Microsoft's new Copilot app, metered agent credits

• FACT (CONFIRMED — primary blog): On Sept 25 Microsoft introduced "the new Copilot": one redesigned app organizing AI work into three surfaces — Home (Copilot Chat + Cowork together as the default landing experience, plus Office in Copilot, which embeds the full Word/Excel/PowerPoint editors inside Copilot), Code (natural-language building of apps, trackers, dashboards and automations, "powered by the same underlying technology as GitHub Copilot," sandboxed, hostable inside the customer tenant), and Autopilot (a persistent, proactive, personal agent "that keeps working even when you're not" — previously called Scout — with its own identity, memory, computer and workspace in the tenant, addressable by @mention in Teams, Outlook, chats, channels and documents).
• FACT (CONFIRMED — primary blog; pricing blog; Learn docs): Microsoft restructured Copilot pricing into two tiers: the user subscription license (USL) — fixed per-user-per-month price covering everyday AI (Chat; Copilot in Word/Excel/PowerPoint/Outlook/Teams; model selection; Auto routing that "weighs accuracy, speed and cost on each request"; GPT-5.6 and Sonnet 5 included, Opus 5 to be included with limits, fair-use limits with warning-then-switch-to-Auto-or-credits behavior) — and usage-based billing (UBB) in Copilot Credits for agentic work: Cowork, Code and Autopilot, advanced in-app experiences (e.g., SharePoint) and frontier models ("Astra and Fable"). UBB builds on and requires the USL. Both tiers run models from OpenAI and Anthropic today; "more labs and open-weight models coming."
• FACT (CONFIRMED — pricing blog + Learn docs + June GA blog): Billing controls are the gating design: for enterprise tenants UBB services stay off and nothing is billed until an admin creates a spending policy in the Microsoft 365 admin center (org/group/user-level budgets, alerts, credit-request routing); Copilot Credit pay-as-you-go is priced at $0.01/credit (stated in the June 16 Cowork GA blog; Learn documents P3 pre-purchase and prepaid capacity packs as alternatives); per-task cost is computed from four inputs — model use, context retrieval, tool calls, runtime.
• FACT (CONFIRMED — primary blog + techcommunity/new FinOps post): New FinOps for AI capabilities: Agent 365 cost management expands beyond Cowork and Work IQ APIs to cover Code and Copilot Managed Runtime (Copilot Studio agents planned for October); admins can restrict model families per group — which also constrains what Auto may select; Cowork consumption insights (which tasks drive credits and what they produced); Microsoft Graph APIs for policy management; end-user visibility of credit usage/balance/history inside Copilot.
• FACT (CONFIRMED — Managed Runtime blog, David Blyth, Sept 25): Microsoft Copilot Managed Runtime introduced in public preview: Microsoft-operated hosting for AI-generated code inside the Microsoft 365 tenant boundary — Entra identity, org policies on connectors/data/endpoints, Git-backed versioning, lifecycle/version controls, central inventory in the M365 admin center, an SDK+CLI for the full dev lifecycle (typed connector services in TypeScript), and openness to third-party/pro-code tools (testimonial from Lovable: "an app made with Lovable can now run inside your Microsoft tenant… same sign-in, same policies, same app inventory"). It already powers apps built in Cowork, Code and Copilot Studio. (The main blog says "in preview"; the dedicated blog says "public preview.")
• FACT (CONFIRMED — primary blog): Context expansion: Microsoft IQ named as the unified grounding platform; Fabric IQ (incl. 20+ million Power BI semantic models) GA today in Chat and Cowork, Code via Frontier soon; Dynamics 365 / Power Platform data and workflows grounding rolling out in public preview over the next month; a new plugin registry (unified Microsoft/partner/custom catalog, IT approval, publish-once) rolling out with GA across surfaces "in the coming weeks."
• FACT (CONFIRMED — primary blog): Sneak peeks with dates: intent-based routing across Chat/Cowork/Code ("coming soon"); Today (proactive command center) private preview in Copilot in October, later Outlook/Teams; @Copilot in Teams with shared channel context/permissions in private preview by end of the month. Rollout: Home + Code start rolling out in the Frontier program in the coming weeks; Code to Frontier end of month, broad availability "in the coming weeks," preview for Microsoft 365 Premium and Pro subscribers later this year; Autopilot expands to private preview at end of the month. Ignite: Nov 17–20, San Francisco.
• INDEPENDENTLY VERIFIED (Reuters, same-day interview): Each Autopilot "will carry its own identity in the company directory and have specific permissions that users can control" (CVP Annie Dix-Pearl); Reuters frames the cost-management move against enterprise fear of "rapid AI usage racking up unexpected costs." MSFT shares rose about 3% in early trading Sept 25.
• REPORTED/CONFIRMED-LINEAGE (GeekWire, Reworked, thelettertwo, from the Sept 23 briefing + primary posts): Nadella promised the super app at Build in June ("come summer… coding to all knowledge work within one Copilot super app") and on the July 29 earnings call said "this quarter" — delivery lands with under a week left in the quarter, and (GeekWire) "most of the new features won't be widely available until later." Autopilot's lineage: OpenClaw — the open-source personal-agent framework that went viral in early 2026; Nadella called it a security risk akin to "a virus" in March (told Morgan Stanley he couldn't launch it himself), Scout debuted at Build "powered by OpenClaw" per Nadella in July; OpenClaw founder Peter Steinberger confirmed Autopilot is also built on OpenClaw (X post, reported by Reworked) — Microsoft's launch announcement does not mention OpenClaw. 404 Media-reported internal Scout planning docs labeled its first rollout phase "Make people addicted" (Reworked). Andreou at the briefing: agents' "autonomy and flexibility… is the same thing that makes them absolutely terrifying to an IT admin"; on cloud-vs-local posture (apparent dig at Anthropic's Claude Cowork local VM): "sending your most sensitive files down to a local machine into a VM was a complete nonstarter."
• COMPANY CLAIM (event briefing, via GeekWire/thelettertwo): Andreou reported Copilot quality metrics over recent months — speed +25%, per-user engagement and satisfaction doubling, daily usage up 8× in the last year — and the consumer-commercial merger context (~90 million people pay for Microsoft 365 personally; several consumer features — Copilot Podcasts, Group Chat, Deep Research — were shut down during the merge). Nadella (X, Sept 25): "We're building Copilot as a new OS for work that spans every model, every form factor, and every task"; Spataro: "Just as Office defined work for the PC era, the new Copilot is built to define it for the AI era" (both via The Verge).
• CONTEXT (pre-window, primary): Cowork itself went GA June 16, 2026 (Charles Lamanna post) after a March–June Frontier preview — it has been a separately-toggled experience billing in Copilot Credits since then; the Sept 25 launch folds it into Home and extends the same billing rail to Code and Autopilot. The consumer/commercial Copilot app merge happened in August 2026 (The Verge).

