Kesslernity · M365 Copilot Field Guide
Microsoft 365 People Skills
A skills profile you did not type.
People Skills maps every user to a 16,000+ skill library and fills their profile three ways: skills they confirm, skills the organization imports, and skills an AI infers from documents, emails, chats and meetings. We ran it in a purpose-built test tenant from activation to org-wide deletion and measured what actually happens. Every claim below carries its source.
Licensing, the part everyone gets wrong
No separate license. Microsoft 365 Copilot users get everything, with AI inference on by default. Viva Suite / Insights / Learning users get inference but no Copilot experiences. Microsoft 365 E3/E5 without Copilot ("Foundation") get manual profiles only; inference is off by default and needs an all-or-nothing admin opt-in. Office 365 E3/E5 is not eligible; Microsoft 365 E3/E5 only. Education and government are excluded.
At a glance
What it does
Builds a per-user skills profile mapped to a library of 16,000+ skills (16,287 displayed in our tenant), powered by LinkedIn, and feeds it to the profile card, people search, Org Explorer, Copilot people queries and Viva.
How skills arrive
Three ways: you confirm them, your org imports them from HR systems, or an AI infers them from your documents, emails, chats and meetings. Inference refreshes about every 30 days and ignores activity older than 90.
The defaults
For Copilot-licensed users, inference ships on. Sharing ships on too: Microsoft's own deck says all skills, including AI-inferred ones, are shared with all M365 experiences by default.
The new exit
An org-wide delete switch went GA the week of 25 August 2026: one control permanently deletes the library and everyone's confirmed, inferred and imported skills. Most data gone in 48–72 hours.
01 · What it is
A data layer, not an app
People Skills is an AI-driven service that generates personalized skill profiles mapped to a customizable, built-in taxonomy, in Microsoft's words. You never open "People Skills"; you meet its output wherever people look each other up: the Microsoft 365 profile card (new Outlook, Outlook web, Office.com, SharePoint, Copilot, People Companion; Teams "coming soon"), people search, Org Explorer, Copilot people queries, Viva Learning recommendations, and the Copilot Analytics skills-landscape report. Two agents (Learning Agent, Workforce Insights Agent) sit in the Frontier preview program.
Confirmed
You add or endorse skills in your profile editor. The only kind Microsoft documents as exportable via the Graph Profile endpoint.
Imported
Ingested from HR or third-party systems via the org data ingestion tool. Labelled "Imported by your organization", but only in the editor; the profile card does not visually distinguish imported from inferred.
Inferred
Assigned by the AI engine from your work signals. Refreshed ~30 days (Copilot path), first arrival documented at 48 hours to 5 days.
The engine
Microsoft's page names the stack: Graph activity signals plus job title, processed by "advanced OpenAI LLM models" with skills matched to the closest taxonomy entry. Processing stays inside the tenant boundary.
02 · What we measured
Four days in a live tenant, activation to deletion
Purpose-built 3-user Copilot-licensed test tenant, 28 August to 1 September 2026. One operator, thin signal, stated as such: these are field observations from one tenant, not universal product behaviour.
Day 0
Activation is one click, with no confirmation
The admin quick setup reported People Skills active in seconds: no confirmation dialog, no defaults summary. The Learn-documented admin path (Copilot → Settings → Data access) did not exist in our tenant; the working path was Settings → Viva → Data management.
Day 1
A skills editor appeared, and it was the wrong one
~26 hours in, a Skills field surfaced in the profile editor's About section, accepting free text with no library lookup. It was the legacy UPA skills field. Learn confirms the mechanism: legacy skills are hidden from the card and editor "when you deploy People Skills", and our Graph export later showed those legacy saves persisting invisibly under source "UPA".
Day 3
The real editor, and the first inferred skill
The library-driven Skills section arrived (closed vocabulary, library entries only: "FinOps" is not among the 16,287 skills; the closest is "IT Cost Optimization"). The first AI-inferred skill, "Cloud Administration", carries a Graph timestamp 69 hours after activation, inside the documented window of 48 hours to 5 days. Our seeded signal was dense in Terraform and AKS, both of which ARE library entries; the engine generalized instead, which matches Learn's own description of closest-match tagging.
