---
title: "You're (probably) shipping wrong AI numbers (MTN #62)"
---

74% of business users have. Here's why it keeps happening, and the check that catches it. ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­    ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­  

| View in browser You're (probably) shipping wrong AI numbers.   I don't say that to be provocative.   In June we surveyed more than 100 business users for a report called Using AI You Don't Trust.   74% had shipped a decision, report, or shared output based on a gen AI number that turned out to be wrong. 39% had done it in the last 30 days alone.   If you use AI for analysis you hand to leadership or clients, this is happening in your reports right now, whether or not you've caught it yet.   Most people blame the prompt when a number comes back wrong. The prompt is just the only part of the exchange you can see. The answer is made from what sits underneath it, the data, the definitions ,and context the model is working from: The data it was given. A pasted CSV is one tool's view at one moment. The model cannot know what is missing, or whether it matched your other systems. How the metric is computed. What counts as pipeline, and whether it should be summed or averaged. The model guesses your definition, and it can guess differently the next time you ask. The business context. Your target, the fact that it changed in July, the two deals that are test records. Nothing in the paste says what "good" means. The analysis itself. A trend, a comparison, or an anomaly needs to be computed. A model working from a paste estimates them on the fly. You can hand the model a perfect prompt, and if those four things aren’t in place, you’ll still get a wrong number.   It will fill whatever's missing with plausible defaults, and plausible defaults are where that 74% comes from.   Without Databox, the honest fix is to supply all four yourself:   1. Export from every source and reconcile. 2. Write the metric definitions down and paste them in. 3. State the target and the exclusions. 4. Specify how to compare the periods. 5. Then do it again next time, because the model does not remember, and a colleague asking the same question gets a different answer.   Most people in our survey skipped that work.   37% verified an AI number by asking a second AI, which inherits the same gaps. Only 23% went back to the source and recomputed, and in our planted-error test those re-derivers caught nearly three times as many real errors.   We built the new Databox so all four are already in place before you type anything. Your sources stay connected and reconciled. Metric definitions and calculations are set once and used every time. Goals, targets, and what changed live alongside the data. Trend analysis, anomaly detection, and forecasting are built in and run on your governed metrics. Genie, the AI Analyst, works from that, so the question is the only part you type, and every answer shows the source it came from. It's included in every plan, so if you'd rather try it than watch it, connect one source and ask it a question. If you'd rather see it before you try it, we're showing it live. On September 29 at 11am ET, Emil, who leads customer success here, is showing the new Databox LIVE, including how our AI Analyst answers a question with every source shown and turns the finding into a report you could hand to leadership. One hour, free, no account needed. Save your seat Can't make 11am ET? Register anyway and the recording arrives within one business day.   - Pete P.S. You'll notice this issue looks different. Move the Needle is now focused on how to actually use AI in your marketing, sales, and reporting work. One problem per issue, with a fix you can put to use right away. If there's an AI problem you're stuck on, hit reply and tell me, and we'll try to solve it in a future issue. Read more Full report: Using AI You Don't Trust: How Business Users Actually Run Analytics in 2026  Databox, Inc. 6 Liberty Square, PMB #471, Boston, MA 02109, USA   You received this email because you signed up for Databox or subscribed to receive our content. Update your email preferences to choose the types of emails your receive. Unsubscribe from all future emails. |
| --- |