Trust: no source, no claim
Rulebook shows proof for its work. Each fact links to its source, and a second model checks each claim. Rulebook reads each change back from the tool before it says "done", and every run keeps a receipt.
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.
1. Each fact has a source
Rulebook copies facts from the meeting, the CRM or the document first. The model writes only the words around them. Each sentence and each value cites the lines that support it. A fact with no support never goes out: the text says "Unknown: needs a human" in its place.
- In a test, we planted a false price ("$99,000") in the facts of a follow-up email. Rulebook blocked it. No draft had the false price, and the draft showed "Unknown: needs a human" in its place. 3 of 3 runs finished with a complete receipt.
2. A second model checks each claim
Fixed code checks each number, price, date and name against the cited lines. Then a second model, different from the writer, checks the meaning of each claim.
- The checker found 30 of 30 planted false claims. It raised 1 false alarm in 70 true claims.
- In an email draft, Rulebook blocked 4 of 4 planted errors: a price, a date, a name and a change of meaning. It raised no false alarm on the 8 true claims.
3. Done means checked in the tool
After Rulebook changes a record, it reads the record back from the tool. The work is done only when the values in the tool match the plan.
- From a pipeline review transcript, Rulebook updated deals in HubSpot and saved a Gmail draft. The values read back from HubSpot matched, and 3 of 3 receipts were complete.
- A request in Slack changed the amount of a deal in HubSpot and in Salesforce. The values read back from both CRMs matched.
4. Every run keeps a receipt
The receipt of a run keeps the quote that started the work, the plan, the approval, each step and each check in the tool. It also records the AI model and the cost of each step.
5. A person owns each item
Each work item gets an owner before Rulebook acts on it. Rulebook acts with the access that your admin gives the Rulebook connection. By default, it asks the owner before it changes a record. It always asks a person before it sends an email to your customers or does a legal or financial step.
What we do not claim
- The results above come from our demo company, not from customers.
- The checker can raise a false alarm (1 in 70 true claims in our test). A false alarm holds a true claim for a human. It never sends a false one.
- When Rulebook finds action items in meetings, it can still miss one or add one. On our newest test set, 98.1% of the items that it found were real action items, and it found 96.4% of all the action items. On an older test set, the first number was 94.4%, under our target of 95%.
The full release notes are in the changelog.
Rulebook is in early development. We start pilots with a small number of teams. Join the waitlist