When the system isn't sure, it asks

Confident? It commits silently. Uncertain? One tap. Unclear? It writes nothing at all.

The problem

The worst thing an automated assistant can do is invent. You ask it to capture a call and it cheerfully writes down "client agreed to £50k" when what they actually said was "could be around fifty, let's see what the board makes of it." A hedge has been promoted to a fact. Now there's a load-bearing falsehood in your records, and the next time you act on it you look bad — or you've left a small bomb for future you.

The reverse failure is just as bad. The assistant is overly cautious, asks you to confirm every trivial extraction, and within a week you're rubber-stamping cards without reading them. The signal is gone.

A working capture system has to thread a narrow needle: write down the things it's sure about, ask only when it genuinely doesn't know, and stay silent when the source isn't clear enough to be worth anything.

What changes

Every fact the CRM extracts from one of your captures carries a confidence score it computed from the original quote. Three things happen:

Very confident — the fact lands silently in your records. The original quote and the score are both kept, so you can audit it later if you ever wonder where a claim came from.

Uncertain — the fact lands in a short review queue on your Review page. You see the proposed claim, the exact quote it came from, and three buttons: Approve, Reject, or Edit-and-Approve. Most decisions take under two seconds. You can optionally leave a comment explaining why — future-you reviewing the audit log will thank present-you.

Unclear — the system writes nothing. Hedging language ("maybe," "around," "let's see") that doesn't reach the uncertain threshold is dropped on purpose. Silence is better than noise.

Whatever you decide is durable. If you reject a misextraction, the same fact never asks again. The CRM doesn't flip-flop and doesn't pester.

Why we built it this way

Three pillars hold the trust contract together, and the review queue is where all three meet.

We cite evidence. Every claim, every score, every suggestion traces back to the original quote that produced it. You will never wonder where something came from; if you disagree, the source is one click away.

We let you override, durably. Your decision is permanent for that specific fact. Correcting the CRM once should be enough — non-stick corrections are the worst-possible failure mode for a tool you're supposed to trust.

We forget on purpose. Low-confidence interpretation is silence, not noise. The CRM does not invent, does not guess, and does not store character judgements, sentiment scores on close contacts, or anything else that would feel surveilling if the roles were reversed.

Together these mean you can trust what's in your records because you know exactly how it got there — and what's not in your records because the CRM had the discretion to stay quiet.

What to read next

→ Back to all guides — six short walkthroughs of how the CRM thinks between your visits.

→ See the CRM in your own week — thirty days free, then £8 a month.