Capacity & Technical Questions/Discussion
Responses to the four areas raised after the technical discussion: sensitive IP and frontier AI, usage and scale, shared project storage, and next steps, with researched market numbers for Ireland, the UK, Europe and Australia, checked against the May 2025 feasibility study. Every section has a copy button for pasting into email, Excel or a document.
Executive summary
The headline answers, in short form.
AI use & IP protection
Comfortable using Claude, GPT and similar tools through approved business accounts or APIs. Credentials, raw student and patient records, confidential contracts and commercially sensitive strategy stay out of general AI tools.
EDUCATE capacity
Launch at 1–2 campuses (100–500 students, 50–100 simultaneous), scaling to 25–50 campuses, 10,000–25,000 students and 1,000–2,500 simultaneous users.
REGULATE capacity
Launch at 5–20 pharmacies (25–200 staff, 10–30 simultaneous), scaling to 500–2,000 pharmacies, 2,500–20,000 staff and 500–2,000 simultaneous users.
AI volume
500 / 1,000 / 2,000 words per complete scenario, and 100–500 words of feedback per attempt. Combined at scale, roughly 206M words a month at base, so a full low / base / high cost model is the next step.
AI use and protection of our information
Where we stand on model use, IP exposure and provider risk.
Approved tools, approved accounts
We are comfortable using Claude, GPT and similar tools for planning, design, coding and testing through approved business accounts or APIs.
Review before proceeding
| Area | What we're checking |
|---|---|
| Training use | Whether our inputs and outputs can be used to train or improve the provider's models. |
| Retention | How long prompts and outputs are kept, and where. |
| Access | Who at the provider can access our data, and under what controls. |
| Hosting | Where processing and storage take place, including data residency. |
| Contractual protections | Zero-retention terms, DPAs, indemnities and termination rights. |
What stays out of general AI tools
- • Credentials and secrets.
- • Raw student and patient records.
- • Confidential contracts.
- • Commercially sensitive strategy.
Share only the code or requirements needed. 'No-training' terms alone do not eliminate disclosure risk.
Provider competition: a strategic risk we could accept, with safeguards
- • Keep our code, content and workflows under our control.
- • Use replaceable model integrations rather than depending on one provider.
- • Evaluate an alternative provider.
Where our differentiation comes from
- • Pharmacy academic and practice expertise.
- • Trusted content.
- • Useful workflows.
Comparison to the May 2025 feasibility study
Where the numbers agree, where they differ, and what the study didn't attempt.
The May 2025 feasibility study (Table 1) gave education-only counts for four countries. Our numbers add Ireland's 2025/26 school expansion, community pharmacy and pharmacist counts, and separate the education opportunity (EDUCATE) from the medication-error / practice opportunity (REGULATE).
| Market | Study's figure | Our figure | Read |
|---|---|---|---|
| UK: pharmacy schools | 30 | 37 | We use the Pharmacy Schools Council's current count; the study's 30 looks like an older figure. |
| UK: students | ~14,000 | 18,000–24,000 (mean ~21,000) | The study's figure sits inside our range, near the low end. |
| Australia: pharmacy schools | 18 | ~20 | Close agreement (the current list of accredited programmes). |
| Australia: students | ~5,400 | 5,500–8,000 (mean ~6,500) | The study's figure sits at the very bottom of our range. |
| Ireland | Not broken out | 6 schools, ~200→350 intake, 900–1,400 enrolled | We add this explicitly; the study's Table 1 only covers USA/UK/Canada/Australia/India. |
| Europe (EU/EEA-wide) | Not estimated at this scale | ~250–300 institutions, 80,000–180,000 students | The study only names individual countries via stakeholder interviews, not an EU-wide figure. |
| Global cost of medication errors | $42B/year (WHO) | $42B/year (WHO) | Aligned: both use the same headline figure for the REGULATE opportunity. |
From the study only: healthcare simulation market
$2.34B (2024) rising to $6.89B (2034) at an 11.4% CAGR. Useful EDUCATE market context we hadn't included in our numbers.
From the study only: per-student pricing anchor
Stakeholders suggested €15–25 per student per year is bearable, a useful pricing anchor for EDUCATE's subscription model and consistent with our scale estimates.
Net read: our AI-usage, concurrency and cost figures are new work; the study stops at market sizing and revenue models and doesn't model AI token volume or infrastructure load. Our market-size figures broadly corroborate the study's UK and Australia numbers, sit a little higher on UK schools, and add Ireland and an EU-wide estimate the study didn't break out at that scale.
