Every fact needs a receipt

There is an obvious version of this product that we deliberately did not build.

The obvious version reads your email, asks a language model to summarise each deal, and shows you the summary. It demos beautifully. It is also unusable, and the reason is worth explaining.

Confident and wrong looks exactly like confident and right

A model asked to summarise a messy thread will produce a fluent paragraph either way. If it misread a forwarded quote as a new offer, the paragraph is just as smooth. Nothing in the output tells you which one you are looking at.

In most software a small error is a small problem. In a sales pipeline it is not. If the system says a deal is worth $24,000 and the real figure was $14,000 in a quoted reply the model misattributed, that number goes into a forecast, the forecast goes to a board, and nobody finds out until the quarter ends.

The rule we work to

Every fact stored links back to the exact message it came from, plus a fingerprint of that message's contents.

If it cannot be traced, it does not go in the database. Not stored with a low confidence score. Not stored with a warning. Not stored.

This sounds restrictive and it is, which is the point. It means the interface can always answer "how do you know that?" with the sentence, the sender and the date. And it means that when the system is wrong, you find out in one click instead of at the end of the quarter.

What it costs

Two things, and both are worth it.

The product knows less. Plenty of things are obvious to a human reading a thread that cannot be pinned to a specific sentence. We leave those out, or mark them clearly as a guess with a number attached.

And it has to say "I do not know" quite often. Every summary ends with what it could not see: the phone call it was not on, the corridor conversation, the WhatsApp thread. That is less impressive than a screen that appears to know everything, and considerably more useful.

Why it matters more than it sounds

The whole proposition is that you can stop maintaining the pipeline by hand. That only works if you trust the automatic one. And trust in software is not built by being right most of the time. It is built by being checkable, and by being honest about the edges.

A number you can click through to its proof is worth more than a smarter number you cannot.