Work Trend Index

This page is for Copilot programme leads and sponsors who want to use Microsoft's 2026 Work Trend Index as a guide. It maps the report's questions (is Copilot becoming a habit, are agents spreading into useful work, are managers taking part, and what should an enablement programme do next?) to the figures Microsoft 365 Advanced Analytics can supply for your own tenant. It also says plainly where tenant data can't reproduce Microsoft's research.

It is a guide, not a tenant edition of Microsoft's report, a readiness index, or a productivity score, and you don't need to read any code to use it. For using the page itself, see the Copilot Adoption User Guide; for every formula, the metrics reference.

On this page: From Microsoft's research to your tenant · Choose the question before the chart · Before you quote a figure · Questions the product cannot answer · Sources

From Microsoft's research to your tenant

Microsoft's 2026 Work Trend Index Annual Report: Agents, human agency, and the opportunity for every organization, published on 5 May 2026, discusses three connected ideas: AI can expand individual capability; leaders must redesign how work happens; and organisations need shared learning practices, not just individual experimentation.

Those ideas come from different kinds of evidence. Keep them separate when discussing your own organisation:

Research evidence	What its denominator and method mean	What this solution can supply instead
Microsoft's aggregated, anonymised Microsoft 365 telemetry	Agent-growth telemetry across Microsoft 365 Copilot Agents and SharePoint, March 2025–March 2026. The monthly-active-agent definition includes an agent with at least one day of user-initiated use or at least one autonomous run in a rolling 28-day period; other research measures have their own population filters	Imported tenant audit observations, with the narrower origin/coverage rules below. Separate billing evidence cannot complete or be added to that definition
The Work Trend Index survey	Self-reported experiences from 20,000 AI-using knowledge workers in ten markets, fielded February–April 2026. People answering “never” or “don't know” about AI use were screened out. Survey findings describe those respondents, not all tenant accounts	Usage and organisational metadata. No imported answers about confidence, training, incentives, human agency or work quality
The report's work-activity analysis	Intermediate Work Activities mapped to O*NET Generalized Work Activities, using fractional shares when one conversation contains several activities. The methodology describes a one-week North American commercial sample, excluding education	Optional, single-category prompt classification against a local administrator taxonomy. It is not that O*NET analysis

The report's 15× year-on-year growth in active agents describes Microsoft's customer base. It is not a growth requirement for one tenant, a benchmark applied by this product, or evidence that an organisation below it is failing.

The report also presents an illustrative four-mode framework: asking, delegation, exploration and collaboration, with human and agent involvement considered independently. It is not a ready-made scoring taxonomy. The solution's optional Directing/Supervising annotations do not calculate those modes.

Decide who your local question is about. “Knowledge workers using AI” is not interchangeable with every Entra account, every Microsoft 365 licence holder or every Copilot seat holder. Agree an appropriate departmental/pilot scope, document exclusions and keep the actual report denominator visible. Telemetry cannot identify the Work Trend Index's survey population or “Frontier Professionals” for you.

Choose the question before the chart

Start with a question you can act on. Each measurement's source, unit, window and limits are in the metrics reference, and Copilot Adoption Setup lists which import each answer needs.

Local question inspired by the report	Where the operator goes	Action the evidence supports	Do not conclude
Is everyday AI use becoming a repeatable habit?	Insights → Copilot → Copilot Adoption → Executive / Analyst view: adoption funnel, engagement mix, frequency/intensity and enablement plan	Separate onboarding, re-engagement, second-app scenarios and advocate recruitment	A high score proves skill, good judgement or productivity
Where is demand outside our paid-seat population?	Unlicensed usage and Licence opportunities	Review frequent Copilot Chat use alongside collaboration/document workload before reallocating seats	Every unlicensed user needs a paid seat, or the candidate score is measured ROI
Are agents spreading, or concentrated in a few teams?	Analyst view department agent columns and Agents reach/growth	Distinguish reach from repeated use; find workflows worth enabling elsewhere	Every agent exists in the usage inventory, or raw audit rows equal turns
Are people building reusable workflows?	Agent builders beside department agent use	Identify departments with create/publish/share activity and discuss enablement support	The counted builder owns a particular usage or billing agent
Are managers and leaders modelling use?	Manager comparison and Leadership group compared with the tenant	Compare like-for-like licensed/unlicensed manager groups; discuss support with teams	Manager use caused adoption, or the result measures employees' perceptions of support
What kinds of work are people asking AI to do?	Optional prompt categories on Insights → Copilot → Prompt insights	Review the reported sample's goals, taxonomy version and failed/capped classifications	The mix reproduces Microsoft's work-activity percentages or measures actual delegation
What safety and grounding evidence accompanies use?	Insights → DLP impact, including governance signals	Review reported flags, label/resource use and policy matches through approved governance processes	A rate is a security-effectiveness rating, or an event proves protected content leaked
Did an enablement intervention coincide with change?	Excel report, matched settings/periods/scopes and an approved external survey	Keep a baseline, compare comparable facts, and investigate changed usage and experience together	A before/after increase proves causality, profit or better work

Before you quote a figure

The detailed rules live on the pages linked below. These are the cautions that matter most when you put one of your figures next to a Work Trend Index finding.

