Skip to main content
Question

AI / Omni hallucinations -> Useless for data analysis

  • March 14, 2026
  • 4 replies
  • 61 views

Omni doesn’t seem ready to be a real, live feature that promises to do “Quick data analysis and pattern recognition across records” as even the most basic capability that Airtable describes (per https://support.airtable.com/v1/docs/using-omni-ai-in-airtable#faqs).

Background: I have built an Airtable base for volunteer shift tracking for a special event, with a multiple volunteer areas, and multiple shifts per area. Each volunteer has a record, and there is a field where after the event, it was logged if they showed up or not.  I am now asking Omni AI to analyze how many volunteers showed up, per volunteer area.  

Issues/Process: First, Omni completely hallucinates data and presented a summary table of the show up rates -- but upon probing deeper, it arbitrarily picked volunteers and included them as part of areas they were not in.  Next, I uncover that the analysis did not actually even use the boolean/checkmark field that describes whether a volunteer showed up or not. After those two issues, I find out Omni was ignoring some shifts for no reason. After teeth-pulling I eventually got Omni to a point where it addressed these issues, and acknowledged “Thank you for pointing out the discrepancy… a complete and accurate analysis requires a thorough aggregation of all relevant records.” (i.e. it did not have this requirement beforehand!). 

So, then I ask for it to run it’s analysis again, but there is a hard stop (built-in by Airtable?) where it will always create a summary or partial “demonstration” and just gets stuck in a loop asking for confirmation of a full analysis, even after I explicitly ask for the complete analysis. This approach might be great for initial checking for issues, like what came up, so that is appreciated, but is also is cutting off getting the full analysis with “showing your work” so I can spot further hallucinations. 

Conclusion/question: Am I missing something about the expectations or limitations of Omni, or Airtable’s intents / roadmap?  It appears Airtable has shifted to focus on being an “app-building platform” now (vs online database service,  years ago when I started using Airtable for this) -- but with such an emphasis on AI capabilities in and on this platform, these basic data analysis hallucinations and walls seem very problematic. 

Addendum: 
After more teeth-pulling, I think Omni is willing to give and show me the full analysis now, across multiple messages where I have to confirm for each volunteer area.  I would hope to think that Omni would be doing this data analysis anyway, if it were truly doing a “complete and accurate” analysis to eventually show me an accurate summary table based on my data. But perhaps somehow splitting such analysis across multiple messages and needing multiple prompts to ask to get that analysis is something that Airtable has deemed better, e.g. to track / account for the computational effort to actually achieve accurate “data analysis” results? 

tl;dr

Omni AI hallucinates basic data analysis results, even though Airtable says the feature is useful for “Quick data analysis and pattern recognition across records“ and more. 

4 replies

omf
Forum|alt.badge.img+5
  • Participating Frequently
  • March 16, 2026

Omni, like most LLMs, does have its limitations. It’s good for getting a general overview or summarizing items, but for getting exact numbers and percentages, it’s not 100% accurate. From the horse’s mouth (chatGPT): “think of LLMs like a very advanced autocorrect… [Omni] is still powered by generative AI models that produce answers by predicting likely text, not by “proving” each statement the way a calculator, database query, or audited report would.” Again, this can be useful in some contexts, but not all.

If the fields you’re analyzing aren’t text based, I’d suggest using the Dashboard interface page. You might be able to achieve the result you’re looking for with a pivot table, depending how the table(s) are set up.


mserkerr
  • New Participant
  • March 16, 2026

I’ve had similar issues where I’ll ask it to do basic data manipulation - some days it works great and all of the records are updated, other days it tells me it can’t do it. I’ve learned you have to be incredibly specific and narrow in your requests and take it one step at a time. 


nroshak
Forum|alt.badge.img+9
  • Inspiring
  • March 16, 2026

In general, when prompting LLMs, it’s helpful to get incredibly specific and detailed with the prompt, and give an example of what you would like it to do. Something like this:

I want you to look at the data in the Volunteers table and give me a summary of how many volunteers attended the Fun Fair in each area. For example, if I ask you how many volunteers showed up in each area at the “Fun Fair” event, you would first find the rows where the “Fun Fair” checkbox field is checked, and then you would look at the “Area” field for each one of those rows where the “Fun Fair” field is checked. If there were 13 rows checked under “Fun Fair”, and 11 of them had “Hot Dog” in the “Area” field, and 2 of them had “Dunk” in the “Area” field , then you would respond “13 volunteers showed up for the Fun Fair. 11 of them worked Hot Dog and 2 of them worked Dunk.” Now, look at the Volunteers table and give me a summary of how many volunteers showed up to the “Car Wash” event (that is, where field “Car Wash” is checked) in each area. 

But to be totally honest I wouldn’t use Omni for this task. I would create a view on the Volunteer table, with a Filter condition on attendance for your event, then group by volunteer area. That’s likely to be faster and easier than making Omni do this right.

-Natalka


VikasVimal
Forum|alt.badge.img+12
  • Inspiring
  • March 19, 2026

You’re using AI, and AI is just as good as its context. 
And quite likely, all the records across all its tables are too much for AI’s token window to process.
And it seems like it is not there yet to create its own transient formulas and lookups to answer your questions. The best way I think would be to create relevant lookups/rollups and point it to the field for analysis.

Here’s a note I wrote on Context that AI needs: https://blog.opstwo.com/ai-in-smbs-needs-three-things-context-context-and-context/