AI assistants over your business data
An AI assistant on your ERP or books is only useful if its numbers match the reports. The way to get that is to keep the model away from the numbers: it picks which fixed look-up answers the question, and code does the maths. Here is how to tell whether an assistant works that way, a free check for made-up numbers, the questions to ask it and the tests to run on any vendor.
Pujan Motiwala · Checked against the assistants' own code on

Tools
- Check an AI answer for made-up numbers
Paste what the assistant said and the report it should have used. Every number that isn't in the report is marked. Nothing leaves your browser.
- Questions to ask an AI assistant
82 questions by industry and role, each with what should answer it and what to check it against, including the ones a good assistant refuses. Free under CC BY 4.0, also as JSON.
- How to test an AI assistant before you buy it
12 tests for any vendor's demo, with pass and fail for each and a test sheet for Excel.
Why assistants get numbers wrong
Most assistants on business data do one of two things: the model writes a database query, or it reads rows and summarises them. Both let the model produce the figure, and sooner or later it produces one that is nearly right: a total that misses a credit note, a share worked out from the wrong base, a figure rounded to look tidy. In a business system one wrong number is enough for nobody to trust the assistant again.
How an assistant should answer
Keep the model away from the figures. A small model only chooses which question is being asked; code runs a fixed, read-only look-up with the same maths as the screens; a written sentence with any number that isn't in the result is dropped; and what the data can't answer gets “the data doesn't have that”. The four steps, for buyers, are in how an AI assistant answers from your records; the engineering note is an assistant that only answers from your own records.
By business
- Manufacturing
Ember AI, at a boiler manufacturer: 12 of its 21 questions answered today.
- Trading and distribution
Ohm AI, at LS Technologies, an electronic component trader: 24 of its 24 questions answered today.
- Project businesses (interiors, fit-out)
Atelier AI, at Savistar & Saag, an interior design studio and furniture workshop: 37 of its 37 questions answered today.
Guides
- How an AI assistant answers without inventing numbers
The design that keeps figures right has four parts: a small AI model only chooses which question is being asked; code runs a fixed, read-only look-up with the same maths as the app's screens; any written sentence with a number that isn't in the result is dropped; and questions the data can't answer get a plain "the data doesn't have that". The model never produces a figure.
- What an AI assistant sends out, under India's DPDP Act
Ask what is sent to the AI provider with each question. It is one of three things: the question only, the question with the rows of the answer, or your records themselves. Ahrom Labs' assistants send the question only by default. Under India's DPDP Act the business stays responsible for safeguards, breach notice and erasure, including at its AI provider; the Rules putting those duties into effect start in May 2027.
Paste an assistant's answer and the report it should come from, and see every number that isn't in the report.
Ask Atelier AI two questions on the interiors demo · how Ahrom Labs builds assistants
Official sources
- Digital Personal Data Protection Act, 2023, section 8. Ministry of Law and Justice, Gazette of India, October 25, 2026.
What this page is not
A guide written by Ahrom Labs from the assistants it runs. What each assistant answers is stated as it stood on the date checked; a vendor's assistant should be judged by the tests, not by this page. Not legal advice on data protection. Last updated .
Talk to us about your systems.
Tell us what's held together with workarounds right now. We reply to every message ourselves.