Ignite 2026 for Customer Service and Process Architects
Photo by Roman Kraft on Unsplash

Table of Contents

  1. Key announcements and their status
  2. Copilot, Microsoft IQ, and Agent 365 around Dynamics 365
  3. Microsoft Contact Center Intelligence
  4. Customer Conversation harness in Copilot Studio
  5. The three harnesses in Copilot Studio
  6. Dynamics 365 Customer Service plugin in Copilot
  7. Agent Accounts, memory, and triggers
  8. Microsoft IQ
  9. Copilot Managed Runtime
  10. Windows 365 for Agents
  11. Microsoft-Decision-1.1
  12. Governance and security for agents
  13. Cost management for usage-based AI
  14. Dynamics 365 Activate and Salesforce integration
  15. Reference architecture for customer service
  16. My recommended next step
  17. Sources

Ignite 2026 takes place from November 17 to 20 in San Francisco and online, and the Book of News is long again. Most of it is about Azure, Foundry, security, and Windows, and very little of it mentions customer service by name. Still, when I read it with my customer service and process glasses on, a lot of it changes how I would design a service organization today. This article is my reading of it. I wrote it for the people who decide where the money goes and for the people who have to build what was decided, because after this Ignite those two groups need to talk to each other more than before.

I wrote this article on November 19, while Ignite is still running. More sessions and announcements will follow until the event closes on November 20. I will update the article after that, so it gives you the full picture.

Key announcements and their status

These are the announcements I think matter for customer service and process work, along with the status Microsoft provided.

Announcement Status at Ignite Why I care
Microsoft Contact Center Intelligence (CCI) Generally available Microsoft AI on top of the contact center you already run
Customer Conversation harness in Copilot Studio Generally available (preview since October) A builder made for customer-facing voice and chat agents
Microsoft Copilot harness in Copilot Studio Generally available The natural home for longer process agents
Dynamics 365 Customer Service plugin in Copilot Generally available Case work can happen in Copilot, with a review step
Microsoft IQ Generally available, some parts preview The shared context your agents will read from
Agent Accounts, memory, and triggers Agent Accounts public preview in December 2026 Agents that own a step in a process
Copilot Managed Runtime Generally available One place to host and govern apps, agents, and workflows
Windows 365 for Agents Partly available, several parts preview A way to automate old screens with no API
Microsoft-Decision-1.1 Generally available A small, cheap model for routing and classification
Agent 365 Cost Management and Business Analytics Preview and Frontier Control over usage-based AI cost
Dynamics 365 Activate CRM generally available, ERP private preview Help with moving off Salesforce or SAP ECC

In short, the question is no longer whether we can build a good service agent on Microsoft technology. The questions now are which runtime it should use, what it can know, whose identity it acts under, and who pays for it.

Copilot, Microsoft IQ, and Agent 365 around Dynamics 365

Jeff Teper’s Ignite post for Dynamics 365 explains the thinking better than the Book of News does. He says business applications now need three things: an intelligent user experience, context, and governance. His phrase for the first is that “Copilot is the front door to your business, and plugins teach it how you work.” Microsoft IQ supplies the context, Agent 365 supplies the control, and Dynamics 365 sits underneath as what Microsoft now calls the “agentic system of record”.

For contact centers, this is a familiar structure. We always had channels, an interaction layer, knowledge, a system of record, and some kind of oversight. What changed is that every one of those layers now has AI in it that can act, not just suggest.

I like the direction. A service representative who can read a case history and draft a reply in Teams or Copilot doesn’t need to jump between three applications. What I keep an eye on is the system of record. When an agent changes a case outside Dynamics 365, the business rules, audit trail, and permissions still have to apply. Microsoft says they do, through review checkpoints and existing permissions, and that is the first thing I would want to see working in a real tenant.

Microsoft Contact Center Intelligence

Microsoft introduced Contact Center Intelligence in public preview on October 21, 2026. At Ignite, Teper wrote that it “is now generally available”, and on November 18 Omar Aftab followed with a full general availability announcement.

The idea is simple. Microsoft took the AI built into Dynamics 365 Contact Center and made it work with other contact center platforms through what the October post called “an open connector”. You keep your telephony, routing, and agent desktop. You add Microsoft’s agents, quality management, and analytics on top.

According to the Microsoft Learn “What’s new” page, updated November 17, 2026, the generally available scope covers integrations with supported CCaaS platforms (with initial guidance for Genesys Cloud and a universal connector for others), quality and operational insights, voice biometric authentication, and the reuse of agents built in Copilot Studio together with extension through Foundry. The October launch also mentioned a Service Operations Agent as “coming soon”. It does not appear in the November list, so I would not count on it yet.

The November 18 announcement adds several points that matter for planning:

  • Partnerships with Avaya, Cisco, and Five9. Microsoft says it is working directly with these vendors “to build and maintain out-of-the-box connectors”. The open connector stays available for every other platform.
  • Handoff to people. Agents handle self-service on your existing voice and digital channels and then hand off to representatives who keep working in the desktop they know. According to Microsoft, context and voice authentication carry over when a call moves to a representative.
  • A new speech model. CCI will offer Microsoft AI Realtime 1, a speech-to-speech model, and Microsoft says its own labs built smaller models tuned for service work such as quality assurance.
  • Governance from the first minute. Every agent built in CCI appears in the Agent 365 registry as soon as a maker names it, before it is published.
  • A path to Dynamics 365 Contact Center. Microsoft describes the later move from CCI to Dynamics 365 Contact Center as a seamless upgrade.

