The Breeze AI Assistant Production Playbook

Why this playbook exists

AI assistants inside HubSpot are easy to demo and hard to ship into daily executive use. The Lagan Aviation Executive Pipeline AI Assistant turned 3 to 5 hours of board reporting prep into a 30 second query. This is the methodology INSIDEA uses to get there for any customer running on HubSpot Breeze.

Phase 1: Find the right job

A general-purpose AI assistant is hard to trust and slow to adopt. A scoped assistant with one clear job gets used daily. The first decision is which job. INSIDEA's intake covers three questions.

• What recurring report or query is currently the most expensive to produce, measured in senior hours per week
• Who needs the output, how often, and in what format do they want the answer
• What data does answering it require, and is that data already in HubSpot or in a connected system the assistant can read

The right first assistant is usually executive pipeline reporting, customer health summaries, or weekly deal triage. Sales rep enablement and support escalation are strong second picks. Anything that runs for the team weekly or quarterly and costs senior hours is a candidate.

Phase 2: Data readiness

Breeze rewards good data and punishes bad data. The first project on any Breeze rollout is CRM hygiene, not AI configuration. INSIDEA audits the property model, the pipeline data, the activity logging, and the contact ownership before touching the assistant. A messy CRM produces messy AI output every time.

The minimum data the assistant needs

• Complete pipeline stage history on every active deal (no stages skipped, no stale closed-lost without a reason)
• Consistent activity logging on meetings, emails, and calls
• Clean property model on contacts and companies with type-discipline (no free text for anything we want to segment)
• ICP segment and lifecycle stage populated on every active record
• Owner assignment on every contact, company, and deal

Phase 3: Configure the assistant

Configuration is the visible part. It is also the smallest part of the engagement. INSIDEA configures the assistant's scope (which data it sees), the output format (a board-style summary, a structured list, a single number, a markdown table), and the access permissions (who can ask, who can act on what it returns).

For the Lagan build, the configuration produced a board-style pipeline summary in under 30 seconds, with key figures pulled from the four business-unit pipelines. The output format was negotiated with the directors before configuration started, so the first thing they saw matched their expectations.

Phase 4: Evaluation harness

AI assistants need an evaluation harness or they drift silently. INSIDEA builds a small golden set of questions with known-good answers. We run the assistant against the golden set on launch and on a regular cadence after. Drift gets caught early, not after a wrong number appears in a board meeting.

Phase 5: Adoption

Adoption is the difference between a working assistant and a shelf-ware demo. The Lagan rollout had day one adoption because the directors were trained the same week the assistant went live, the output matched what they actually wanted, and the assistant lived inside HubSpot where they already worked.

• Train the leadership users individually, not in a group session. Two people in one screen-share is fine. Twelve people on Zoom is a guarantee of weak adoption.
• Pre-populate the first three sessions so the experience is fast from the first try, not slow while they figure out the right question.
• Have INSIDEA on call for the first two weeks for ad hoc questions and quick tuning.
• Measure usage at week two, week four, and week eight. If usage drops, the assistant needs tuning, not retraining.