hallucination is a property of how models predict text, so attempts to "switch errors off" are doomed from the start.

In quoting, hallucinations take specific shapes:

• A miscalculated discount based on an ambiguous instruction, distorting margin across the whole table.
• Invented product features or support tiers that do not exist in the portfolio.
• Fictional contract clauses or wrong readings of regulations.
• Invented case studies and references that sound credible but never happened.

Why nobody catches it

A salesperson under target pressure sees a tidy table, consistent language and a finished-looking document. Verification feels like lost time, especially when every hour of delay increases the chance the client picks a competitor.

"I drive to referral meetings myself, so I have no time to actually run the company" - I have heard that sentence in different forms from a dozen SME owners in recent months. When the founder does not review every quote, small inconsistencies stay invisible until the client catches them.

Here the client caught it on first reading. He asked whether the four-hour SLA really was standard. The salesperson had no good answer, the owner learned about it a day later, and the PLN 200,000 contract was gone. The referral channel from that circle froze for a quarter.

A principle from IBM training materials decades ago still holds: a computer can never be held accountable, therefore it must never make a management decision.

What it costs beyond the 200,000

Legal responsibility sits with the company - regulators treat misleading a client the same whether the error came from a human or a model. I break the cost into items, because only then is the full scale visible:

• Direct lost revenue - PLN 200,000 from a contract already on the table.
• The cost of acquiring a new client in the referral channel, which is built over months, not per click.
• Lower funnel efficiency - the client tells the story around, freezing further recommendations.
• Potential administrative penalties and regulatory risk tied to communication and domain reputation.
• The cost of rebuilding domain reputation if the faulty communication went out at scale and spam rate climbed.
• The cost of rebuilding trust in the team - sales churn costs more than one contract.

In cost-per-completed-task terms, verification raises the cost of a single task by a few percent and removes a PLN 200,000 loss from the balance sheet. In TCO logic it is a safeguard, not an expense.

A checkpoint instead of a better prompt

Better prompting treats the symptom, because the model will still guess when data is missing. Zero-Hallucination architecture assumes verification is part of the process. In practice that means:

• A verification checklist before every quote goes out, covering numbers, clauses and scope.
• Automatic comparison of figures with the current price list, flagging anything outside approved ranges.
• Sign-off on key clauses by whoever owns legal compliance.
• A send block until the checkpoint is closed.

The honest other side: AI still cuts quote preparation time by 60-70 percent, confirmed by my own measurements from August 2026, so the checkpoint has to be light enough not to kill that gain. That is why you design it as a gate, not a second full-time job - in AGAPE Octopus verification takes four minutes per document on average.

An analogy from real life: a checkpoint works like B30 concrete in a structure. You cannot see it in the quote and the client never asks about the concrete grade at handover, yet without it the structure collapses under the first serious load.

An example from last month: the system stopped a quote where the model had calculated an 18 percent discount while the approved range topped out at 10. The cost of that stop was four minutes. The alternative cost was potentially another PLN 200,000 and another frozen referral channel.

How many documents go out unverified

AI risk in sales will not disappear, because models will keep guessing with full confidence. Hallucination is managed systemically - exactly like credit or operational risk. Losing a contract to AI is not a conference scenario: it already happened, in a company like yours, in the second quarter of 2026.