• One app, three modes of work — Microsoft's "convergence" moment. After years of Copilot sprawl (Bing sidebar → consumer app + M365 Copilot as two brands/apps, merged in August), the enterprise AI surface is now explicitly three-mode: ask (Chat), delegate (Cowork), build (Code), plus autonomous staffing (Autopilot). This mirrors the pattern OpenAI (Codex merged into ChatGPT this summer) and Anthropic (Claude chat + Cowork unified Sept 16) established — Microsoft's late but distribution-strong copy (thelettertwo; GeekWire frames the launch as "Microsoft's answer to OpenAI and Anthropic… selling directly to businesses").
• The unit of knowledge work broadens from "file" to "file + purpose-built software." Code declares small tenant-hosted apps/automations a fourth artifact class alongside document/spreadsheet/deck, built by non-developers, with developers staying in GitHub Copilot. Combined with the Managed Runtime, Microsoft is betting that vibe-coded internal software gets an official, governed home instead of escaping to shadow tools — and it is opening that same runtime to third-party builders (Lovable first on the record).
• Pricing crossed the seat/consumption line at the center of the enterprise AI market. "Per-seat plus consumption" — the model hyperscalers use for cloud — now applies to agent labor on the platform that owns most corporate productivity seats: fixed USL for everyday AI, metered Copilot Credits ($0.01 each) for Cowork/Code/Autopilot and frontier models. Microsoft's rhetorical framing (pricing blog): metering everyday AI is "the wrong deal… companies may ration access," so everyday stays flat while long-running autonomy carries the meter. Budget ownership shifts from "licenses bought" to "spend governed."
• Agent supervision becomes an IT-governance discipline with teeth. The design is inversion-of-default: nothing agentic is billable/enabled until an admin creates a spending policy; model families are allow-listed per group (and that constrains Auto); credit requests route into existing approval workflows; Graph APIs make policy scriptable; Agent 365 supplies identity, audit and inventory; every Autopilot is an assignable directory object. That is a compliance story (Reuters' framing), not just a product story — and it is Microsoft's explicit differentiator against "almost impossible to bring into the enterprise" OSS agents (Andreou).
• Autopilot legitimizes the always-on personal agent for the enterprise — and buries its origins. Scout's OpenClaw engine was never un-asked (Nadella publicly credited it in July; Steinberger confirms it for Autopilot), yet the launch material is silent on it. What changed is the wrapper: same autonomy class, now with tenant identity, memory, its own computer, email address, @mention reachability, audit, and a meter — "the autonomy you want, the controls IT requires" is the whole product pitch.
• Grounding got a name and a pricing sheet. "Microsoft IQ" as the unified context layer (Work IQ, Fabric IQ with 20M+ semantic models, Dynamics/Power Platform) plus a governed plugin registry converts "context advantage" into the moat argument — while UBB's four cost inputs (model, context, tools, runtime) make that same context usage directly billable.

