Ops Log topic
AI & Automation field notes
Reviewed AI & Automation field notes by Michal Jatczak. Each public note includes its evidence, test context and a check you can run.
7 reviewed notes
A Graph meeting export can return an empty page that still carries a next link
During a planned Microsoft Graph service update, paginated calls to getAllRecordings or getAllTranscripts can return HTTP 200 with an empty collection alongside a next link, then restart and re-serve items. An export loop that exits on no items stops there and reports success over a short archive.
Anthropic in Excel and PowerPoint is a separate setting, on by default for EU tenants created after 25 March 2026
Copilot in Word, Excel and PowerPoint has its own Anthropic setting, a different object from the global Anthropic subprocessor control, and it defaults to on for EU, EFTA and UK tenants created after 25 March 2026. Whether anything actually leaves the EU Data Boundary turns on the second control, so read the tenant creation date and then read both.
Shadow AI is a governance-readiness problem, not a deadline
The first artifact three separate AI frameworks all demand is the same: a register of the AI systems in use, their data posture, and an owner. Build that register now for the governance value, and let the shifting AI Act dates be a secondary driver.
Onboarding automation: the role profile, the joiner event, and the five parts that break
An operator runbook for automated joiner provisioning in Microsoft Entra ID: the role-profile file, a normalised joiner event, a poll-for-readiness Graph sequence, the Temporary Access Pass trap that locks out day-one starters, and the formula to compute your own time saving.
Self-hosted LLM serving in the EU: runtime choice, GPU sizing, and the sovereignty argument
A runbook for deciding whether to self-host an open-weights model in the EU: the runtime field after TGI entered maintenance mode, GPU sizing computed from the model config instead of copied from a table, a corrected token-throughput cost model, and the DORA, NIS2 and AI Act record you will be asked for.
Retrieval over a permissioned corpus: ACL handling, the update pipeline, and what to measure
A bench specification for retrieval over a corpus that carries permissions: the four access-control failure modes, why post-filtering starves an approximate index, where the cache breakpoint belongs, and the six numbers to record on every run.
Choosing a workflow engine: what a self-hosted n8n actually costs to run
A decision memo comparing Power Automate licence-and-request-ceiling pricing against the real running cost of a self-hosted n8n, with vendor list prices checked on 2026-07-28, the figures from one operated instance, and a break-even script you fill in with your own flow counts and operator hours.