Ram Maree · Copilot & AI Architecture
An architect's briefing  ·  compiled & verified August 2026

The Microsoft Copilot & AI Stack — who uses what, and why

Microsoft's AI estate now separates cleanly into four planes: an experience plane (M365 Copilot, Copilot Chat, Cowork, agents-in-apps), a build plane (Copilot Studio low-code, bridged to Microsoft Foundry pro-code), a data plane (the Graph semantic index, Azure AI Search, Fabric/OneLake), and a governance plane (Purview, Entra Agent ID, Agent 365). Since Ignite 2025 and Build 2026, all four are grounded through a shared "IQ" layer — Work IQ, Fabric IQ, Foundry IQ, Web IQ — the connective tissue that lets a low-code agent and a pro-code one draw on the same semantic knowledge. This page maps the stack, decodes the 2026 feature vocabulary, and shows five real configurations organizations are actually buying — including the cost-conscious route in.

Experience Agent build Grounding — the "IQ" layer Data Governance
Illustration: the five planes of the Microsoft AI platform as stacked translucent layers
The centerpiece

The stack, layer by layer

Five layers, four coded planes plus one connective grounding spine. Reused color coding carries through the scenario recipes and comparison table below — the eye should link them on sight.

Illustration: one shared grounding spine feeding every layer of the stack
1 Experience — where people work The surfaces end users actually open
M365 Copilot$30/user/mo · agent mode reached general availability April 2026
Copilot ChatFree tier, web-grounded — not grounded on tenant data inside Office apps
Copilot CoworkAgentic tasks, GA 16 June 2026, credit-metered on top of the seat
Custom apps / TeamsAgents surfaced inside Teams or bespoke line-of-business apps
who builds the agents behind these surfaces
2 Agent build — who builds it, and how Low-code ↔ pro-code, bridged both ways
Copilot StudioLow-code. Priced in Copilot Credits — $0.01/credit, or a $200/mo pack of 25,000 credits.
Microsoft FoundryPro-code. Azure consumption pricing. Renamed from Azure AI Foundry, January 2026.
Bidirectional bridge — a Studio agent can call into Foundry; a Foundry agent can be surfaced back into Copilot Studio / Teams
MODEL TRAFFIC CONTROL — the AI gateway sits in the request path: routing, token metering/chargeback, content safety, spend caps. Microsoft-native: Azure API Management AI gateway (llm-* policies). Alternatives: LiteLLM (open source), Portkey; packaged portals like Cloudforce nebulaONE bundle a gateway + chat UI + agent builder.
both ground on the same IQ layer
IQ Grounding — the "IQ" layer (Ignite 2025 → Build 2026) The connective tissue tying every plane together
Work IQGraph semantic index, productized — API reached GA 16 June 2026
Fabric IQSemantics over business data — the Ontology layer is still in preview
Foundry IQKnowledge bases built on Azure AI Search — core reached GA via API
Web IQWeb grounding for public/current information
the IQ layer reads from these data sources
4 Data — what gets grounded on Tenant content, lakehouse, custom RAG, external
Microsoft GraphMail, files, chats, meetings, SharePoint — permission-trimmed at query time
Microsoft Fabric / OneLakeThe data intelligence platform underneath Fabric IQ
Azure AI SearchCustom RAG over anything — blobs, SQL, PDFs — powers Foundry IQ
External dataGraph connectors (synced) & MCP federated connectors — federated needs the full $30 license
access & oversight wrap every layer above
5 Governance — identity, policy, oversight Applies across all four planes above
PurviewLabels, DLP, eDiscovery — extended to cover agents
Entra Agent IDIdentity for agents, same model as identity for people
Agent 365Agent registry — GA 1 May 2026, $15/user/mo
Illustration: four governing arches over the stack, each a different job

Four different jobs people conflate

Purview governs data (labels, DLP, lineage, eDiscovery). Entra Agent ID governs agent identity. Agent 365 is the agent fleet registry. The AI gateway governs model traffic — and unlike the other three it sits in the request path, not in the oversight plane. Keeping these four apart is the fastest way to sound (and be) precise in a governance conversation.

What's new in 2026

Cowork, Skills, and who picks the model

The three things worth walking into the room already fluent in.

Copilot Cowork

GA 16 Jun 2026

What it is

An agentic execution layer: hand it a whole task and it plans, works across Outlook, Teams, Excel, Word, PowerPoint, OneDrive and SharePoint, and returns a finished deliverable — not a draft. Triggered by an email arriving, a Teams @mention, or Power Automate. Human approval checkpoints gate sensitive actions.

