TL;DR — for the AI risk team

Model: Claude (Anthropic) — current frontier model. Coachy uses Claude via authenticated API, not a self-hosted weight.

Scope: ST1512 standard and KSBs only. System prompt scopes Claude away from generic-AI behaviour.

Guardrails: Coachy explains and coaches; it never writes portfolio evidence. EPA integrity is a hard product line.

Data: Apprentice conversations are not used to train AI models. UK data residency, GDPR-compliant, OneFile-aligned governance.

Why this post exists

If you're an L&D Director, an HR Director, or you sit on an AI risk committee, you've probably been asked the same question half a dozen times in the last quarter: "what AI tools are we letting our people use, and on what basis?". AI tutors used by apprentices are squarely in scope. This post is the document you can forward to your internal AI risk team to answer their questions about Coachy — in detail, in plain English, with the lines we hold and the lines we don't.

You can also read the public Coachy product page for the buyer-facing summary. This post sits underneath that — aimed at the technical and governance reviewer who needs more.

1. The model under the hood

Coachy is built on Claude, the frontier AI model from Anthropic. Specifically:

Model family
Claude (Anthropic) — production tier. Coachy moves to newer Claude versions as they release; we don't fork or self-host weights.
Access pattern
Authenticated API. Each apprentice conversation is a server-side call to the Anthropic Claude API; the model itself runs on Anthropic infrastructure, not in your tenant or ours.
Why Claude
Claude leads peer models on the qualities that matter for tutoring: nuanced explanation, careful reasoning, respecting safety guidelines, and resisting the temptation to give straight answers when a Socratic question is the better teaching move. We benchmarked Claude vs alternatives during build and continue to compare.
Versioning
Coachy currently supports apprentices on the AI & Automation Practitioner Level 4 (ST1512) standard only. Other apprenticeship levels (L5 leadership, L6 ML fellowship) are on the Coachy roadmap. Coachy follows Anthropic's stable production model line. Test cohorts evaluate new model versions before they reach apprentices, so the model behind Coachy is always vetted, never untested.

2. The scope we hold Coachy to

One of the biggest mistakes AI tutors make is being too generalist. Coachy is the opposite by design.

Standard-scoped, not topic-broad

Coachy's system prompt scopes Claude to the AI & Automation Practitioner Level 4 (ST1512) standard, its KSBs, and the TESS pedagogy. When apprentices ask off-scope questions, Coachy redirects them rather than answering generically.

Standard-deep, not internet-broad

The system prompt is grounded in the ST1512 standard from Skills England, the KSB framework, and TESS's own learning materials. Coachy speaks the apprentice's curriculum, not the general internet.

British English, UK context

Apprentices are UK-based, working in UK organisations, against UK qualifications. Coachy is configured for British English spelling and grammar, UK regulatory context, and UK funding mechanics where these come up in conversation.

3. The guardrails that protect EPA integrity

This is the single most important section of this post. End-point assessment (EPA) is independent: the EPA organisation, not TESS, awards the grade. Anything written by AI on an apprentice's behalf can be flagged at gateway, can fail at EPA, and damages the apprentice's credentials for years afterwards. Coachy is built so that doesn't happen.

The Coachy EPA-safety pledge

Coachy will explain a concept, help an apprentice work through their thinking, point at relevant KSBs, and ask Socratic questions — but it will not write evidence on the apprentice's behalf. Portfolio evidence has to be the apprentice's own work to pass EPA, and Coachy is built to support that, not undermine it.

What Coachy does

  • Explains concepts in plain English, multiple ways
  • Asks Socratic questions to help the apprentice arrive at the answer themselves
  • Points at the relevant KSB an apprentice is working on
  • Helps apprentices practice for EPA observation and discussion
  • Is available 24/7 across mobile and desktop

What Coachy will not do

  • Write portfolio evidence on the apprentice's behalf
  • Generate fabricated examples of workplace activity
  • Replace the named human skills coach
  • Pretend to be authoritative outside the ST1512 scope
  • Act as the formal grader of any apprentice work

These guardrails are enforced through the system prompt, the conversation policy, and ongoing review of conversations by the TESS quality team. We treat any drift from these guardrails as a P1 product defect.

4. Data handling, privacy & governance

The questions every L&D risk team asks. Here are the answers.

