One row per decision unit
Choose a stable appointment or prospect grain and resolve duplicates before training.
Is your CRM data ready for assignment modeling?
A credible assignment model needs appointment-level outcomes, stable agent identity, pre-assignment prospect context, and enough overlap to compare eligible agents fairly.
Data guideQuality beats field count
Ten reliable fields available at assignment time are more useful than hundreds of sparse or outcome-revealing fields.
In brief
Use this CRM data-readiness checklist to evaluate appointment history, outcomes, agent identity, leakage risk, missing values, and model feasibility.
The model can only learn the decision captured in the history
CRM exports often mix leads that never reached an appointment, appointments still being worked, contracts created after a win, duplicate contacts, and fields updated throughout the sales cycle. That shape must be normalized before modeling.
The core analytical row is usually one prospect or opportunity with a real appointment, an assigned sales agent, a mature won or lost outcome, and only the context that would have been known when assignment occurred.
Outcome maturity
Open appointments need a documented aging rule based on the actual time-to-win distribution.
Leakage control
Contract value, final status, post-appointment notes, and other future information cannot be predictive inputs.
Build the decision around usable evidence.
Minimum viable history
These fields establish the assignment decision and its result.
- Stable opportunity or prospect ID
- Lead and appointment dates
- Assigned agent ID and display name
- Won, lost, and open status
- Sale date when explicitly won
- Address or market for appointments
Useful enrichment and context
These fields can improve ranking when they are available before assignment and have sufficient coverage.
- Lead source and campaign
- Project or service category
- Property and location context
- Prospect name-derived categories
- Local demographic characteristics
- Appointment timing and seasonality
Data-readiness checklist
| Area | Ready looks like | Warning sign |
|---|---|---|
| Appointments | Real appointment date and address | Leads without appointments mixed into training |
| Outcomes | Explicit wins plus consistent mature losses | Large unexplained open backlog |
| Agents | Stable source IDs with readable names | Names reused as identifiers |
| Features | Known before assignment with useful coverage | Post-sale or mostly missing fields |
| Volume | Repeated outcomes across agents and contexts | Tiny isolated agent-segment cells |
| History | Enough recent and older data to assess drift | One short period or major undocumented process change |
Prepare the data without distorting it
Profile the raw export
Count records, nulls, duplicates, dates, statuses, agents, and source values before transformation.
Define the appointment cohort
Exclude leads that never became appointments and apply a documented outcome-maturity rule.
Build leakage controls
Classify every candidate field by when it becomes available and remove future information.
Evaluate feasibility
Measure coverage, class balance, agent overlap, segment stability, and enough volume for honest holdout testing.
See what your own appointment history supports.
Isotope Labs provides a complimentary CRM integration and historical evaluation before recommending a live rollout. You receive the evidence, limitations, operating requirements, and a clear next step.
Common questions
Do we need perfect CRM data?
No, but the decision grain, assigned agent, outcome, and key dates must be recoverable. Missing values can be modeled explicitly when their meaning and coverage are understood.
How are nulls handled?
Numerical values can use training-fold imputation plus missingness indicators; categorical values can use an explicit unknown category. All preprocessing must be fitted only on training data.
Do leads without appointments belong in the model?
Not when the target is whether an appointment will be won. Leads that never reached an appointment answer a different qualification question and should be modeled separately.
How much history is enough?
There is no universal row count. Feasibility depends on wins, losses, active agents, feature coverage, overlap, process stability, and the complexity of the model being tested.
Continue your evaluation
Assignment guide
Machine learning lead assignment built around the agent-prospect fit.
Continue with practical guidance, evaluation criteria, and next steps.
CRM guideAdd performance-based assignment to your MarketSharp workflow.
Continue with practical guidance, evaluation criteria, and next steps.
CRM guideBring appointment-agent intelligence into JobNimbus.
Continue with practical guidance, evaluation criteria, and next steps.
Complimentary fit review
Get a complimentary CRM data-readiness review.
We will inspect a representative history, identify the usable appointment cohort, flag leakage and quality risks, and explain whether a credible assignment evaluation is possible.
- CRM data-readiness review
- Historical model evaluation
- Constrained assignment backtest
- Plain-language opportunity review