Pricing

  • The $30/user/mo M365 Copilot seat is a prerequisite — Cowork isn't sold standalone.
  • Execution itself is metered in Copilot Credits on top: $0.01/credit pay-as-you-go, or a "P3" commitment plan.
  • Cost per task varies with model, retrieval depth, tool calls and runtime — independent reviewers flag that ROI is hard to predict without piloting.

Adoption vendor claim

Microsoft claims more than half the Fortune 500 are touching it, naming Accenture, Avanade, Capital Group, Koch and Zurich Insurance as adopters.

Demo use cases

  • Comparing ~4,000 files for a product-version comparison
  • Sales-pipeline review with drafted follow-ups

"Skills" — the extension vocabulary

glossary

Microsoft's terminology here is genuinely fragmented across products. This table decodes it — same word can mean different things depending which surface you're standing on.

TermWhereWhat it means
Declarative agentsM365 Copilot / Copilot StudioConfig-driven agents (instructions + knowledge + actions) on Copilot's native orchestrator
Custom engine agentsAgents SDK / FoundryPro-code, bring-your-own orchestration and model
ActionsCopilot StudioAnything an agent can invoke: connectors, flows, REST/OpenAPI, MCP tools
PluginsCopilot Studio / declarative agentsThe packaging layer grouping actions (OpenAPI plugin, Teams message extension)
MCP toolsCopilot Studio (native)Point at an MCP server; its tools auto-surface as Actions
Copilot Skills (new, 2026)PowerPoint, Excel, SharePoint, VS 2026Packaged one-click task workflows (e.g. Excel "Brand Kit", deck-formatting skill)
Fabric data agentsMicrosoft FabricFormerly "AI Skills" — conversational Q&A over lakehouse data
Copilot Tuning (Build 2025) — low-code fine-tuning on tenant data, inside Copilot Studio. Requires roughly 5,000+ M365 Copilot seats (medium confidence — verify before quoting) — an enterprise-only play in practice.

Model choice — who gets to pick, where

as of Aug 2026
M365 CopilotMicrosoft picks. Default is GPT-5.6 (since 9 Jul 2026), with Claude Sonnet 5 selectable (since 2 Jul 2026) and an "Auto" router choosing per task. Admin toggles per provider — Claude is off by default in EU/UK tenants. Users choose per-surface, not globally.
Copilot StudioYou choose from a menu: Auto, GPT-5.5/5.6, Claude Sonnet 5, Claude Opus 4.8; per-topic model assignment arriving mid-2026. Off by default in EU/EFTA/UK, absent in GCC High/DoD.
Microsoft FoundryYou choose anything. Full catalog: OpenAI, Anthropic, Mistral, Meta Llama, xAI Grok, DeepSeek, Microsoft MAI/Phi, Cohere and more — 11,000+ models marketed. PAYG per-token, Provisioned Throughput Units, or Batch.

Indicative model prices — per 1M tokens, third-party price trackers

GPT-5.6 Luna
$1.20 out
GPT-5.6 Terra
$12 out
GPT-5.6 Sol
$30 out
Claude Sonnet 5
$10 out
Claude Opus 5
$25 out
Claude Fable 5
$50 out
ModelProviderInput / 1MOutput / 1MFoundry status
GPT-5.6 LunaOpenAI$0.20$1.20GA (Azure price parity from 1 Aug 2026)
GPT-5.6 TerraOpenAI$2.00$12.00GA
GPT-5.6 SolOpenAI$5.00$30.00GA
Claude Sonnet 5Anthropic$2.00$10.00GA (Azure-hosted)
Claude Opus 5Anthropic$5.00$25.00GA (Azure-hosted)
Claude Fable 5Anthropic$10.00$50.00Preview (Anthropic-hosted only)

~ indicative, per 1M tokens — third-party trackers, not Microsoft or Anthropic list prices. Bars scaled to output price, the usual cost driver.

Moves fast — OpenAI cut Luna 80% on 30 Jul 2026 (cached input $0.02/1M) — the price war is real; model prices are the fastest-moving number in this document.

The rule of thumb

In M365 Copilot, Microsoft picks (with narrow user choice). In Copilot Studio, you choose from a menu. In Foundry, you choose anything.

Which route

Where's your data? What's your budget? Who builds?

Three questions that sort almost every real deployment into one of five configurations.

