Learning ledger

A changed category might mean the first choice was wrong, or that new information arrived that nobody could have known at first. The log doesn't distinguish them. I'd sample 50 changed tickets and read them, to estimate what share were genuine errors before using the change rate…

Confidence is the field people are least comfortable with. Some will leave it blank or always write 50%. I'd offer five fixed options (50, 60, 70, 80, 90%) rather than free text, and see whether that helps. Once there are enough records…

Some methods overlap, especially the risk register and the risk map. I'm keeping both for now because they answer different questions, but I'd like feedback on whether the list is easier to use shorter.

One planned method, the Decision Diagnostic, is listed in this collection as coming. I haven't written it up in full yet, and I'd rather wait than publish a half-version.

Microsoft's documentation says scores "can change based on how you phrase or order questions and options" and suggests testing option order. I want to see how large that effect is on a real routing task before I'd trust a single threshold across wording changes.

I haven't used Autopilot or Code myself, because both are in early access. Everything above about them comes from Microsoft's own descriptions. I'll update this page when I've tried them.

Microsoft's message centre posts about this change (MC1479276 and MC1479277) say usage-based billing rolls out in Copilot during the fourth quarter of 2026, and that Code and Autopilot need a Copilot licence plus a spending policy. I've only read those posts through a…

I'd like to confirm how the new Home experience, which merges Chat and Cowork, changes what unlicensed Copilot Chat users see. Microsoft's pricing table currently shows Cowork as unavailable in Copilot Chat and metered for licensed users, but the experience is changing quickly.

I haven't confirmed how a single agent that mixes public web content and SharePoint content is billed when an answer only uses the public part. Microsoft's rates are per event, so I'd expect only grounding events to be charged, but I'd test it in a pilot environment before…

Time saved is the weakest number in the example, because people are poor at estimating it. I'd rather measure it on a few real tasks before and after than rely on a survey. I'm also watching how the September 2026 changes to Home and Cowork shift what the licence alone includes.

I haven't found a Microsoft report that shows how much of someone's Copilot use touched their organisational content compared with the web. That would answer my first question directly. Until I find one, I ask people.

I haven't tested how reliably a notebook says "that isn't in the references" when asked something outside them. I'd test that with a question I know the answer to before trusting it on a real decision.

I'm not sure how sharing works when the notebook owner later adds a new reference. My reading is that sharing is evaluated per linked file, but I haven't found Microsoft guidance on whether new references are automatically shared with existing members. I'd test it before relying…

I haven't checked how version history on a page compares with a Word document. For a decision brief that matters, I'd want to see who changed the recommendation and when. Until I've confirmed that, I'd convert the final version to a Word document or PDF, which Microsoft's…

I haven't tested how shared links from other people's pages behave during the retention period, before deletion. Microsoft's guidance says links stop working after deletion, but it isn't clear to me whether collaborators keep access during the retention period without a…

Microsoft notes that auto-sharing SharePoint files with an agent only works when sharing with specific security groups, so file permissions might need updating by hand. I'd like to test what a user sees when the agent can't open its own knowledge file for them.

The table shows what can be added. I haven't found Microsoft guidance on how answer quality changes as an agent approaches 100 files. I'd test it with real questions before assuming more sources means better answers.

Microsoft's pages don't agree on who can create agents. The SharePoint admin page says creating needs a Copilot licence, while the Microsoft 365 agents admin guide says people can create them if the tenant has pay-as-you-go for SharePoint or they have a licence. I'd test it in…

Microsoft's pay-as-you-go pricing page was last updated in April 2025 and describes "messages", while newer pages talk about Copilot Credits at the same $0.01. I believe they're the same unit renamed, but I haven't found a Microsoft page that says so directly. Check the meter…

I haven't tested what carries across when an Agent Builder agent is copied to Copilot Studio, such as embedded files and sharing settings. I'd check that on a test agent before moving a popular one.

I haven't found a simple way to see credits per conversation for a specific agent, which is the number I'd want to compare with the estimate. The environment-level view tells you the total. I'd run a short pilot with known conversation counts to back out the real figure.

