Which path?
An honest read — this is about your tenant and your QA posture, not which model is “better.”
Copilot-free, on a local frozen SLM
- Reproducibility matters — you can freeze the model (tag+quant+SHA-256+seed) into a re-runnable validation artifact.
- You want to own the reliability controls (schema-constrained decoding, allowlist validator, human gate) in your own SAS/R code.
- You want to avoid the Copilot-specific IT frictions (DLP-connector block, the 01-Nov-2026 AI-Builder credit clock, Copilot Studio licensing + agent-publish gate, cloud egress).
- Honest cost: a small (7–8B) model hallucinates more than a frontier model, and it is a real, qualified instrument to validate, own, and maintain (its own IT hurdles).
M365-native, on Copilot + AI Builder
- You are already entitled, the DLP question is sorted (the three connectors in one group), and the content is genuinely de-identified ops-only.
- You want it M365-native and fast — for the ops/drafting layer Copilot is genuinely capable, at a few dollars/month.
- You cannot freeze the model (no reproducible validation artifact), and you carry the DLP / credit-clock / licensing / agent-publish frictions — configuration states you can resolve, but real.
Track 1 · The Copilot-free local-SLM path (new)
Local SLM operating wiki
The full Copilot-free build: a local scheduled Graph job, an on-prem frozen SLM doing schema-constrained extraction (LOCALMIND), a human-gated parse-guard, SharePoint store, a local RAG “ask the KB”, and a deterministic-first weekly digest — with the honest reliability story and the IT-readiness contrast kept in the headline.
The engine & the IT playbook (reused)
The LOCALMIND %slm_* library (loopback-guard, schema-constrained, allowlist-validator) that does the extract+validate+gate, the on-device worked example, and the IT/Security/QA enablement runbook with the hurdles + open-weight-licence landscape to hand to IT.
Track 2 · The M365 Copilot path
Biostatistics × Copilot operating wiki
The original: the self-updating Outlook→Power Automate→AI Builder→SharePoint→Copilot-agent pipeline, with build screencasts (the flow grows action-by-action), the worked example, the prompt library, and governance.
Copilot IT readiness
What to ask IT for the Copilot path (the DLP-connector group, the AI-Builder credit clock, the agent-publish control), and the ships-Monday no-AI fallback.
Why the AI step is kept reproducible — the evidence
Field evidence — Copilot on a real QC task
A verbatim, page-cited record of the M365 Copilot Analyst Agent on an SDTM→ADaM→TLF QC task: empty “FINAL” files, walked-back “submission-ready” claims, an unstable error rate, a fabricated appendix — and its own concession “you shouldn’t [use this].” Read fairly: one high-stakes QC session, not the ops-KB task — it shows why to keep anything you must reproduce off a non-pinnable cloud model.
Copilot trial-ops feasibility
The 75-automation re-assessment: where Copilot is genuinely feasible (19/75 fully, most buildable) and where it falls short — the gaps tracing to no-on-prem-model + no-model-freeze, exactly what the local path closes.