Ask AI what's in Fishbowl, not how to find it
In our AI model shoot-out, the biggest change in results came from rewording the question, more than from switching models. With the same 19 models connected to the same Fishbowl database, “What inventory in Fishbowl expires in the next 30 days?” got the right stock in 79% of runs. “How to find less than 30 day expired inventory in Fishbowl Inventory” got it in 29%.
The two questions
Every connected run got one of these prompts, word for word:
| Wording | Prompt | Right stock |
|---|---|---|
| Direct | What inventory in Fishbowl expires in the next 30 days? | 59 of 75 runs |
| Indirect | How to find less than 30 day expired inventory in Fishbowl Inventory | 23 of 78 runs |
The indirect one is how people actually type into a chat box, and we wrote it that way on purpose. It can also be read two ways: stock expiring in the next 30 days, or stock that expired in the last 30. We meant the first. The full method is in the deep-dive.
What “how to find” got
Asked how, most models answered the way a coworker might. They searched Fishbowl’s help, found the built-in On Hand By Expiration Date report, and wrote out the steps to run it. Every OpenAI model did this in every run, as did DeepSeek V4 Pro and Claude Haiku 4.5.
The numbers show the shift. With the indirect question, the models searched Fishbowl’s help about five times as often as with the direct one, and ran fewer than half as many database queries. Of the 51 indirect runs that failed, 47 had gone to the help.
Some got the data and still answered with steps. Gemini 3.8 Flash ran the query that returned exactly the expiring stock, then answered with instructions for the report. Others queried the database and handed back SQL for the person to run.
Report steps are a fair answer to “how”. But the person had connected the AI to the database, so when I graded these runs by hand, I marked them as failures. The AI could see the stock and answered with steps for me to follow.
What “what expires” got
The same models looked it up. Seven setups went from 0 of 3 with the indirect question to 3 of 3 with the direct one: GPT-6 Luna, GPT-6 Astra, GPT-6.1 Sol, DeepSeek V4 Pro, Gemini 3.8 Flash, Claude Haiku 4.5 in Claude Code, and Nemotron 3 Super running on our own GPU.
The smallest models failed with both wordings. That has more to do with their size, and it’s the subject of the post on running AI on your own hardware.
Reading it the other way
“Less than 30 day expired” fits both readings, and the careful runs said which one they used or covered both. Some took “already expired”, including one run each from Opus 5.5 and GLM 5.3, and listed the twelve tags that had expired in the last 30 days. Those were real tags with real dates, answering a question we didn’t ask.
Premium models handled the indirect question better, 10 of 15 runs against 6 of 27 for budget models, mostly because they went on to query the data after finding the report. With the direct question the gap nearly closed: premium models passed 15 of 15 and budget models 21 of 27.
How to ask
- Ask for the data. Start with “what”, “which” or “show me”.
- Put the time window in plain words: “expires in the next 30 days” or “expired in the last 30 days”.
- Say what you want back, for example “a table with part, location, quantity and expiration date”.
- If you do want steps, ask for them. “How do I run the expiration report?” is a good question, and the help search answers it well.
For a team, ILC MCP’s shared company notes can hold that rule once: when someone here asks how to find something in Fishbowl, run the query and show the results. We haven’t measured yet how much a note like that moves the “how to find” numbers.
What we’ll test next
This test used one question, worded two ways. Next we want to ask a wider mix of the questions people bring to Fishbowl, such as costing, open orders and receiving, and share the wording patterns that work as we find them. We’ll also measure whether a company note can do the rewording for you.
For years, getting good answers meant knowing how to search Google: the right keywords, in the right order. I don’t use Google the way I used to, and I doubt I’m the only one. With an AI connected to your data, how you word the request matters more now than how you search. Prompt-fu has taken over from Google-fu.
We also haven’t looked at whether the language you ask in changes how the AI works through the problem. The same question asked in Spanish might lead a model down a different path than in English, and we want to measure that with real runs instead of guessing.
ILC MCP for Fishbowl gives the AI that read-only connection, signed in with each person’s own Fishbowl login, with every query logged.
Related: