AI Read His Customers’ Praise as Complaints and Cut His Loan

Nabil had asked for $60,000. Two reach-in freezers needed replacement, and the equipment estimate came to $41,800 before delivery. He wanted the balance for inventory and enough cash to cover the weeks when installation would disrupt the store.

The first offer arrived through the lender’s dashboard. He printed it and set the page beside the equipment estimate. The loan amount was $28,500. The fixed fee was $6,270, bringing the total repayment to $34,770.

The page did not show a credit score. It listed broad factors used to make the offer, including uneven deposits and an elevated level of customer dissatisfaction online. That second factor stopped him. The shop had an average rating of 4.

8 stars across 214 public reviews.

Nabil had owned the neighborhood grocery for nine years. Customers posted in English, Spanish, Arabic and Vietnamese, depending on which language came more easily after carrying milk or rice to the register. Many of the reviews were brief. Some thanked him for special orders.

Others praised employees for carrying bags outside or keeping an item behind the counter until payday.

The stars said they were pleased. The profile said otherwise.

How praise became a risk signal

The lender did not employ people to read hundreds of reviews for every applicant. It bought an automated business profile built from public and financial data. According to the general description Nabil received, the profile analyzed customer comments for sentiment and topics, then converted those findings into signals that could be used alongside revenue and payment history.

For a review written outside English, the system first had to identify the language or translate the text. Another model then classified the words as positive, neutral or negative and looked for subjects such as service, price or product quality. Those classifications were combined across reviews. The lender saw the resulting profile, not the original conversation between a customer and the shop.

This matters because a five-star rating and a sentence are not always treated as one piece of evidence. A scoring system may keep the star average as one variable while turning the written text into a separate complaint rate. If the text classifier marks a five-star review as negative, both signals can remain in the record, even though they contradict each other.

Nabil could see the broad factor on the lender’s page, but he could not inspect the complaint rate, the individual labels or the weight assigned to them. He also could not tell whether the system had low confidence in a translation. The printed offer gave him a dollar result without the intermediate steps.

A financing consultant offered to challenge the profile for $1,250. The fee covered a review of the public comments, human translations of disputed passages and submission through the lender’s support process. There was no promise that the offer would change.

Nabil put the consultant’s estimate beneath the printed loan offer. The fee was almost one month of payroll for a part-time employee. It was also less than the difference between the equipment estimate and the amount he had been offered, though paying it did not guarantee access to that missing money.

The words the classifier missed

Before hiring the consultant, Nabil and his niece copied the non-English reviews into a spreadsheet. There were 31. They asked regular customers to check the meaning of passages that common translation tools rendered awkwardly.

One Arabic review used a polite expression that literally referred to tiring or troubling the shopkeeper. In context, the customer was thanking Nabil for taking extra time with a special order. The translated text retained words associated with trouble and fatigue, which a sentiment classifier could read as dissatisfaction.

A Spanish customer wrote that there was nothing to complain about after an employee replaced a damaged package. The sentence contained the word “complain,” even though the grammar negated it. Another review described an item as cheap and good. The likely English rendering emphasized “cheap,” a word that can indicate value in one sentence and poor quality in another.

These are common failure points for automated sentiment analysis. Translation can preserve dictionary meaning while losing the social purpose of a phrase. A classifier may also rely on words that often appear in complaints without handling negation or context well, particularly when the translated sentence is short. The system can still assign a category with enough confidence to count it.

The spreadsheet did not prove which reviews had entered Nabil’s score. It showed a pattern. Of the 31 non-English comments, the consultant identified 18 that could plausibly be categorized as negative after literal translation, despite strong star ratings and positive human readings.

Nabil kept returning to the printed offer. The $28,500 would replace one freezer and cover part of the second. The $6,270 fee remained due under the repayment schedule, leaving little room for the disruption he had planned around. Waiting carried its own cost because one freezer was already losing temperature overnight and using more electricity.

He paid the $1,250.

A larger offer, without a visible score

The consultant submitted human translations and pointed out the mismatch between the review text labels and the star ratings. Nabil also supplied recent bank records, but there had been no major change in sales. Deposits during the review period rose about 1.5 percent, not enough to explain a large change in available credit by themselves.

Six weeks later, the lender issued another offer. Nabil printed that page too.

The new amount was $48,000. Its fixed fee was $6,240, for total repayment of $54,240. The online customer feedback factor no longer appeared among the main reasons holding down the offer. No one showed him the old profile, the revised profile or a calculation tying a particular review to a particular dollar.

The difference was still concrete. The first offer carried a fee equal to 22 percent of the amount borrowed. The second fee was 13 percent. After adding the consultant’s charge, the larger loan gave Nabil enough to replace both failing units and restock the frozen-food cases, though he postponed the produce cooler that had been part of his original request.

He accepted the second offer. He did not regard the $1,250 as payment for translation alone. It bought access to a human process around a score that affected his work but was not available for him to examine.

The revised result also left him unsure about the reviews that remained online. Customers still wrote in the language they preferred, and he did not ask them to switch to English. His niece continued adding non-English comments to the spreadsheet, with a plain translation beside each one. Nabil kept the two printed offers in the same folder as the equipment invoice.

Questions people ask

Can customer reviews affect a small-business loan offer?

In Nabil’s case, the lender used an automated business profile that included online customer sentiment alongside financial records. The reviews did not replace revenue data, but their classified sentiment contributed to a negative factor on his dashboard and appeared to affect both the amount offered and the fee.

Why would an

AI system mark a positive review as negative?

Sentiment systems often translate text before classifying it. Literal translation can lose an idiom’s intended meaning, while words such as “complain,” “trouble” or “cheap” may trigger negative labels even when a customer is expressing praise. Short comments give the classifier little context for correcting that first reading.

Could the owner see how the score was calculated?

He could see broad factors associated with the offer, including online dissatisfaction, but not the underlying score, review labels or model weights. The human appeal focused on likely errors found in public comments rather than reproducing the lender’s calculation, which remained unavailable to him.

Did paying for human help change the result?

Six weeks after the consultant submitted human translations, the offer increased from $28,500 to $48,000 and carried a lower fee as a share of the loan. Nabil paid $1,250 for the review, accepted the revised offer and filed both printed pages with the $41,800 equipment invoice.