How to stop AI from inventing facts in customer emails
Do not let the model invent the facts. Copy each fact from its source first, and let the model write only the words. Check each claim with fixed code and with a second model, and hold any claim without a source for a person.
Why a draft goes wrong
A follow-up email after a call has a few facts that matter: a price, a date, a name, the next step. A model that writes the whole email can change one of them, and the email still reads well. The customer finds the mistake, not you.
How Rulebook writes a draft
- Facts first. Rulebook copies the facts from the meeting, the CRM or the document exactly. Each sentence and each value cites the lines that support it.
- Fixed checks. Code checks each number, price, date and name against the cited lines.
- A second model. A different model checks the meaning of each claim against its source.
- No source, no claim. A fact with no support never goes out. The draft says "Unknown: needs a human" in its place.
- A person approves. Rulebook does not send an email to your customers unless a person approves it.
What we tested
- The checker found 30 of 30 planted false claims. It raised 1 false alarm in 70 true claims.
- We planted a false price ("$99,000") in the facts of a follow-up email. Rulebook blocked it, and the draft showed "Unknown: needs a human" in its place.
These results come from Brightloop, our demo company with synthetic data, on real HubSpot, Salesforce, Slack and Google accounts. They are not results from customers.
The cost of the check
The second model adds about two cents to each email draft. We chose accuracy first. A false alarm holds a true claim for a person. It never sends a false one. More results are on the trust page.
Rulebook is in early development. We start pilots with a small number of teams. Join the waitlist