What Is Copilot and How Does It Work?

Microsoft 365 Copilot added 5 million paid seats in one quarter, reaching 20 million by Microsoft’s fiscal third quarter of 2026. However, those licenses covered less than 4.5% of Microsoft’s more than 450 million paid Microsoft 365 seats. Microsoft’s September 25 redesign addresses that gap by combining chat, editable Office documents, app building, and persistent agents into a single Copilot application.

Key Takeaways:

• The redesigned Microsoft Copilot app has three main areas: Home, Code, and Autopilot.
• Home combines Chat, Cowork, and editable Word, Excel, and PowerPoint content in one workspace.
• Code creates tenant-hosted apps, trackers, dashboards, and automations from natural-language instructions.
• Autopilot is a persistent cloud agent with its own identity, memory, computer, and workspace.
• Chat and Office features use a fixed user subscription license, while Cowork, Code, and Autopilot use usage-based billing.
• Enterprises need accurate permissions, usable business data, spending limits, and measurable task outcomes before expanding deployment.

The Copilot Super App Consolidates Work

Microsoft introduced the redesigned app on September 25, 2026, combining conversational assistance, delegated work, software creation, and long-running automation within one interface. Microsoft describes it as “AI built for work,” designed for organizations using Microsoft 365, Outlook, Teams, SharePoint, OneDrive, Dynamics 365, and Power Platform.

The redesign addresses a fragmented lineup: consumer Copilot, Microsoft 365 Copilot, Office assistants, and Cowork previously appeared across separate interfaces. ZDNET called the earlier lineup fragmented, while The Verge reported the redesigned client now combines chat, coding, and agents into one application.

Microsoft aims for the new Copilot to have as much impact on AI-mediated work as Office had on PC productivity. The immediate advantage is simpler access: an employee can start with a question, move into an editable document, delegate a larger task, or build a small internal tool without switching products.

Enterprise Adoption and the Usage Gap

Microsoft 365 Copilot reached 20 million paid seats in fiscal Q3 2026, up from 15 million one quarter earlier, 5 million paying seats added in about 90 days, according to GCN. Customers with more than 50,000 seats reportedly quadrupled year over year.

Accenture is the largest announced deployment: about 743,000 employees in April 2026, starting with several hundred senior leaders in 2023, expanding to roughly 20,000 users, followed by phased training, data-governance work, and role-specific adoption programs.

Scale alone does not guarantee routine use. A survey covering about 200,000 Accenture users recorded 89% monthly active usage, while 84% said they would deeply miss the product if it disappeared. Microsoft also reported 97% of participating employees completed some routine tasks up to 15 times faster. These are company-reported outcomes, not an independent controlled study, so buyers should test the same workflows internally.

The broader installed-base figures are less encouraging. Twenty million Copilot seats represent less than 4.5% of Microsoft’s more than 450 million paid Microsoft 365 seats. Reporting summarized by eTeknix estimated only 20% to 30% of the paying group used it weekly, about 1% of the entire Microsoft 365 paid base.

Home, Code, and Autopilot

The redesign organizes work into three areas, each with different authority, runtime, and financial risk.

Home combines quick conversation with Cowork, the mode for delegated, multi-step work, preparing an RFP response, customer briefing, launch kit, or financial-close package. Office in Copilot adds live, editable documents instead of plain text that must be copied into Word or rebuilt as a spreadsheet.

Code uses the same underlying technology as GitHub Copilot, according to Microsoft, but targets knowledge workers creating small internal applications. Generated software runs in a sandbox and can be hosted through Copilot Managed Runtime inside the customer’s tenant. Professional developers still need to review generated logic, access controls, dependencies, and data handling before operational use.

Autopilot has the broadest authority: it can continue after a user closes a laptop, watch Teams channels, handle follow-ups, and restart work days later. Its persistent identity and audit trail make it easier to govern than an agent operating through a shared employee account, but they turn agent permissions into a significant security decision, a persistent agent can repeat a mistaken action until a policy, budget, or human approval stops it.

