The CRM and Product Usage Data a Marketing Agent Needs Before Launch
The CRM fields and product usage events a marketing AI agent needs before launch, how to tie usage to CRM accounts, and pass marks for completeness and freshness.
On this page
The data requirements before deploying marketing AI agents on CRM and product usage data are narrower than a general AI readiness audit suggests. An agent needs the fields its workflow reads to be filled on the records it will touch, values fresh enough for the action it takes, and every product usage event tied to exactly one CRM account. Below are the fields and events per workflow, a method for connecting usage events to accounts, and pass marks you can adopt or argue with. We read the vendor pages cited in this piece on October 7, 2026, and we keep what they document separate from what we propose on top.
What do vendors document as prerequisites for their own agents?
The data prerequisites before deploying AI marketing agents, on the CRM side and the product usage side, are written down by vendors at the level of single fields and settings. Agents plan and act across tools (our primer on what agentic AI is covers the mechanics), so each one inherits whatever the records say.
The setup guide for the prospecting agent from HubSpot, last updated September 30, 2026, lists four AI settings switches that must be on, three of which grant data access: CRM data, Customer conversation data and Files data. The same page says the agent "considers recent engagements up to the past year that are associated to contact records" and names form submissions, page views, calls, meetings, notes and email opens. By that wording, engagement that was never associated with a contact record sits outside what the agent considers. The play settings add quieter dependencies: sending from the contact owner needs an owner with a connected inbox or a fallback sender, and send times fall back to a default time zone when the contact's time zone is unavailable.
A broader guide to adopting AI features from HubSpot, last updated September 16, 2026, puts the principle in one sentence: "If your data is incomplete or inconsistent, AI outputs may be inaccurate or irrelevant." It recommends property validation rules, marking important properties as required, merging duplicate records, and logging emails and recording calls.
The considerations page for Agentforce Engagement from Salesforce, which notes that the Agentforce Lead Nurturing name has changed to Agentforce Engagement, is stricter: "The Owner field on leads, person accounts, and contacts must be filled before assigning a record to the Engagement agent." For contacts, the owner field on the related account has to be filled as well. The page adds that with Header-only Email Capture turned on the agent can't use the email subject or body to understand prospect intent, and that Send as Seller requires record owners to have email addresses connected to both Inbox and Einstein Activity Capture.
None of these pages puts a number on any of it. We found no fill-rate, duplicate-rate or freshness threshold in the vendor pages we read, which is why the pass marks further down are labeled as our assumptions.
The checklist, workflow by workflow
A data readiness checklist for AI marketing agents, covering CRM and product usage, only works when it is written per workflow, because a prospecting agent and an expansion agent read different fields. The table is our proposal, built on the documented prerequisites above. Field names vary by CRM.
| Workflow | CRM fields the agent reads | Product usage events | Checks that matter most |
|---|---|---|---|
| Lead follow-up and prospecting | Contact or lead: email, owner, consent status, lifecycle stage, job title, time zone. Account: domain, owner, industry. Logged emails, calls and meetings. | None required | Completeness of the fields that gate a send; duplicates on email and domain |
| Trial and product-qualified lead nurture | The row above, plus plan and trial start date on the account | Account created, activation event, first use of a key feature, teammate invited, last active date | Usage-to-account mapping; freshness |
| Expansion and upsell | Account: owner, plan, seats purchased, renewal date. Deal: stage, amount, line items. | Active users against seats purchased, usage against plan limit | Plan and seat fields that agree with billing; freshness |
| Retention and churn risk | Account: owner, renewal date, open tickets. Contact: role, such as admin. | Usage frequency against an expected frequency, last active date | Freshness; an agreed expected frequency per event |
Consent status sits in the first row on purpose, and our first-party data playbook covers how to collect and store it.
Keep the usage event list short. The customer success workspace from HubSpot has a product usage tracking feature whose documentation is tagged Beta and listed for Service Hub Professional and Enterprise. It defines activation as "a one-time event that indicates a customer has started using a product or feature" and states: "Each account is limited to 10 activation and/or usage events tracked per portal." In our view an agent is better served by a few events with agreed definitions than by a raw clickstream.
How do you tie product usage events to CRM accounts?
A data readiness checklist for AI agents that spans CRM and product usage integration turns on one question: can each usage event be traced to exactly one CRM account? Four steps get you to a yes.
- Pick the account key from your own database. The Group spec for Segment, published by Twilio, defines the Group call as "how you associate an individual user with a group, such as a company, organization, account, project, or team" and says "A Group ID is the unique identifier which you recognize a group by in your own database." Use the workspace or tenant ID your product already issues, and store it in a unique field on the CRM account.
- Attach the key to every user and every event. The same spec cautions that "not all platforms support multiple groups for a single user." Decide in advance which account wins when one person belongs to two workspaces.
