Skip to content
Fonsent

A CRM value needs a source

Why a summary is not evidence, what a source-backed value looks like, and why the check belongs in the database.

Updated · 3 min read · Fonsent

In early access today. The AI model isn’t switched on yet. Field updates come from built-in rules, which pick up only sentences that state a value outright. What works today

A CRM fills up with values nobody can account for. The deal value moved from one number to another. The close date slipped a month. Someone, or something, decided that, and the record doesn’t say who or why.

That was tolerable when people typed every value. It stops being tolerable when software writes them, because software writes more of them, faster, and with the same confident formatting whether it is right or wrong.

A summary is not a source

The usual answer is a summary: a paragraph that says what the customer wants. A summary is a reading of the email, not the email. It can be a good reading. It can also drop the condition at the end of the sentence, or merge two people’s positions into one.

When a summary and its source disagree, the source is right. So the useful question about any value is not “does this sound plausible?” but “which sentence says so?”

What a source-backed value looks like

In Fonsent, a value written from an email, a meeting or a transcript is stored with the exact sentence it came from. Open the value and you see the source and its date, the quote in its surrounding text, who applied the change, and an Undo.

FonsentOpportunities
An opportunity record in Fonsent. The deal value, $96,000, is open to show its source: an email dated 12 November 2026 with the sentence that states the budget marked, who applied the change, and an Undo control.

The quote is short and specific on purpose. “Finance confirmed an approved budget of $96,000 for the first year” supports a deal value. A general sense that the customer has money does not.

Why the check belongs in the database

You could ask the model to always cite its source and hope. Fonsent doesn’t rely on that. The database compares the quote with the stored source, character for character, and refuses a value that has none. A value that can’t point to a sentence is not written, whatever the model intended.

Two more rules keep the evidence meaningful over time. A source can’t be edited once it is read in, so a quote can’t be made true after the fact. And evidence is only ever added, never changed or deleted.

What it changes day to day

  • Review gets faster. You read one sentence and decide, instead of re-reading a thread.
  • Disagreements end by reading. “Why does it say that?” has an answer one click from the field.
  • Corrections hold. A person’s own edit always wins over an older source.

Where it stops

A source-backed value is only as good as what was written down. If nobody stated the budget in an email or a meeting, there is nothing to quote and nothing is proposed. That is the intended behaviour, not a gap to paper over.

  • Fonsent doesn’t record meetings. It reads the transcripts and calendar files you import.
  • A source is linked to an opportunity only when the account has exactly one open opportunity. With several, it is linked to none of them and Fonsent says so.
  • Evidence from a private source is readable only by people who can see that source. Everyone else sees a count.

The idea is simple enough to hold a product to: every value the AI writes shows its source. See how it works in the product.

Start with your own records.

Create an account, import a CSV or a few emails, and read the first proposals with their quotes. An import is one change you can undo.

Early access. No payment details are asked for.