Coapro

Legal

Model Training

Effective 24 September 2026

What do the AI models behind Coapro learn, from what data, and who decides? Written plain: only what the code does. The short answer — we train our own models, on redacted features, never on your employees’ verbatim words, and never by sending anyone’s prompts to a third party.

The models

Coapro’s coaching intelligence uses three kinds of models. Two are rules you can read; one is trained by us.

1. The policy engine (deterministic). Decides allow / warn / block. It is rules and patterns you can inspect — every policy category, every pattern. No learning is involved in deciding what an employee may send. Nothing learned ever sits in the send path.

2. The rubric system (deterministic). Progress bands, practice areas, and coaching framing come from a fixed rubric and a published coaching corpus. This is the default coaching brain, and it is fully explainable: any manager can see why an employee is “steady” in one practice area and has an “opportunity” in another.

3. Our trained ranking models (machine learning, ours). Small models we train ourselves on our own infrastructure. They order which coaching topics to recommend to which team, and prioritize which artifacts a reviewer looks at first. They do not decide anything an employee can or cannot send; they do not rate, score, or rank employees.

What we train on

  • Synthetic data. Most of the initial training set is synthetic practice journeys generated from our published coaching corpus — no real person’s data at all.
  • Redacted feature statistics (aggregate, default-on, contract-gated). Outcome-class counts, band histories, recency-weighted category features — the same aggregates the console already shows. Never verbatim text, never re-identifiable. A customer can exclude even this with the Training Exclusion option in their order form, and the product loses no functionality.
  • Opt-in interaction outcomes. A customer may additionally contribute pseudonymized accept/reject outcomes to improve ranking quality. Never required.

What we never train on

  • Verbatim prompt content — under any configuration. Verbatim prompt text is short-retention by design (deleted on the customer’s configured clock). A trained model would be a retention window that never forgets, so we do not build one.
  • Quiz answers (purged after 30 days; snapshots only).
  • Any customer’s data outside the contract terms above. No third-party training runs exist to send it to.

Where the models run

Training and inference run on our own infrastructure. The trained models serve recommendations and review-queue triage only — always outside the send path, and automated targeting of coaching stays off by default (and off for EU tenants pending the EU AI Act intent analysis).

What an employee sees

Coaching language describes practice areas and opportunities — never a score, never “failing”, never “risky”. An employee sees their own data on their own My AI Usage page and can export it without asking anyone.

Runtime text generation

Coaching digest emails and nudge text are drafted at runtime by a large language model hosted by Ollama Cloud, which receives redacted, non-verbatim context and — per its published terms — does not train on API inputs. It appears on our sub-processor list. If that ever changes, this page and the list change first, and customers are notified before the change takes effect.