Trained models for conversion scoring, customer health, and timing, built on your own outcome history, validated before anything acts on them, and explained down to the field. Your agents call them over MCP and API, starting with scoring.
Your agent: judgment, orchestration, language · ax1om: learned prediction, with the factors behind every number
Train on your outcomes, pass the validation gates, predict with the factors behind it. Your agent learns the pattern once.
Which records convert, and the fields behind each score.
Which customers are drifting, learned from your own renewals.
Which months an account is likely to come into market.
What your agent can call over MCP today, and why it needs to.
assess_fit() "verdict": "fit" score_records({ explain: true }) We cannot show your pipeline, and we will not show you someone else's, so here is the readout itself, in the shape it takes on your own first run. This is the dashboard you get, not a result we are promising you. It reports AUC and lift at the top of the list so you can judge the model before you deploy it, conversion by score bucket so you can see where the list stops paying, and a score distribution for setting thresholds. The figures are yours, filled in the moment your model finishes training. Beside it, the factors: every prediction ships with the fields that drove it, because a score with reasons gets used and an opaque one gets ignored.
Each bar is one bucket's own conversion rate. Where the bars drop under the baseline is where the list stops paying.
How your records spread across the score range, which is what you set a cutoff against.
| Field | Impact | Stability | Expand |
|---|---|---|---|
| LeadSource | ↑ Raises the score, relative importance shown as a bar | Stable | |
| Title | ↑ Raises the score, relative importance shown as a bar | Stable | |
| NumberOfEmployees | ↑ Raises the score, relative importance shown as a bar | Moderate | |
| Industry | ↓ Lowers the score, relative importance shown as a bar | Moderate |
Conversion scoring is live on ax1om's MCP surface today. Point an agent at it and it can check fit, prepare a scrubbed export, train a model on your own conversion history, and score records live, with the factors behind each score. Agents authorize with per-user OAuth: you approve exactly what each agent may do, and you can revoke it any time.
These are the questions teams arrive with. Every one is answered the same way: a model trained on your own outcomes, put through the validation gates, returning the factors behind what it predicted.
A model trained on your own conversion history ranks today's inbound, and the fields that drove each score write back to the record. Your reps read why an account sits where it does in the object they already work, so the ranking gets used instead of ignored.
Conversion scoringBack-test fit against the conversions your CRM already recorded rather than an ICP doc written in a workshop and refreshed never. Every scoring rule carries its own performance readout, so you can split one when it turns out to be hiding two different populations.
Conversion scoringInstead of a product-qualified threshold somebody estimated once, the model learns which behaviors preceded payment across your own accounts. Each account comes back with a score and the factors behind it, on the same rails as the rest of your scoring.
Conversion scoringRank the accounts you lost on how they look now, rather than sending the whole list a fresh round of email. Losses are native fuel for this engine: they are the outcome data most scoring tools throw away.
Conversion scoringYour agent calls score_records over MCP and gets a score per record back, with the factors behind it when explain is on. The ranking comes out of a model trained on your outcome history, not out of the agent’s own judgment about which rows look promising.
MCP referenceFour of the MCP tools run entirely on your machine and never touch the network. An agent can check whether your data fits, generate the scrub spec for an export, and lint that export against it with nothing leaving your org.
ax1om for AI agentsAuthorize Salesforce or HubSpot in one OAuth click. Read-only by default, OAuth tokens encrypted at rest. ax1om reads your schema and record counts automatically.
Pick the object and what counts as a conversion. ax1om auto-classifies which fields are predictive by population and stability, so you rarely have to touch it.
ax1om trains a dedicated LightGBM model on your own wins and losses. No cross-customer data, no shared models. Usually under 10 minutes.
Send scores out by CSV, CRM writeback, live API, or scheduled weekly rescoring. SHAP shows exactly why each record ranked where it did.
Agent pilots stall the same way every time: nobody signs off on a number the agent made up. Revenue calls route on a made-up metric until one goes visibly wrong, and the org concludes agents can't do this. The fix is not a smarter prompt; it is a trained model your agent can call.
ax1om runs on providers that hold independent SOC 2 Type II, ISO 27001, ISO 27017, ISO 27018, and PCI DSS Level 1 certifications. Our own SOC 2 Type II audit begins Q4 2026.
You see every factor behind every score, and nothing routes until you say so. Your rules, your thresholds, your writeback.
You know what your team, and now your agents, could do with prioritization they actually trust. Everything you've tried either can't be explained or can't be maintained. That's not a you problem; it's how the tools were built.
Check two things: what the model trained on, and how much of the reasoning you can see. Einstein only builds a model from your own leads once you have about 1,000 recent leads and 120 conversions; below that it scores you with a model built from other companies’ data, and its explanations stop at field names. HubSpot’s own docs describe their predictive score as “blackbox machine learning.” ax1om trains only on your conversion history at any size, and every score shows the factors behind it.
ax1om is specifically built for Series A-C companies. The guided setup adapts to your data volume and trains a working model on as few as 50 converted opportunities. You don't need years of CRM history, and the model improves as your data grows.
That's the point. ax1om handles the entire ML pipeline: feature engineering, model training, scoring, CRM writeback, and weekly retraining. No code, no notebooks, no data science expertise required.
Scores are advisory, never automated actions, so a wrong score costs you a judgment call, not a lost deal. Every score ships with the SHAP fields that drove it, so a rep can see the reasoning and override it. ax1om also down-weights sparse and noisy fields automatically and flags likely target leakage before training, which catches the most common causes of confidently-wrong scores up front.
All of it. Nothing routes until you say so. You set the thresholds, you designate which CRM fields ax1om writes to, and scores stay advisory: ax1om never takes an action on a record by itself. Every score ships with the SHAP fields behind it, so you can audit a decision before you build a rule on it.
That is what the explanation is for. A black-box 91 gets ignored; a score that reads "↑ VP-level title (1.9x) · ↑ 3 meetings last 14d · ↓ free email domain" gets acted on. That line is written into a field on the Lead or Account record your reps already work, so there is no new tool for them to adopt.
Yes. Conversion scoring is live on ax1om’s MCP surface today. Point an MCP client at it - Claude Desktop, Claude Code, and Cursor all work - and your agent can check fit, prepare a scrubbed export, train a model on your own conversion history, and score records live with the factors behind each score. Agents authorize with per-user OAuth: you approve exactly what the agent may do, and you can revoke the grant at any time. Customer health and timing ship in the product; scoring is what agents call over MCP today.
Only what you choose to send, and the workflow is built scrub-first. The export spec defines which field classes may leave your org and which never may, and direct identifiers are dropped or hashed before anything uploads. Four of the eleven MCP tools run entirely on your machine and never touch the network, so an agent can prepare and lint an export with nothing leaving the org. What comes back to the agent is scores and factors, not your records. Your models train on ax1om’s infrastructure, on your data only, and are never shared across customers.
Connect your data, train on your own outcomes, and every number your team and your agents act on is trained, validated, and explained. The free tier shows you on your own data.