Architecture and Enterprise Data

The redesigned application operates above several Microsoft context and execution layers. Microsoft IQ provides organizational context. Work IQ connects Copilot to work data, while Fabric IQ adds semantic context from Power BI. Fabric IQ brings more than 20 million Power BI semantic models into Chat and Cowork. Dynamics 365 and Power Platform grounding was scheduled for public preview during the month after launch.

This makes data readiness a deployment requirement. The assistant can only retrieve information users and agents are allowed to access. Excessive SharePoint permissions can expose material to the wrong employee, while fragmented files, weak metadata, duplicate documents, and outdated records reduce answer quality. A rollout should include permission review, information ownership, retention rules, and tests using actual employee roles.

Microsoft is also creating a unified plugin registry for Microsoft, partner, and custom plugins. Central approval can reduce uncontrolled extensions, but each plugin adds another path to business data or actions. Administrators need to review the plugin’s permissions, authentication model, logging, and failure behavior rather than relying on the Copilot interface as the security boundary.

Business Value and ROI Measurement

The business case is strongest when Copilot changes a repeated workflow with a measurable output. Accenture’s Avanade joint venture built an internal system that aggregates 8-K and 10-K filings for sales teams; Accenture reported the tool helped generate 43% more sales opportunities. Marketing teams used Copilot to compare new material with previous campaigns and draft storyboards, while other employees began building agents without traditional coding skills.

NHS England provided access for 505,000 clinicians and support staff in June 2026, testing Copilot where time savings have value but inaccurate summaries or inappropriate access carry serious consequences. Healthcare organizations still require qualified review, approved data boundaries, and clear rules for tasks the assistant must not perform.

Microsoft added ROI tracking to Foundry in July 2026, linking agent costs with task completion, time savings, and business value, allowing teams to compare agent versions and inspect traces behind weak outcomes. A useful internal scorecard should track completed tasks, human correction time, failure rate, total agent spend, and cost per accepted result.

Licensing and Usage-Based Costs

Microsoft separates everyday assistance from long-running agent work. The user subscription license covers Chat and Copilot inside Word, Excel, PowerPoint, Outlook, and Teams. Microsoft’s Auto router balances accuracy, speed, and cost when selecting a model. Cowork, Code, Autopilot, and frontier models such as Astra and Fable use usage-based billing.

TechTarget reported Microsoft 365 Copilot enterprise pricing at $30 per user per month, billed annually, in April 2026, with a small-business range of $18 to $21 per user per month. These charges are in addition to qualifying Microsoft 365 subscriptions. Usage-based agent costs require separate controls and cannot be estimated from seat price alone.

The following Python utility combines the fixed license expense with metered agent charges exported from a billing system. It requires the operator to provide actual usage rather than estimating a per-agent cost.

Note: The following code is an illustrative example and has not been verified against official documentation. Please refer to the official docs for production-ready code.

#!/usr/bin/env python3
"""
copilot_monthly_cost.py

Example:
 python3 copilot_monthly_cost.py \
 --licensed-users 500 \
 --license-price 30 \
 --metered-agent-cost 1840.75
"""

import argparse

parser = argparse.ArgumentParser(
 description="Calculate fixed and metered Copilot monthly cost."
)
parser.add_argument("--licensed-users", type=int, required=True)
parser.add_argument("--license-price", type=float, required=True)
parser.add_argument("--metered-agent-cost", type=float, required=True)
args = parser.parse_args()

if args.licensed_users < 0:
 raise ValueError("licensed-users must be non-negative")
if args.license_price < 0 or args.metered_agent_cost < 0:
 raise ValueError("cost inputs must be non-negative")

fixed_cost = args.licensed_users * args.license_price
total_cost = fixed_cost + args.metered_agent_cost
cost_per_licensed_user = (
 total_cost / args.licensed_users
 if args.licensed_users
 else 0.0
)

print(f"fixed_license_cost=${fixed_cost:,.2f}")
print(f"metered_agent_cost=${args.metered_agent_cost:,.2f}")
print(f"total_monthly_cost=${total_cost:,.2f}")
print(f"cost_per_licensed_user=${cost_per_licensed_user:,.2f}")

# Note: production use should import billing records through an
# approved API, distinguish taxes and negotiated discounts, validate
# reporting periods, and alert when department budgets are exceeded.