What is actually available on the event date (announced vs live) — critical for grading claims:

• The three surfaces (FACT as described; capabilities COMPANY CLAIM until shipped): Home = activity recap + suggestions + Chat (instant Q&A/drafting) + Cowork (delegate an end-to-end job: RFP response, launch kit, briefing, financial close package) + Office in Copilot (real, editable, live-synced docs; @mention teammates; "not just some simplified viewer" — Andreou). Code = NL spec → Copilot picks an approach → builds widget/dashboard/cloud-hosted internal app in a sandbox, hosted in-tenant, grounded by Microsoft IQ, plugins auto-sync, sharing "as simple as saving and sharing a Word document." Autopilot = named agent with role+goal; watches channels, follows up on threads, runs recurring work, resumes projects days later, unattended in a cloud computer inside the tenant; @mention in Teams/Outlook/docs; identity in the company directory with user-controlled permissions (Reuters); Microsoft demo: "Dot," a regional manager's agent that tracked Black Friday inventory across Teams channels, email threads and inventory data, sent daily summaries and flagged a shipment problem affecting 18 stores.
• Model/routing layer (FACT — pricing blog; primary): USL includes manual model choice (GPT-5.6, Sonnet 5 in; Opus 5 arriving with limits) and Auto, which scores each request on accuracy/speed/cost and routes to a model and reasoning-effort level; fair-use caps with warn→(Auto free | credits) behavior. UBB exposes the frontier end (Astra, Fable) with full model choice. Admins' model-family allowlists propagate into Auto's candidate set — so routing policy is a governance object, not just a product setting.
• Metering mechanics (FACT — primary docs): UBB is denominated in Copilot Credits = common currency across eligible services ($0.01/credit standard PayGo; P3 pre-purchase with discount; Azure-subscription linkage for billing; consumption reporting by policy/user/group/agent/service/funding source; read-only finance roles). Per-task credit cost = f(model use, context retrieval, tool calls, runtime) — the runtime term is what makes unattended agents financially different: spend accrues while you sleep.
• Microsoft IQ / grounding (FACT as stated): unified intelligence layer across Work IQ (emails/docs/meetings), Fabric IQ (semantic models), Dynamics/Power Platform data and workflows; plugins = skills + connectors, governed via a single registry with central IT approval.
• Managed Runtime (FACT as stated — its blog): the tenant-boundary execution plane for generated code: Entra identity, policy-governed connectors/endpoints/auditing, Git-based version/source control with preview/current versioning, TypeScript SDK typed connector services, CLI full-lifecycle, M365 admin-center app inventory (access/usage/health/policy). "Open build, managed run" — third-party tools (SDK-compatible, e.g., Lovable) can deploy into the same governed host.
• Under the hood (REPORTED, Microsoft-silent at launch): Autopilot is built on OpenClaw (founder confirmation via X, reported by Reworked; Nadella's July "powered by OpenClaw" for Scout; Verge's Build-season Scout/OpenClaw piece). The open question Microsoft didn't resolve at launch is what wrapper-vs-engine governance boundaries look like when the engine is an OSS framework IT departments may themselves block.

