The models
CoaPro uses two deterministic systems you can read, and small machine-learning models that we build and run ourselves.
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 coaching framework (deterministic). Practice areas, progress bands, and coaching framing are deterministic and explainable: any manager can see why an employee is “steady” in one practice area and has an “opportunity” in another.
3. Our own models (machine learning, built and run by us).
- CoaPro Lens reads a prompt after it has been sent and assesses it against the practice areas: did it give the AI enough context, check the output, fit the task? Those assessments are the evidence behind each employee’s progress bands. Lens’s security reader also flags prompts that appear to share sensitive internal material, which your published policy can make a reportable signal.
- CoaPro Coach (drafts coaching lessons) and CoaPro Compass (reads your AI policy and drafts rules for an administrator to review) are built but not yet used in the product. When either goes live, this page changes first.
None of these models decides what an employee may send. Lens produces coaching evidence, not a performance rating: its output is shown as practice areas and bands, never as a score, and CoaPro is built for skill development only, never employment evaluation.
What the models read when they run
Lens runs after the prompt has gone, on the stored copy of the prompt: the verbatim text while it is inside your retention window, otherwise the redacted copy. It runs on CoaPro-operated hardware, never on a third-party AI service. What it keeps:
- Scores (practice area and band) — kept as long as the event record.
- A short rationale — treated as prompt content: deleted on the same clock as the prompt and covered by the same legal hold.
- Verbatim fragments — never written anywhere.
What we train on
No customer data has been used to train any CoaPro model. The training data is:
- Synthetic examples — practice prompts, lessons, and AI policies written from our coaching corpus and from made-up company specifications.
- Public, license-checked text — for example, research datasets of prompts people chose to publish, and public government AI policies. Every source is checked against a license allowlist and scrubbed for personal data and secrets before use.
- Third-party AI models as drafting and labelling assistants — large language models helped write the synthetic examples and label the public ones. They were given synthetic or public text only, never customer data.
- Our own team’s prompts — used only to test model quality, never to train.
Our contract allows one further source, and we say so because it is in the DPA: aggregate, de-identified feature statistics (outcome counts, band histories — never text) from customers who have not chosen the Training Exclusion in their order form. We do not use this today. If we start, this page will say so first, and the Training Exclusion removes it with no loss of functionality.
What we never train on
- Customer 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 sent to a third party for training. No such training run exists.
Where the models run
Training and running all happen on CoaPro-operated hardware, always outside the send path. 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
One optional feature uses an outside model: coaching nudge text, when an administrator turns nudge generation on (it is off by default). Nudges are drafted by a large language model hosted by Ollama Cloud, which never receives the prompt text, redacted or otherwise, and, per its published terms, does not train on API inputs. It appears on our sub-processor list. Coaching digest emails are built from templates, not generated. If any of this changes, this page and the list change first, and customers are notified before the change takes effect.