Is use becoming a habit?

• The adoption and habit rates divide by the people actually scored, not by the seats you bought. On a very large tenant a cap can make them describe a subset, so keep the scored count and any cap warning beside the rate. See headline rates.
• Habitual means a weighted engagement score at or above Established, not just frequent use: someone can open Copilot every day and still not be habitual. See engagement bands.
• Dormant and Never used mean no use was observed in the history the solution holds. They aren't proof of no use, or of no business value.

Are agents becoming part of work?

• Only verified customer-built agents count towards agent breadth, depth and growth. Agents of unknown origin are counted separately, so the figures are a floor. See agent breadth, depth, builders and reach.
• A turn that Microsoft audits twice counts once, and Copilot Studio maker testing is left out. See how turns are counted.
• Agent growth compares your organisation with itself, over 14 closed 28-day windows. The report's 15× describes Microsoft's whole customer base, not a target for you. Never add the audit series and the billed-agent series together.
• Agent builders counts the people who created, published or shared a Copilot Studio agent. They aren't matched to the agents they built.

Are managers and leaders participating?

• The manager comparison is an association, not a cause: teams whose manager uses Copilot can differ in role, seniority and seat coverage. Managers whose own use is unknown are never counted as inactive. See manager modelling.
• The leadership card compares a group an administrator chose with the whole scored tenant, leaders included. It needs at least ten licensed leaders and, unless the group includes everyone, ten licensed people outside it; See PII can't override that.
• Neither tells you whether employees feel supported by their manager. That's a survey question; see pairing usage with an employee survey.

What work are people asking AI to do?

• Optional prompt categories classify a sample: the first eligible new prompts in each import cycle, up to a cap. It is neither a census nor a random sample, and a prompt can't be reclassified later because its text isn't kept. Groups of fewer than ten people are withheld.
• The categories are your administrator's taxonomy, not the report's O*NET analysis. Directing and Supervising are labels an administrator attaches to a category, not something the model infers, and they don't calculate the report's four modes of working.

What governance signals accompany that use?

• On Insights → DLP impact, a missing flag means "not reported", not "clean": each rate is over the interactions where Microsoft reported the flag. See Copilot governance signals.
• A flag or a policy match isn't proof that protected content leaked, or a measure of how well your safeguards work.

What can we say about time and productivity?

• The three time-saved figures are modelled ranges of hours with visible assumptions: never measured time, money or return on investment, and never added together. The low end is an assumption (50% of the high end by default), not a statistical confidence interval. See time saved and the licence and Cowork estimates.
• A surface credited at 0 minutes, such as Copilot Chat, is uncredited by the model, not worthless.
• To show change over time, keep a baseline export and compare like with like: see Comparing two exports. To understand people's experience, pair usage with an approved survey.

Questions the product cannot answer

The evidence does not establish:

• Individual AI skill, confidence, human agency, critical thinking, training effectiveness, psychological safety or employees' feelings about managers.
• The Work Trend Index's organisational/individual readiness zones, survey impact scores or which employees are Frontier Professionals.
• That usage, manager participation or an intervention caused productivity, work quality, revenue, profit or business returns.
• Exact autonomous-run totals, a complete agent registry from audit activity alone, or verified authoring/audit/billing ownership joins.
• A security-effectiveness score, guaranteed safe output or proof of protected-content leakage from a flag/policy event.

These boundaries also apply to proposals:

Proposal	Current boundary
#649 — behavioural Transformation Paradox/readiness	Not shipped as a readiness quadrant. This guide makes no product decision about it
#652 — O*NET-specific intent methodology	Formal/fractional mapping and verified accuracy are not delivered by the local categories
#683 / #684 — tuning	33 rules are editable on Settings. Activation-window reporting isn't available, and history windows and row/performance caps are fixed

Use the report as a set of research questions and guidance. Do not claim a survey replica, a 15× tenant target or a match to publisher-global percentages.

Sources

Official research and API contracts, checked 11 October 2026:

• 2026 Work Trend Index report and methodology: global telemetry versus self-reported survey, agent definition, O*NET fractional sample and conceptual modes.
• Graph Copilot user-count summary: permission, D28/D30 versions, prompt fields and available surfaces.
• Purview Copilot audit schema: resource, agent, message, model and policy evidence.
• Agent 365 package list: catalogue licence, application permission and API scope.
• Foundry data privacy: processing boundaries, filtering and service-side data handling.
• Microsoft Copilot assisted-hours methodology: context for a time-credit model, not measured tenant savings.

---

Related: Copilot Adoption User Guide · Interpretation · Metrics reference · Copilot analytics