Training
Apprentice conversations with Coachy are not used to train AI models. We do not consent to model-training use of apprentice data with Anthropic; conversations are processed for inference only.
Storage
Conversation history is stored securely in TESS-controlled infrastructure within the UK / EEA. Apprentices have visibility of their own conversations.
Sharing
Apprentice conversations are visible to the apprentice and (where appropriate, with their consent) their TESS skills coach. They are not visible to other apprentices, other employers, or third parties.
Data residency
UK / EEA. Anthropic's API endpoints used by Coachy are located in regions compliant with UK GDPR; data does not transit beyond contracted data-processing agreements.
GDPR
TESS is a UK-registered organisation operating under UK GDPR. Coachy inherits the same data governance, retention policy, and subject-rights workflow that the wider TESS Group uses for OneFile-held apprentice data.
Authentication
Magic-link authentication using the apprentice's existing OneFile email. No new account, no separate password, no separate identity store. Authentication audit logs are retained per the TESS data retention policy.
Right to delete
Apprentices can request deletion of their conversation history at any time, in line with UK GDPR right-to-erasure. Standard turnaround is 30 days; usually much faster.
Sub-processors
Anthropic (model inference). TESS-controlled UK/EEA infrastructure (storage, retrieval, application logic). No further third-party sub-processors involved in handling apprentice conversation data.

5. Pedagogy — how Coachy decides how to respond

This is where the system prompt does most of the work.

Resist the answer; reward the question

Coachy is configured to resist giving direct answers to apprentices. When an apprentice asks "what's the answer to X", Coachy is more likely to ask a clarifying question, or to walk the apprentice through the reasoning, than to provide the answer outright. This is borrowed directly from the Khan Academy / Khanmigo "Socratic guidance" approach.

Point at the KSB

Coachy is aware of the ST1512 KSB framework. When relevant, it nudges apprentices toward thinking about how the question they're working on maps to a specific KSB they need to evidence. This is a pedagogical move, not a portfolio-writing move; the apprentice still has to produce the evidence themselves.

Cite when uncertain

Where Coachy is asked about external facts (regulatory dates, qualification details, exam mechanics) it tries to be explicit about what it is and is not certain about, and points apprentices at the canonical source. AI hallucination risk is real; we don't pretend it isn't.

6. Ongoing oversight

Coachy is not "set up and forget". It is operationally owned, with a named team and a measurable review cadence.

  • Named owner: a TESS senior leader is accountable for Coachy's pedagogy, safety, and policy compliance.
  • Conversation sampling: the TESS quality team samples Coachy conversations weekly, looking for drift from the guardrails (especially anything that looks like Coachy starting to write evidence).
  • Defect escalation: any apprentice-flagged or quality-flagged conversation that breaches the guardrails is treated as a P1 product defect and triaged into the next release window.
  • Model version review: when Anthropic ships a new model version, the TESS quality team runs a fixed test set against it before promoting Coachy to that version.
  • Annual review: the entire Coachy policy — system prompt, scope, governance — is reviewed annually and republished. The version of this document is fixed for transparency.

7. What's next on the Coachy roadmap

Active development items for the rest of 2026:

  • Persistent sign-in & conversation history. Direct response to apprentice feedback; resumes conversations between sessions.
  • Voice mode. For apprentices on commute, walking the warehouse floor, or wanting to practice EPA observation out loud.
  • OneFile portfolio sync. Coachy reads the apprentice's KSB progress directly so it can suggest evidence opportunities precisely.
  • Save-as-evidence. Turn a reflective conversation into a portfolio entry the coach can review and approve. A capability unique to Coachy.
  • Employer dashboard. Cohort engagement, KSB heatmap, EPA readiness traffic-light, all visible to the L&D team responsible for the cohort.

None of these change the lines drawn in this post. Coachy will still explain and coach; it will still not write portfolio evidence; and it will still treat apprentice data with the same governance the TESS Group already operates.

Want this in a 30-minute conversation with your AI risk team?

We'll happily walk your AI risk reviewers through Coachy — live demonstration, sample conversations, the system prompt approach, and the data flow. Most of these conversations conclude in a single call.

Book a 30-minute Discovery Call

Related Reading

Book a Free Discovery Call4.9/5