- Land each event on a record with the right timestamp. The custom events guide from HubSpot checks identifiers in priority order: a custom matching ID property, the record ID (marked recommended), then the contact email address and the contact usertoken, both for contacts only. It also says: "By default, HubSpot sets the event occurrence timestamp to the time the request was sent." A nightly batch or a backfill therefore needs
occurredAton every occurrence, or last week's logins will read as activity from the day they were sent. - Sync account attributes the other way and know the lag. The Salesforce V2 source documented by Amplitude imports Salesforce fields as group properties and needs a Group Name Field that "should identify a row in the table uniquely." By default the update process "runs once a day" and an hourly schedule is offered. We read that as up to a day of lag on a plan or owner change under the default.
Then count three things across the last 30 days of events: events with no account key, events whose key matches no CRM account, and events whose key matches more than one. The third number is your duplicate-account problem showing up in usage data. Matching anonymous visitors and devices to known people is identity resolution, a separate job that this checklist leaves aside.
Pass marks we use, stated as assumptions
A data readiness checklist for AI agents that work across marketing and CRM integration needs pass marks, and since the vendor pages gave us none, the figures below are our working assumptions. Tighten them where a wrong action is expensive.
| Check | How to measure it | Pass mark (our assumption) |
|---|---|---|
| Send-gating completeness | Share of in-scope records with email, owner and consent status all filled | 95% |
| Personalization completeness | Share with job title, industry and plan filled, and a fallback written for blanks | 80% |
| Duplicate accounts | Accounts sharing a domain or account key, as a share of in-scope accounts | Under 2% |
| Usage-to-account mapping | Share of the last 30 days of events matching exactly one CRM account | 98% |
| Freshness | Age of the newest usage value, and of the last plan or owner sync, when the agent acts | 24 hours or less |
Freshness carries more weight than its place in the table suggests. The Salesforce page says that once a prospect is assigned, the Engagement agent begins working the prospect immediately and schedules the first email to be sent within an hour, so a wrong field has to be caught before assignment.
A worked example (hypothetical)
A hypothetical B2B software company wants a trial-nurture agent on 4,000 trial accounts. Owner is filled on 3,720 of them, or 93%, which misses the 95% mark. The last 30 days hold 310,000 usage events: 297,600 match exactly one CRM account (96%), 9,300 carry no account key (3%) and 3,100 match two CRM accounts (1%). The newest usage values are 6 hours old when the agent acts, which passes.
Two checks fail. The 9,300 keyless events point to a part of the product that sends events without the account key, and the 3,100 double matches point to duplicate accounts waiting to be merged. Launch waits on the mapping fix, since a message built on another account's usage does more damage than a late one. The owner gap is easier to handle: scope the agent to the 3,720 owned accounts and repair the assignment rule in parallel.
Approval points, failure modes and what to measure
Both vendors document a review mode, and a new agent should start in it. The HubSpot prospecting agent offers "Review before sending" alongside "Send automatically" as its two Autonomous mode options. With Require Manual Approval on, the Salesforce page says "the record owner has 14 calendar days to approve each email" before the agent cancels it. Ask reviewers to tag each rejection with the field that caused it. Rejections per 100 drafts, broken out by field, is the most direct production measure of data readiness we know.
The failure modes are predictable:
- A stale plan field produces an upgrade pitch to an account that upgraded yesterday.
- A duplicate account gives one buyer two owners and two streams of outreach.
- Batched events sent without
occurredAtmake old activity read as a surge on the day of the batch. - A record with an empty Owner field fails the Salesforce requirement for assignment to the Engagement agent.
For measurement, record a manual baseline before launch: reply rate, meetings booked, qualified pipeline per 100 records worked, and cost per record. Track the same four after launch, plus data-caused rejections and the share of in-scope records excluded for failing a check. Our breakdown of what a marketing agent costs per contact and the AI ROI calculator are further resources for the cost side. Reading data is half the picture; what agents can write back to a CRM covers the other half, and the governance operating model covers who is accountable for each workflow. When a team wants help wiring the path from usage data to the CRM, that work sits inside our agentic AI automation service.
Readiness gets decided one workflow at a time. A CRM full of half-empty properties can still be ready for an agent that reads six fields, provided those six are filled, current and attached to the right account. The quickest way to find out is to list the fields for one workflow and count them before anyone turns the agent on.
Sources
- HubSpot, Set up and use the prospecting agent (last updated September 30, 2026)
- HubSpot, Set up a HubSpot account to adopt AI features (last updated September 16, 2026)
- HubSpot, Track activation and product usage in the customer success workspace (Beta, last updated September 16, 2026)
- HubSpot, Send custom event occurrences
- Salesforce, Considerations for Using Agentforce Engagement
- Twilio, Segment Spec: Group
- Amplitude, Salesforce Group source