From your CRM to a model your agents can call
ax1om is a modeling engine for AI agents: trained on your own outcome history, validated, with the result surfaced before anything acts, and explained down to the field. The same models are called by your team in the CRM and by your agents over MCP and API.
Four steps, one sitting.
The same flow every user walks, from first login to scores landing in Salesforce. No consulting engagement, no data science team.
Connect your data
Authorize Salesforce or HubSpot in one OAuth click.
Define what you predict
Pick the object and what counts as a conversion.
Train on your pipeline
ax1om trains a dedicated LightGBM model on your own wins and losses.
Deploy where your team and your agents work
Send scores out by CSV, CRM writeback, live API, or scheduled weekly rescoring.
Where agents come in.
Your agents call all three model families over MCP, the same models your team works with in the CRM. Scoring is live on the API and the CRM writeback today.
Behavior is a signal, not a footnote.
Most scoring stops at the lead's own fields. ax1om also learns from how an account behaves - campaigns, activities, and cases - and how recently.
Every object, not just fields
Campaign responses, activities, tasks, and cases roll up into per-record engagement signals - the behavior most scoring ignores because it lives outside the lead row.
Recency that decays
A touch from last week isn't a touch from last quarter. Engagement is weighted across multiple half-lives, so recent activity counts for more, automatically.
Broken out by type and window
Counts and response rates per activity type across 30, 60, and 90 day windows - the model sees the shape of engagement, not just the total volume.
What actually happens when you click Train.
No black box, no proprietary magic number. The pipeline is boring on purpose: well-understood gradient boosting, published explainability, and no cross-customer data ever.
- 01Pull
Fresh data, every run
A Cloud Run Job pulls your CRM live through the OAuth connection. Nothing cached, nothing warehoused - your CRM stays the source of truth.
- 02Engineer
Features per field type
Numbers used directly, categories one-hot, dates turned into timing, and text fields like title turned into readable level and token features.
- 03Train
LightGBM, one per org
A gradient-boosting classifier trained on a single organization's data. No cross-customer training, no shared feature store, tuned per dataset.
- 04Explain
SHAP + Feature Stability
SHAP TreeExplainer attributes every prediction, then 5-fold cross-validation measures how stable each feature's importance is across retrains.
- 05Store
Organization-scoped
Model, transformers, scores, and SHAP land in storage under an org-scoped path. The Live API serves the same artifacts at 5-20 ms.
How each field type becomes a feature.
Each field type gets the strategy that's right for gradient-boosted trees - including the messy text fields most tools quietly drop.
| Field type | Strategy | Notes |
|---|---|---|
| Numeric | Used directly | Null handling per column. No scaling required for tree models. |
| Categorical | One-hot encoded | Each value becomes its own binary feature the model can weigh independently. |
| Text (title, industry) | Levels + unigrams + bigrams | A seniority tier plus the most predictive word and phrase tokens, ranked by lift. Every resulting feature stays readable - no embeddings, no opaque vectors. |
| Boolean | Binarized | Converted to 0 or 1 with explicit null handling. |
| Date | Timing + recency | Drives conversion cutoffs and days-since-event features, not raw values. |
We diagnose your data before training.
Population, Information Value, correlation, and a predictive-strength label on every field - plus redundancy and leakage checks - so you fix problems before they reach the model.
Field diagnostics
Population rate, Information Value, correlation, and a predictive-strength label on every field - before you commit a training run.
Redundancy clusters
When several fields carry the same signal (revenue, employees, contacts all moving together), ax1om groups them and names one to keep - so importance is not split across aliases.
Leakage detection
A field that looks too good - usually one populated after the deal closes - gets flagged before it quietly inflates your AUC.
| Rank | Field | Population | Signal |
|---|---|---|---|
| 01 | LeadSource | Complete | Strong |
| 02 | Title | Mostly filled | Strong |
| 03 | NumberOfEmployees | Mostly filled | Medium |
| 04 | Industry | Partly filled | Medium |
| 05 | AnnualRevenue | Sparse | Weak |
| 06 | Rating | Sparse | None |
Not just a score.
A full view of how the model performs, why each record landed where it did, and whether the explanation will survive your next retrain.
AUC and lift curve
Standard classification metrics so you can judge whether the model is worth deploying before you wire it into anything.
SHAP feature attribution
Per-record SHAP values showing exactly which fields drove each prediction, and by how much. No mystery scores.
Every feature traces to a field
Each engineered feature maps back to the exact CRM field and value it came from - "Title: VP", "Country: United States", "campaigns in the last 30 days". Nothing opaque.
Feature Stability Score
A 0-to-1 measure of how consistent each feature's importance is across validation folds. High FSS means the explanation holds up when you retrain next quarter.
Score distribution
Histogram of conversion likelihood across your records so you can set sensible routing thresholds.
Scores land where your reps work. Nothing else is touched.
ax1om reads your CRM read-only by default. The only thing it writes back is the score - and an optional SHAP summary - and only to the fields you explicitly designate. OAuth tokens are Fernet-encrypted before they reach the database, and PII is stripped from stored model artifacts, so only aggregated features and scores persist.
The full posture on infrastructure, encryption, and isolation lives on the trust page. The privacy policy covers retention, GDPR and CCPA rights, and the subprocessor list.
Common questions.
How long does training actually take?
Most models finish in under 10 minutes. Larger datasets above 100,000 records can take up to 30 minutes on the first run. Retrains are usually faster because the schema and feature transformers are cached.
How does ax1om use a messy field like job title?
It turns title into readable features: a seniority tier (C-suite, VP, director, manager, senior IC, IC, intern) plus the word and phrase tokens that actually predict conversion in your data, ranked by lift. You see "Title level: VP" contributing a specific amount in the explanation - not an embedding, not a black box. The same approach handles other high-cardinality text like industry.
What if my CRM data is messy?
That's the normal starting condition. ax1om's field diagnostics call out population rate, correlation, and stability per field before training, flag redundant fields that carry the same signal, and warn on fields that look like target leakage. You can exclude problem fields, and the model down-weights sparse or noisy columns automatically.
Do you sync my CRM data into a warehouse?
No. ax1om pulls data fresh from your CRM on every training run. No nightly sync, no warehouse middleware, no stale feature store. The CRM stays the single source of truth, and your training data reflects whatever it looks like at the moment you click Train.
How does writeback work in Salesforce and HubSpot?
You designate a custom field on each object, and ax1om writes the current score (and optionally a SHAP summary) to that field. It only touches the fields you map - it never modifies your existing data.
What runs behind the scenes when I click Train?
A separate Cloud Run Job (`ax1om-train`) spins up, pulls your data fresh, runs feature engineering, trains LightGBM, computes SHAP with TreeExplainer, measures Feature Stability across folds, and writes artifacts (model, scores, SHAP values, feature builders, schema, metadata) to organization-scoped storage. Then it updates your run status in the dashboard.
Ready for numbers your agents didn’t make up?
Connect your data, train on your own outcomes, and every number your team and your agents act on is trained, validated, and explained.