• The pricing model for agentic work just got standardized for the biggest installed base in enterprise software (FACT + INTERPRETATION). USL-for-everyday + Credits-for-autonomy is now the default mental model for roughly 450M commercial seats' worth of organizations, and Microsoft explicitly invites the industry to treat agent spend as FinOps. Finance teams get a new line item class ("where are we spending, are we on track?" is verbatim admin-docs language); rivals (Google, Salesforce, OpenAI, Anthropic) will be priced against this rail.
• Long-running agents got a first-class identity + governance template (INDEPENDENTLY VERIFIED direction of travel; FACT as to Microsoft's design claims). Own directory identity, scoped permissions, audit, email address, @mention reach, spend caps, model allowlists — the package converts "scary autonomy" (Andreou's own framing) into something procurement can actually sign, and sets the checklist security vendors (cf. Okta's Oktane "Blueprint Alliance" for agent security, Sept 22) and regulators will reference.
• The autonomy/cost coupling is the honest economics story (FACT mechanics; INTERPRETATION on impact). Frontier agent work is unbounded in runtime; per-task pricing on model+context+tools+runtime with a $0.01 credit is Microsoft saying agent work is a consumption service, not a seat. Combined with off-by-default gating, it is the first big attempt to industrialize agent budgeting before the bill-shock wave — but Reworked's point stands: an Autopilot that keeps operating after you log off "opens a business up to unexpected bills down the line," and cap-setting becomes a new admin art form.
• Convergence of chat/coding/agents is now contested ground everywhere (INDEPENDENTLY VERIFIED pattern). OpenAI folded Codex into ChatGPT; Anthropic unified Claude chat + Cowork on Sept 16; Microsoft's answer adds the one thing labs lack: the tenant, the identity fabric, the Office files, and 30M+ paying seats to attach agents to. The "AI built for work" positioning (and Verge's skepticism: consumer Copilot "has failed… compared to Claude or ChatGPT") makes this a distribution-and-trust play rather than a raw-capability play.
• The OpenClaw arc is a live case study in how enterprise AI absorbs open-source (INDEPENDENTLY VERIFIED lineage; INTERPRETATION). From "a virus" (March) to product engine (June, credited in July) to unmentioned launch substrate (Sept): autonomy innovation came from OSS; Microsoft's contribution is wrapper — identity, tenancy, meter, audit. Expect that wrapper-value argument to echo through every agent-platform discussion, and expect Steinberger/OpenClaw licensing terms to matter.
• Context is becoming the moat with a price tag (FACT + INTERPRETATION). 20M+ semantic models, Work IQ, Dynamics data: grounding breadth is the stated differentiator (Andreou's "hallucination" reassurance framing) — and the same retrieval volume is a billing input. Vendors selling enterprise context (Glean et al., per thelettertwo) just got told the platform owner competes with them at zero migration friction.

• Budgeted autonomy: organizations can now formally delegate open-ended, multi-day, unattended work to an employee-like agent (name, role, goal, directory identity) with pre-set financial and model guardrails — instead of shadow-deploying OpenClaw builds on laptops.
• A governed home for citizen-dev software: purpose-built trackers/dashboards/apps from natural language, tenant-hosted, IT-inventoried, Git-versioned — and now a published path for third-party builders (Lovable pattern) to land in the same runtime.
• FinOps for AI as a concrete practice: spend policies per group/user, model-family allowlists, credit-request approval flows, Graph API automation of policy, outcome-vs-consumption insights for Cowork — the vocabulary and tooling for agent budgets exists in-product, not just in whitepapers.
• Model-portability politics at enterprise scale: Nadella's "private eval… substitute a model" hedge, multi-lab USL+UBB, promised open-weight support — buyers get a credible switching story inside one bill; labs get metered distribution; MAI models stay demo-menu-absent for now (GeekWire observation).
• @mention-to-agent workflows: humans and agents share the same collaboration surface (Teams/Outlook/docs), enabling "delegate in channel, review in channel" patterns — with @Copilot in Teams giving whole channels shared-context AI.
• A testable template for agent governance (identity + permissions + audit + meter + kill switch via policy) that other platforms, standards efforts (Okta's alliance; S10's rumored lab standards body), and regulators will measure against.