Usage and scale: EDUCATE and REGULATE
Capacity planned separately for each platform across launch, grow and scale.
For capacity planning we plan EDUCATE and REGULATE separately across launch, grow and scale. The figures below are proposed sizing assumptions to validate through pilots, not customer forecasts or commitments.
EDUCATE
: pharmacy education and scenario generationSynthetic pharmacy cases, lecturer review before release, reusable scenarios and supervised feedback.
Across modules and campuses, with broader integrations as we scale.
Student identities remain separate from AI prompts, with human oversight of assessment.
Suggested capacity across the platform
| Stage | Campuses / teaching sites | Registered students | Simultaneous users |
|---|---|---|---|
| Launch | 1–2 | 100–500 | 50–100 |
| Grow | 5–10 | 1,000–5,000 | 250–500 |
| Scale | 25–50 | 10,000–25,000 | 1,000–2,500 |
AI assumptions (low / base / high)
| Assumption | Low | Base | High |
|---|---|---|---|
| Words per complete scenario (including model answers and feedback) | 500 | 1,000 | 2,000 |
| New scenarios per lecturer, per module, per term | 10 | 25 | 50 |
| Student-specific feedback per attempt (words) | 100 | 250 | 500 |
| Attempts per active student per month | 10 | 25 | 40 |
Derived output volume
| Measure | Low | Base | High |
|---|---|---|---|
| Words per lecturer, per module, per term (scenario generation) | 5,000 | 25,000 | 100,000 |
| Words per active student per month (feedback) | 1,000 | 6,250 | 20,000 |
| Monthly output, launch (500 students) | 0.5M | 3.1M | 10M |
| Monthly output, grow (5,000 students) | 5M | 31M | 100M |
| Monthly output, scale (25,000 students) | 25M | 156M | 500M |
Stage totals use the top of the registered-student band and assume all registered students are active. Approved scenarios are reusable across students, so a student opening a case does not trigger fresh generation. If live AI dialogue is introduced, cost it separately.
REGULATE
: medication error reporting, SOP management and regulatory functionalityMedication-error and near-miss reporting, SOP versioning and staff acknowledgement, action tracking and audit trails.
Pharmacy-group dashboards and reporting.
Wider integrations. AI can help draft summaries and SOP updates for pharmacist approval.
Suggested capacity across the platform
| Stage | Pharmacies | Registered staff | Simultaneous users |
|---|---|---|---|
| Launch | 5–20 | 25–200 | 10–30 |
| Grow | 50–200 | 250–2,000 | 50–200 |
| Scale | 500–2,000 | 2,500–20,000 | 500–2,000 |
AI assumptions (low / base / high)
| Assumption | Low | Base | High |
|---|---|---|---|
| AI-assisted tasks per pharmacy per month | 20 | 50 | 200 |
| Output words per task | 200 | 500 | 1,500 |
| Words per pharmacy per month | 4,000 | 25,000 | 300,000 |
These tasks are summaries, SOP drafts and queries. They are not predicted incident counts, and routine record saving should not require an AI call.
Derived output volume
| Measure | Low | Base | High |
|---|---|---|---|
| Monthly output, launch (20 pharmacies) | 80,000 | 0.5M | 6M |
| Monthly output, grow (200 pharmacies) | 0.8M | 5M | 60M |
| Monthly output, scale (2,000 pharmacies) | 8M | 50M | 600M |
Stage totals use the top of the pharmacy band.
Combined at scale (top of band)
| Measure | Low | Base | High |
|---|---|---|---|
| Output words per month | 33M | 206M | 1.1B |
| Output words per year | 400M | 2.5B | 13.2B |
| Output tokens per year (×1.33) | 530M | 3.3B | 17.6B |
| Indicative model cost per year ($5 / 1M output tokens) | ~$2,600 | ~$16,400 | ~$88,000 |
Indicative only, and output tokens only. It excludes input tokens, retries and live dialogue, so the full low / base / high cost model is a next step.
Privacy and anonymisation
Masking student and patient records before queries reach a model.
A proposed route is to mask student and pharmacy patient records before queries reach a model. Maya Data Privacy's AISafe is the first candidate to evaluate, alongside comparable tools.
mayadataprivacy.com/aisafeA real-time privacy layer that sits between users and any LLM. It detects and replaces personal data in prompts before they reach the model, then re-identifies the response so the user sees real context. It is LLM-agnostic (ChatGPT, Claude, Gemini and custom enterprise models), can run as middleware intercepting API calls, and the vendor states it is ISO 27001:2022 and SOC 2 certified and headquartered in Dublin. The vendor also offers AppSafe, which creates anonymised copies of production databases.
