Every company sits on thousands of invoices, orders and customer records. Anyone with a simple question either digs by hand or uploads the company's numbers to a cloud. I thought both options were wrong. So I built Souprise.
But before I explain what it does, I need to talk about failure.
Why enterprise AI projects actually die
I've watched a lot of AI projects over the past years. Very few fail on technology. They fail on the day the tool invents a number for the first time and somebody notices. A language model, when in doubt, guesses an amount that looks right. The answer sounds completely confident. By the third incident like that, nobody asks the tool anymore. They go back to the colleague with the Excel sheet. The project survives another six months on a slide deck, then dies quietly.
I reproduced this effect in my own tests. The model quoted an invoice amount that existed in no record at all. Literally none.
And there's a second layer almost nobody talks about. The legal one. Who is liable when an employee issues a credit note or sends a payment reminder based on an invented figure? The insurance industry learned this lesson faster than most IT departments. Newer policies increasingly carry exclusions for damage arising from AI-generated content. In plain terms: the cost of a hallucinated number ends up with the company. A tool whose mistakes can't be insured is an operational risk.
And there's a third reason that has nothing to do with technology at all. The projects are rolled out the wrong way. Many companies believe AI works instantly and overlook that employees need to be trained and brought along first. A tool introduced over people's heads doesn't get used, no matter how good it is. Acceptance is decided during the rollout, not in the feature list. That deserves its own post, and I'll write one.
What Souprise does about it
Souprise is a free open-source tool. You ask in plain language. Which invoices are overdue. What do we know about this customer. The answer arrives in seconds, with sources attached, and everything runs on your own machine. No account, no cloud.
The decisive difference sits deeper, though. Souprise takes the numbers away from the language model.
In the default mode every value is copied verbatim from the record, never generated. Sums and averages are computed in code, not by the model. If the system doesn't know a customer, it refuses instead of answering about the closest lookalike. The language model is only allowed to phrase the sentence, and every figure in the finished text is checked against the records. If one deviates, the text is discarded.
That's not a claim, it's measured. Every acceptance bar was locked and committed before the tests ran. Value accuracy of 1.000 across sixty questions. Zero wrong values under conflicting data. Zero invented figures shipped in the language mode. The results that hurt are published too — fine-tuning measurably added nothing for factual lookups, and the report says so.
For the insurance question, this approach means something concrete. A system that demonstrably only copies values, always cites its sources and stays silent when unsure shifts the risk profile. An unquantifiable AI statement becomes an auditable database extract with a receipt.
What's still missing, in public
For enterprise use, three things are still missing. They're firmly next on the plan. Record-level access control, so sales sees different data than accounting. A complete audit trail that captures every question, every source and every answer in an append-only log. And encryption of the index files at rest. Anyone can watch the progress on GitHub — each of these has an open ticket with acceptance criteria that get locked before the work starts.
Companies have good reasons to be wary of AI. Souprise takes that seriously and answers with tooling you can check.
Trying it costs one afternoon and one CSV export.
Souprise doesn't sell answers. It sells the certainty that the answer is correct.