A finance team can run this calculation by department and pair it with accepted task counts from Foundry. The relevant metric is cost per completed, reviewed outcome; raw token use or agent hours say little about value if employees repeatedly correct the output.

The billing change follows GitHub Copilot’s June 2026 move to usage-based AI credits. Our 2026 GitHub guide explains how long agent sessions generate many model calls behind one user request. The same principle applies: enable metered functions for a limited group, impose spending limits, and compare charges with completed work before expanding access.

Copilot, Gemini, and Anthropic

Microsoft’s advantage lies in distribution and integration: an organization using Microsoft 365 can keep documents, identity, collaboration, agents, and generated applications within a familiar administrative stack. Its weakness includes repeated redesigns and repositioning, while consumer adoption trails ChatGPT and Claude.

Google Gemini is the natural alternative for organizations focused on Google productivity tools. A Computerworld writer who used Copilot since its 2023 public release reported switching to Gemini, showing that Office integration does not guarantee a better experience for every user. Enterprises should compare products with the same documents, permission structures, and task rubrics rather than assuming suite ownership ensures output quality.

Anthropic offers a different model through Claude Code and its agent products. Claude Code runs locally in isolated environments such as containers or virtual machines, while Microsoft Code is designed around a cloud sandbox inside the Microsoft 365 tenant. Microsoft emphasizes centralized governance and integration; Anthropic’s approach appeals to developers who prefer terminal-centered work and local control. The better option depends on data location, developer workflow, administrative policy, and total task cost.

Limitations and Deployment Checklist

The all-in-one Copilot reduces interface fragmentation, but it does not fix inaccurate source data, excessive permissions, weak process ownership, or vague task definitions. Generated documents can contain incorrect claims. Generated software can pass a superficial review while mishandling authorization or edge cases. Persistent agents can repeat errors at machine speed.

Start with bounded workflows and measurable acceptance criteria:

• Review access first: Inspect SharePoint, OneDrive, Teams, Outlook, Dynamics 365, and Power Platform permissions before connecting agents.
• Choose repeatable tasks: Begin with document summaries, controlled report preparation, meeting follow-ups, or internal tools with deterministic checks.
• Separate creation from approval: Let Copilot draft documents or create branches, but keep financial approval, publishing, merging, and high-impact actions under human review.
• Set spending policies: Leave usage-based functions disabled until owners, limits, alerts, and reporting periods are defined.
• Measure accepted outcomes: Record completion time, correction time, failures, agent cost, and user adoption by workflow.
• Test alternatives: Compare Copilot with Gemini, Claude, a conventional automation, or a simpler search and rules system where the task allows.

Microsoft’s 2026 redesign gives Copilot a clearer structure and a stronger enterprise proposition. Home turns conversation into editable Office work, Code creates tenant-hosted tools, and Autopilot continues recurring tasks without a live user session. The opportunity is substantial, but the adoption figures show that access alone does not create habitual use. The next phase will depend on operational evidence: clean data, controlled permissions, accepted outputs, reduced task time, and a bill justified by the completed work.

Related Reading

More in-depth coverage from this blog on closely related topics:

• Understanding the Avast Sandbox
• F-Droid 2.0 App Store Update and Features
• AI Market Cap Comparison
• What Is Vision Pro and Its Features
• AI Inference Cost Trends and Economics

Sources and References

Sources cited while researching and writing this article:

• Microsoft thinks its new Copilot ‘super app’ will be as influential as Office
• Microsoft 365 Copilot added 5 million paying enterprise seats in a single quarter, and the gap buried in Microsoft’s own numbers tells the real story of who is winning the …
• Microsoft 365 Copilot Is Still Below 4.5% Adoption
• Introducing the new Copilot with Home, Code and Autopilot
• Microsoft’s new Copilot app puts everything in one place – but the price is ‘evolving’