• Autopilot's actual billing rail and rate card: does it move fully onto Copilot Credits from GitHub Copilot credits (Reworked: unstated), and what does an always-on Autopilot-hour cost in credits at GA? Microsoft has published no per-task credit prices for Code or Autopilot yet.
• Revocation semantics: when a human's role changes mid-task, what happens to their Autopilot's delegated access (Reworked's practitioner question)? Does Agent 365 publish agent-identity lifecycle docs before Ignite?
• Consumer UBB design: what do "later this year" usage-based billing consumer M365 Premium/Pro users actually see — caps, warnings, or hard stops — and will consumer agent bills become a support/social-media story?
• GA timeline discipline: do Home/Code land "within weeks" and Autopilot private preview by Sept 30 as stated? Does the plugin registry ship end-of-September (briefing) or "coming weeks" (blog)?
• Model economics: will anything move back from UBB into USL (pricing blog promises some will as "economics allow")? When do open-weight models join, and does MAI ever reach the demo menu? How tight does Auto actually route once Opus 5-with-limits lands?
• Third-party runtime adoption: beyond Lovable, do Cursor-class/agent builders deploy Managed Runtime SDK targets, and does GitHub Copilot feed the same host (blog says Code reuses its technology)?
• Liability stance: with metered, supervised agents, how does Microsoft answer "developer liable for agent conduct" (FTC chair) — does audit/evidence packaging for Autopilot actions become part of the compliance offer?
• Does "powered by OpenClaw" survive contact with enterprise security review — do any Autopilot private-preview participants get the "same autonomy, plus a meter" pitch tested against OSS supply-chain concerns Nadella himself raised in March?

Days–weeks: Autopilot private preview opens (≈ Sept 30); Code reaches Frontier end of month; Home/Office-in-Copilot to Frontier "coming weeks"; D365/Power Platform grounding enters public preview through October; plugin registry GA. Expect preview-billing anecdotes and first real-world credit-cost reports (and first cap-misconfiguration horror stories) within weeks of Autopilot preview landing. October: Copilot Studio agents join Agent 365 cost management; Today private preview. Nov 17–20: Ignite — the credibility checkpoint where preview→GA conversion, per-service credit rate cards, and Autopilot's GitHub-credits-vs-Credits billing question should be answered; Q1 FY27 earnings (late Oct) will be the first to report under the "Agents and Infra" segment and the first to carry any consumption revenue narrative. PREDICTION: (a) Google's Gemini Enterprise will extend seat-plus-consumption pricing within ~two quarters because Microsoft just gave CIOs a comparison rail; (b) at least one bill-shock incident (unattended-agent overrun) will surface publicly before GA, and Microsoft will respond with per-agent hard caps/step budgets rather than policy docs; (c) the Managed Runtime will add a marketplace-like discovery surface by mid-2027, pulling Copilot Studio and third-party builders into one inventory; (d) expect a rename/consolidation of "Autopilot" naming collisions (Windows Autopilot) to be clarified at Ignite; (e) Auto-router telemetry published later will show the allowlist-constrained picker materially shaping lab mix — an early read on whether "multi-model hedge" changes OpenAI/Anthropic enterprise revenue share.

The app redesign is the headline; the meter is the story. Microsoft quietly did what nobody else had: made long-running autonomy a billable, budgeted, IT-gated utility for the platform that owns the world's office seats — everyday AI stays a flat subscription ("metering everyday AI is the wrong deal"), while agents that work while you sleep run on runtime-metered credits with names, identities, email addresses and — crucially — spending policies that are off by default. It's also the most candid launch in Copilot's history about where the good ideas came from: the autonomy engine is OpenClaw, the thing Nadella called a virus in March, now tenant-wrapped and un-mentioned in the blog post. So give the demo its due — @mention your agent like a colleague, it flags a shipment problem across 18 stores — but hold the keynote confidence against the calendar: Home, Code and Autopilot are previews and "coming weeks," the Learn admin inventory didn't even list the new services on launch evening, and Copilot's brand promise has been redesigned into the ground before. Judge Microsoft on two dates, not one: Sept 30 (does Autopilot preview actually open) and Ignite in November (does anyone's bill survive it). If both hold, seat-plus-consumption becomes the enterprise AI standard and agent governance becomes this decade's cloud cost management. If they slip, this joins the long list of Copilot relaunches — with a meter nobody's paid yet.