Illustration: five routes diverging from one starting point
The heart of it

Five ways organizations are actually deploying this

Each card's cost tier ($ – $$$$) is derived directly from the dollar figures quoted in its cost picture, compared across all five scenarios low → high — it is a compression of the numbers below it, not a separate estimate. Scale and impact are not put on an invented numeric scale — the brief data doesn't support one across all five, so scale is stated as the actual reach described, and impact is flagged by its evidence source (independent / vendor-claimed / mixed / unproven) rather than scored.

S1

Free tier first

The cost-conscious route — SMBs, cautious enterprises starting out
Zero seat uplift

Configuration

Copilot Chat (free) 1–3 metered Studio agents SharePoint agents, team-site scoped

Why this route

Pay only for usage; no per-seat commitment. Tests real demand before buying seats.

Watch-out

Watch-out Copilot Chat is not grounded on tenant data inside Office apps. Tenant-graph grounding burns ~10 credits/query — the "my own data" use case is the ~10x-cost one. Federated connectors are locked; they need the full $30 license.

Reality check independent

Gartner: only ~5–6% of Copilot pilots reach large-scale deployment — this is where most orgs stall.

Cost / scale / impact

Cost
$$$$
$0 licenses. PAYG credits: FAQ agent ~$50/mo · generative + doc-search ~$150–300/mo · complex autonomous agent ~$500–1,000+/mo (~ estimates, at 1,000 sessions/mo) — or a $200/mo 25k-credit pack.
Scale
Individual / small-team pilot — no seat commitment
Impact
Independent — Gartner pilot-to-scale data
S2

Seat rollout

Productivity-led enterprise — Microsoft standardizers, wealth/advisory firms
Governance-first

Configuration

M365 Copilot seats, broadly SharePoint Advanced Mgmt — oversharing cleanup first Training program Viva Insights dashboard

Why this route

Single security boundary, works in the apps people already live in, no engineering required.

Watch-out

Watch-out Activation is typically ~36%; 57% of orgs see engagement decline post-launch. Value concentrates in meetings and drafting. Microsoft's own Work Trend Index 2026: only 16% of users progress past individual assistance.

Examples

UBS — 50,000 seats, largest FinServ deal MS-announced · Northern Trust — enterprise-wide, 150+ use-case backlog unverified secondary · Australian government trial, 7,600 users independent — ~1hr/day saved on admin, 86% wanted to keep it, gains strongly training-dependent.

Cost / scale / impact

Cost
$$$$
$360/seat/yr list. 100 seats → $36k/yr · 1,000 seats → $360k/yr · UBS's 50,000 seats ≈ $18M/yr at list (pre-negotiation). Plus a governance program: 4–6 weeks if mature, 8–16 if not.
Scale
Org-wide — 100 to 50,000+ seats
Impact
Mixed — vendor deal announcements + one independent field study
S3

One instrumented agent

The evidence-first route — the smartest small-budget play
Proof before scale

Configuration

A handful of M365 Copilot seats ONE Copilot Studio / SharePoint agent Scoped to a single function's content Copilot Studio analytics

Why this route

The instrumented small deployment beats the uninstrumented big one — it discovers real demand and produces the business case for (or against) S2.

Watch-out

Watch-out No control over chunking in Studio's managed RAG. Precision ceilings eventually push you toward S4.

Impact evidence

Copilot Studio analytics is the only place you get question-level telemetry: AI theme-clustering of what users actually asked, resolved / partially-resolved / escalated / abandoned outcomes, transcripts (29-day retention). Native Copilot Chat gives you none of this. Moody's ran the bottom-up version of this at scale — internal copilot to 14,000 staff harvested 800+ use-case ideas.

Cost / scale / impact

Cost
$$$$
~$200/mo credit pack + a few $30 seats + advisory/build — industry benchmark $10–40k for a 2–6 week pilot.
Scale
Single function / team — scoped pilot
Impact
Independent — first-party question-level telemetry, not a claim
S4

Pro-code vertical copilot

Product-grade, on Foundry — where AI IS the product
Engineering-led

Configuration

Microsoft Foundry — model of choice Azure AI Search / Foundry IQ knowledge bases Custom-engine agent + evals/tracing Optional: bridged into Teams/M365 via Studio

Why this route

Full control — chunking, citations, model choice, evals — and differentiation you can actually sell.

Watch-out

Watch-out A different order of cost and team entirely. Don't start here for internal productivity.

Examples

BlackRock Aladdin Copilot — Azure OpenAI, supervised agentic · Ally.ai — 30% less after-call work, 85%+ summarization accuracy unverified secondary · Morgan Stanley (OpenAI-direct variant) — 98% advisor-team adoption, retrieval 20%→80% independent · Morningstar's own Intelligence Engine — Azure OpenAI since 2023.

