Training
Teach the business once. Reuse the same approved knowledge, voice and corrections across every AI Assistant that should know them.
Open Training in the current dashboard. A Training set is the shared source of truth for a business, brand or client. Assistants use it, but each assistant keeps its own account, job, actions, boundaries and activation state.
1. Create one source of truth
Choose New training set and name it after the business, brand or managed client. Keep separate sets when facts or voice genuinely differ; do not duplicate the same business merely because it has several channels.
2. Add trusted knowledge
Use Knowledge for approved facts and sources that assistants should share. Start with the information that repeatedly decides a customer conversation:
- Returns, shipping, operating hours, locations and contact rules.
- Qualification questions, pricing boundaries and escalation policy.
- Trusted files, knowledge entries and connected business sources.
- Corrections learned from reviewed conversations.
Live commerce, scheduling, CRM and support connections have different read/write capabilities. Check all data sources before assuming a source is readable or actionable.
3. Define the shared voice
Voice controls the recognizable way the business speaks: tone, formality, reply length, vocabulary, emoji policy, language behavior and forbidden patterns. Put brand-wide voice here; keep job-specific instructions—such as which fields a booking job must collect—inside the relevant AI Assistant.
4. Review learning
Treat corrections as product changes, not casual chat. Review the proposed learning, make sure it fixes the general rule instead of one isolated wording, and approve only facts you want reused. Your private DMtoLead Assistant can propose and test instruction improvements from a screenshot or field report before you apply them.
5. Assign Training and test the job
Open AI Assistants, choose the assistant job and attach the correct Training set. Test:
- A normal question the knowledge should answer.
- A live-data question that requires the connected source.
- A missing-information case where the assistant must clarify.
- A risky request that should escalate or refuse.
Activate only after the answer, source use and handoff match the policy for that specific job.