Day 4
The delete switch, run live
Microsoft's flyout lists what deletion removes: the library, and everyone's confirmed, inferred and imported skills, permanently. The admin surface flipped to "We're deleting People Skills from your organization" within seconds. Minutes later, the user editor still showed all four skills and a Graph export still returned every record, consistent with the documented 48 to 72 hour pace. Deletion is a process, not a click.
03 · Findings your DPO will ask about
Where the docs and the product disagreed
All of the following were measured in one tenant on a feature that moves quickly; treat them as field reports to verify in your own tenant, not permanent product truths.
1
The inferred skill is indistinguishable from confirmed ones in the editor list. Same pill, same icon, no label. Learn separately confirms the profile card does not differentiate imported from inferred. The three-flavour taxonomy is real in the data model and invisible to the person it describes.
2
The Graph export returned the inferred skill. Learn states twice that only confirmed skills can be exported and AI-generated skills cannot. In our tenant, the beta Profile skills endpoint returned the AI-inferred skill alongside hand-confirmed ones, with identical provenance ("SkillsInViva") and no inference marker. Either the export boundary does not hold, or the pipeline records inferred skills as confirmed. Both readings undercut the documented boundary; verify in your tenant before relying on it.
3
The user inference opt-out was missing. Learn and Microsoft's deck both describe a user-level opt-out of AI-generated skills in the profile editor settings. Our tenant's Data-and-privacy settings contained exactly two sharing toggles (both defaulted on) and no inference opt-out at all.
4
The AI-restricted skills control was missing. Learn documents marking sensitive skills as AI-restricted with 24-hour removal. No such control existed anywhere in our tenant's admin UI.
5
Copilot Chat never saw any of it. Through day 4, Copilot answered skills questions with no access to the skills store, improvising from other sources, including presenting a Copilot Memory entry as a "confirmed skill". Early Copilot answers about skills are improvisation, not skills data.
6
One editor quirk to know: in our tenant only the first library search per editor session returned results; every later query said "No results found" until the dialog was closed and reopened.
04 · The controls
Who can turn what off
Users
Add, reorder and delete skills in the profile editor's Skills section. Two sharing toggles under Data and privacy: "Share your skills on your profile" and "Share imported skills automatically". Both ship on. The documented inference opt-out was absent from our tenant (finding 3 above).
Admins
Defaults are set through Feature Access Management access policies, tenant-wide or targeted at users, groups or regions: skills profile visibility, People Skills user access, and the E3/E5 inference opt-in (all-or-nothing, up to 15 days to first skills, ~180-day refresh). Plus the org-wide deletion control below.
The delete switch (GA, week of 25 August 2026)
Settings → Viva → Data management → People Skills → Settings → People Skills removal and deletion. Microsoft's own warning: "Removing People Skills is permanent." Deleted data is not restored on re-setup. Most data is deleted within 48–72 hours; Viva Learning skills data can take up to a week, though its skills experience turns off immediately. Microsoft recommends waiting at least a week before setting People Skills up again.
05 · The habit
Open yours this week
1
Users
Open your Microsoft 365 profile, then Update your profile → Skills. Read what is listed. If your tenant runs Copilot, something may be there that you never added, and nothing in the list will tell you which entries the AI wrote.
2
Admins
Know your posture before the questions land: which license tiers you hold, whether inference is running, what the sharing defaults are, and where the deletion control lives. If you deployed over legacy Delve/UPA skills, know that the old data persists in Graph even though the UI hides it.
3
Governance
Microsoft's responsible-AI page says People Skills is not for behavioural monitoring, performance management or employment decisions. Test the export boundary in your own tenant (finding 2), and treat "inferred skills cannot leave" as a claim to verify, not a control to rely on.
Watch the four days compressed to three minutes: the episode "Microsoft 365 Is Building a Skills Profile of You" on the Kesslernity YouTube channel, with every capture from this guide on screen.