I don't yet have a feel for how many credits a typical task like the weekly note uses. Microsoft links a Cowork cost estimator from its admin documentation, and I'd compare that estimate with /cost after a few weeks of real use.

Foundry has changed quickly, including a new portal and a move from hub-based projects. Microsoft says some platform capabilities, such as built-in monitoring dashboards, are in preview and points to a general availability overview for current status. I'd check that page before…

Model availability differs by region and deployment type, and Microsoft keeps a separate region availability page for each category. I haven't checked which current models are offered as regional deployments in UK regions. That's the first thing I'd look up for a real project.

The formula is my own, and I haven't tested it against how a group would rank the same rows by discussion alone. I'd like to try that comparison with a real team. If the ranking matches discussion most of the time, the formula is saving effort. If it often disagrees, the formula…

Partial reversibility is hard to score. An apology letter partly undoes a badly worded reply, but not entirely. For now I score those as 3 and write a note. I haven't found a cleaner way that doesn't add another column.

I'm unsure how to set the "error rate moves" trigger without enough data to know the normal range. For a new agent, I'd set a deliberately cautious limit for the first two months and widen it once the weekly sample shows what normal looks like.

Adaptive cards that wait for a response can be left unanswered. I haven't settled on a standard for what a flow should do if nobody responds within a set time: escalate, repeat or default to the safe option. My current view is to default to the safest choice and tell the owner's…

Microsoft's feature table is detailed but changes often, and a note on the same page says Agent 365 meets FedRAMP High controls with final authorisation pending, which suggests the service is still settling. I'd re-read the table before any licensing decision.

Microsoft now has two versions of DSPM: a classic DSPM for AI and a newer DSPM that it says replaces it. The oversharing assessments appear in both. I haven't confirmed which licences each requires, so I'd check the Purview service description before planning around them.

I haven't found a clean way to allocate seat licences to decision types. Asking people to estimate their split works, but it's an estimate, and I'd rather label it as one than pretend the number is more precise than it is. I'm also checking how Microsoft's new cost reports group…

I don't yet know how well a design-time estimate holds up for agents with open-ended conversations, where the number of questions per conversation varies a lot. I'd like a month of real logs from a pilot before trusting any range for those.

For Copilot experiences billed through the Microsoft 365 admin center, such as Cowork, I haven't yet confirmed what forecasting the new Cost Management dashboard offers beyond trends and limits. Until I've seen it, I'd build the forecast in a spreadsheet from volumes and compare…

Prices here are list prices checked on 11 October 2026. Enterprise agreements set their own prices, so your figure may differ, and exchange rates move. I also want to confirm how the 12-credit pattern holds for agents with several grounding calls per answer, since that would…

I haven't checked how Microsoft's annual subscription terms treat reassigning a licence mid-term between people. My understanding is that a licence can be moved from one user to another within the tenant, but the number of licences is fixed for the term. I'd confirm that against…

Learning value is the score I'm least sure how to judge in advance. Some topics I expect to be easy turn out to teach me the most, because explaining them exposes a gap. I may start scoring it after writing instead, as a way of calibrating my guesses.

I don't yet have a good way to hear which gaps readers notice. For now I go by what's empty or thin. A simple feedback route, with no sign-up and no sales, would help, and it's something I'm considering for after the current round of articles.

I haven't decided what to measure each quarter beyond delivery against plan. I don't want traffic numbers to drive the writing, and the site carries no offers until 2027, so there are no sales figures either. Something like "questions answered" might be better, if I can find an…

I haven't confirmed how reliably roadmap IDs appear in Message center posts. Many posts reference them, but I don't know if all do. Until I'm sure, I search the Message center by feature name as well as ID.

I don't yet know how well a general model's stated probabilities match reality for this kind of task. That's the biggest open question, and the test set exists to answer it. Microsoft says some Foundry monitoring dashboards are in preview, so I plan to keep my own log as well…