- In Hrishi Olickel's examples, AI data pipelines produced wrong answers that matched row counts and dollar totals, or that several models endorsed. Stripped minus signs, unrecognised dates and mismatched spellings were the causes.
- His fix is to define how a job can fail before it starts, and to use checks from people who know the field.
AI data pipelines can fail in ways that standard checks do not see. Hrishi Olickel of Southbridge AI made that case on the AI Engineer show, in an episode about what cheap intelligence means for data. In one of his examples, row counts and dollar totals matched while most dates were empty. In another, several models endorsed a badly wrong figure. He said mistakes add up quickly when agents work this fast.
What was said
The examples came from a campaign finance project in Ohio. In the first, the pipeline looked for one spelling of the in-kind label. The file used a different one. The model therefore treated 15,658 in-kind contributions as ordinary money and never rechecked.
In the second, the pipeline tried four date formats, but the file used a fifth. As a result, 99.4% of transaction dates came out empty. A speaker on the show said "every row count and dollar matched, so the agent moved on."
In the third, several runs saw that loan amounts were negative. They judged the sign unneeded and kept only the size of each number. A speaker on the show described the result: "A net of $784,000 became $5.3 million of debt in the end, validated by multiple models."
He added that bigger models can be worse here. They bring their own beliefs to a task and defend them harder.
Why it matters
Our reading: a passing check proves only what it measures. In the date example, counts and dollars matched while nearly every date was missing. Checks that compare only counts and totals, or only the pipeline's own output, may miss a dropped sign or an empty field. Teams that buy or run AI pipelines should ask what each check can and cannot catch.
Olickel's practical advice was to define a failure surface before starting. That means setting a baseline to fail against, such as a data slice checked by hand. In the campaign finance case, he said any human who knew the field would catch the missing in-kind money. That is how the team found it.
A second model helped too. The team held back 200 groups of records as a test. Two models, Jev and Luna, gave the same answer on 194 of them. Those 194 held no errors. Every wrong answer fell in the six where the models split.
The other side
Olickel conceded the limits of both fixes. Human review is expensive and slow, and it interrupts the work. Model agreement helps less when models share a blind spot. In one benchmark, four model families landed on an identical total, yet the answer key handled signs differently from all of them.
He also said the hardest part of building a failure surface was putting it into words and tooling. The best current models, he said, are only barely good at it. The examples come from one team's projects. His view that this is the largest class of error across companies is his own observation. The talk gave no figure for how often it happens elsewhere.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI (2026-10-12)