What we need to test before adopting
| Criterion | What we're checking |
|---|---|
| Free-text handling | Clinical narratives, dispensing notes and incident descriptions, not just structured fields. |
| Clinical meaning | Whether masking degrades the quality or safety of the AI output. |
| Deployment location | Where the service runs and where data is processed. |
| Retention | What is kept, for how long, and where. |
| Cost | Per-seat or per-call pricing measured against our volume assumptions. |
| Comparable tools | Reviewed against the same criteria when there is time. |
Important caveat
Reversible masking is pseudonymisation, not anonymisation. It reduces exposure but does not remove our data-protection obligations. Identifiable records should remain in our controlled system, and a privacy assessment must be completed before any live personal data is used.
Shared project storage
Where the team puts project information.
Recommendation: Project specific SharePoint.
| Folder | Contents |
|---|---|
| 00 Project Management | Plans, minutes, decisions log |
| 01 Requirements & Design | PRDs, architecture, wireframes |
| 02 Clinical & Educational Content | Scenario design, evidence framework |
| 03 Technical | Repositories, environments, infrastructure notes |
| 04 Commercial & Legal | Contracts, DPAs, NDAs, restricted access |
| 05 Assurance & Compliance | DPIA, ISO, security, policies |
| 06 Research & Evidence | Papers, guidance, references |
- • No unredacted student or patient data, and no special-category data, in the shared area.
- • Keep sensitive material inside the institutional tenancy, not in personal drives.
Next steps and decisions needed
The cost model, the build-for-launch checklist, and the assurance items to schedule.
Next steps
- Build an initial low / base / high cost model using these assumptions: input and output tokens, retries, and peak AI requests costed separately from simultaneous platform users.
- Confirm the launch targets for EDUCATE and REGULATE, and validate them through pilots.
- Project documents live in Teams. Greg can help with access.
- Build for launch with tenant isolation, access controls, backups, usage limits and queued generation, then expand using pilot measurements and load tests.
Assurance roadmap: the "figure out as we go" list
| Item | Why it matters |
|---|---|
| ISO/IEC 27001 | Information security management, typically required for institutional procurement. |
| ISO/IEC 42001 | AI management system, increasingly requested for AI products in education and health. |
| GDPR DPIA + processor register | A data protection impact assessment, plus a register of every AI provider and DPA. |
| Medical device (MDR) screening | Required if any feature could be read as clinical decision support. |
| WCAG 2.2 AA accessibility | Condition of adoption for most institutions. |
| Clinical & educational governance | Named academic reviewer per scenario, sign-off before learner release. |
Sources and assumptions
Where the numbers come from, and how the estimates were derived.
- •Ireland: IPU, Key Facts on Community Pharmacy (2024): 1,906 community pharmacies, 5,081 community pharmacists; IPU Review, September 2025: six pharmacy schools and 150+ additional graduates annually; PSI register for pharmacist numbers.
- •United Kingdom: Pharmacy Schools Council: 37 pharmacy schools; Pharmaceutical Journal: 4,570 MPharm acceptances in 2024; Pharmacy Market Review 2024: ~13,800 community pharmacies; GPhC register for pharmacist numbers.
- •Europe: EAFP: 195 higher-education institutions providing pharmacy studies in 25 EU states (2011); PGEU: over 500,000 community pharmacists working in 200,000 pharmacies.
- •Australia: Pharmacy Council of Australia accredited programmes; Pharmacy Guild of Australia, Pharmacy Profile 2024: 5,935 community pharmacies; Pharmacy Board of Australia 2024/25: 40,913 registered pharmacists, ~1,444 domestic first-time registrants.
- •Capacity and AI assumptions: The EDUCATE and REGULATE capacity and AI figures are the proposed sizing assumptions, agreed for planning and to be validated through pilots, not customer forecasts or commitments. EDUCATE and REGULATE are planned separately across launch, grow and scale.
- •Derived volumes: Monthly output is the AI assumption multiplied out at the top of each stage band: EDUCATE uses registered students × attempts per month × feedback words per attempt, plus scenario generation per lecturer; REGULATE uses pharmacies × tasks per month × words per task. Stage totals assume all registered students are active. Token conversion uses 1.33 tokens per word; the indicative cost uses $5 per million output tokens and excludes input tokens, retries and live dialogue.
- •Market sizing: Enrolment = intake × programme length. Concurrency assumes 10–15% of enrolled users active at peak in a teaching window. The market figures are planning estimates assembled from public sources, not audited data, and should be refined once pilot institutions are confirmed.