Cost / scale / impact

Cost
$$$$
Engineering-led. Azure AI Search from $75/mo (Basic) / $250/mo (S1) + model tokens. Enterprise RAG deployments commonly ~$5k–50k/mo all-in (broad third-party estimate). Build: $75–500k+, 3–12 months.
Scale
Product-specific — one capability, firm-wide use
Impact
Mixed — strong named cases, some independently reported, some not
S5

Sell into Copilot

The data-vendor play — data & content vendors
Revenue line, not cost center

Configuration

MCP server Federated Copilot connectors Branded Copilot agent + Cowork plugin Agent 365 for governance

Why this route

Distribution — be present inside the agents your clients already use. Monetize data as agent-callable tools.

Watch-out

Watch-out Zero exclusivity — everyone's also in Claude and ChatGPT. Your clients need the full $30 Copilot license for federated connectors. Microsoft is becoming a competitor in the same channel (e.g. embedding financial research natively).

Who else is here

The 2026 table-stakes move: Morningstar shipped Jun 2026 · LSEG May 2026 · PitchBook Jun 2026 · Moody's / S&P Global earlier · FactSet.

Cost / scale / impact

Cost
$$$$
Engineering cost of the MCP server + connector integration — but this is a revenue line, not a cost center.
Scale
External distribution — every client already on Copilot
Impact
Early-stage, unproven at scale — 2026 move, no track record yet
The adoption reality

What actually happens when organizations roll this out

Independent field data, not vendor claims. This is why the evidence-first route (S3) exists.

~5–6%
of Copilot pilots ever reach large-scale deployment
Gartner survey data
~36%
typical seat activation in enterprise rollouts
independent field reporting
16%
of users progress past individual assistance into workflow redesign
Microsoft Work Trend Index 2026
6–12 mo
typical platform-hardening delay while business units drift to shadow AI
practitioner field reports
Illustration: a telemetry funnel narrowing from pilots to scaled deployments
The pattern behind the numbers — organizations stall when asked to commit before evidence. The deployments that survive start scoped, instrumented and evidence-first: one governed agent with real telemetry beats a thousand uninstrumented seats.
Side by side

The five routes, compared

Same facts as the cards above, compressed to one row each.

  S1 · Free tier first S2 · Seat rollout S3 · One instrumented agent S4 · Pro-code vertical copilot S5 · Sell into Copilot
License cost$0$360/seat/yr~$30/seat × a few$0 (no seats needed)$0 (build cost only)
Consumption cost$50–1,000+/mo credits—(seat-based)~$200/mo creditsTokens + AI Search $75–250+/moAzure/MCP hosting, variable
Engineering effortLow, low-codeLow–med (governance program)Low–medHigh, pro-codeMed–high (MCP build)
Time to valueDays–weeks2–4 months (rollout + training)2–6 weeks3–12 monthsWeeks–months
Telemetry you getNone (Chat) / basic (Studio)Viva Insights, usage-level onlyFull: theme-clustered, outcome-tagged, transcriptsFull: your own evals/tracingPlatform connector logs, limited
Ceiling / limit~5–6% of pilots scale (Gartner)~36% activation, 57% engagement declineNo chunking control; precision ceilingCost & team scale — not for internal productivityZero exclusivity; MS is also a competitor
Best-fit orgSMB / cautious enterprise pilotingEnterprise standardizing on MicrosoftFunction/team proving ROI firstFirms whose AI IS the productData/content vendors
Named examples— (early-stage, generic)UBS, Northern Trust, Australian govtMoody's (14k staff, 800+ ideas)BlackRock, Morgan Stanley, MorningstarMorningstar, LSEG, PitchBook, Moody's, S&P
About

Who wrote this

Ram Maree — AI & enterprise architect (Azure Solutions Architect Expert, TOGAF 9.2, CBAP), with production AI deployments across financial services and the public sector. This briefing is maintained as a working document: facts are verified against Microsoft Learn, vendor primary sources and live GitHub telemetry, estimates are flagged as estimates, and vendor claims are labelled as vendor claims.

How an engagement usually starts

  • A lay-of-the-land session — this briefing, applied to your tenant and your data estate
  • One scoped, instrumented agent — the cheapest configuration that produces real evidence (S3)
  • A function capability map — what the telemetry says your organization should build, buy or skip next

LinkedIn · GitHub · Technical depth: the Architect's Reference · August 2026 platform update