Until now, when a customer asked me how to get Microsoft’s contact center AI, the honest answer was usually “move to Dynamics 365 Contact Center”. For an organization in the middle of a long contract with another vendor, that answer ended the conversation. Now the AI layer and the contact center platform are two separate decisions. Teper says it plainly: many organizations “don’t have time to replace their existing contact center infrastructure”. The vendor partnerships make this more concrete. If you run Avaya, Cisco, or Five9, you can expect a maintained connector instead of building the integration yourself.

It also adds a new integration point to the architecture. Between your CCaaS platform and CCI you now pass live audio or messages, events, customer context, and handoffs to people. I would want clear answers on what your platform actually exposes to the connector, whether the handoff really keeps its context in your setup so the customer does not have to start over, where recordings, transcripts, and voiceprints are stored, and which system has the final word on quality scores when both your vendor and CCI can evaluate a conversation.

Voice biometrics is now part of a generally available offer. Biometric data used to identify a person is special category data under GDPR. Bring your data protection officer in before the first pilot, not after it.

Customer Conversation harness in Copilot Studio

Every agent in Copilot Studio runs on a harness. Microsoft Learn describes the harness as the runtime between your agent design and the model. It decides when the model is called, what context it gets, and which tools it uses.

The Customer Conversation harness went into public preview in October, and Jason Moore’s Copilot Studio post from November 18 says it is now generally available. Omar Aftab described it as “a purpose-built harness and builder specifically designed to create powerful conversational agents for customer engagement”. What makes it interesting for service is the way it is put together. One agent handles both voice and chat, so a customer can move from the website to the phone without losing context. The agent can be a team of specialist sub-agents, each with its own knowledge and tools. In October Microsoft said each sub-agent would be able to use its own model at general availability. The November 18 CCI announcement now lists this as “coming soon”, so it did not arrive with general availability. Pro-code agents built in Foundry can join that team, and Microsoft IQ is available to all of them.

The part I like most is that generative and scripted behavior live side by side. A skill can be left to the model, or it can follow a fixed topic step by step. In a regulated contact center that is exactly what you need. You let the model handle the open “my bill looks wrong” conversation, and you script the identity check and the payment word for word.

Two practical notes from Microsoft Learn. You can start building in Copilot Studio or in the CCI portal, but only the portal connects the agent to a live contact center. And agents on this harness are billed in Copilot Credits.

In my view, this is the first time Microsoft treats the customer conversation as its own design problem instead of a variation of an internal assistant. With general availability, it is ready for production pilots. I would still keep the first rollout narrow. Per-agent model choice is not there yet, and it will change the cost picture once it arrives.

The three harnesses in Copilot Studio

Copilot Studio now offers three harnesses. The Microsoft Copilot harness is the one many of us knew as the GitHub Copilot harness. Microsoft renamed it on October 28, 2026, and said existing agents keep working with no change to capabilities or billing. At Ignite, Teper listed it as generally available.

This is roughly how I explain the choice in customer workshops:

Question Standard harness Microsoft Copilot harness Customer Conversation harness
Who talks to the agent Employees or customers Mostly employees and processes Customers
How behavior is defined Topics, triggers, branches Instructions, skills, tools Sub-agents with skills and topics
Best for Predictable, rule-based flows Long, multi-step process work Voice and chat service with human escalation
Example Internal help desk Accounts payable exceptions Billing hotline with a payment step
Billing Included with some licenses, or consumption Copilot Credits Copilot Credits

Microsoft published numbers comparing the Microsoft Copilot harness with the standard one. In Teper’s post, multitool performance rose from 87.0% to 98.9% and input tokens per task fell by 45%. The renaming post reports pass rates 10.8 percentage points higher with an OpenAI model and 22.1 points higher with an Anthropic model. These come from Microsoft’s own evaluations. I take them as a sign of direction, not as a promise for your process. A customer example from Jason Moore’s November 18 post is more useful to me: by adding a Copilot Studio agent to its finance automation, Asurion raised touchless invoice processing from less than 20% to more than 70%.

There is also a preview that upgrades a standard agent to the Microsoft Copilot harness. It carries over instructions, tools, and prompts, rebuilds topics as skills or workflows, and gives you a report of what moved and what did not. Useful, but try it on a few agents before you trust it with a whole estate.

Upgrading from the standard harness to the Microsoft Copilot harness can move an agent from license-included use to usage-based billing. Make that cost decision before the technical one.

Dynamics 365 Customer Service plugin in Copilot

The case form is no longer the only place where case work happens. Microsoft announced Dynamics 365 in Copilot, and the Customer Service plugin is generally available.

Jared Spataro’s post mentions 13 service skills. Copilot can triage cases, gather case history, look up answers in knowledge, draft replies, and prepare escalation summaries, “with case updates held for review before anything is committed”. Agents built on top can go further and create, update, follow up on, and close cases, turn solved cases into knowledge, and rate interaction quality.

The Microsoft Learn article on using these skills in Cowork fills in what an architect needs to know. Representatives approve proposed field updates through checkpoints before anything is written to Dynamics 365. Everything runs inside the permissions the user already has. Teams can add their own skills for their workflows, knowledge sources, and triage rules. And to get started you need Cowork access under your licensing and spending policy, Customer Service licensing, the Service app, and the Customer Service MCP configuration.

From the business side, this helps with the work that happens around a conversation: preparing a callback, writing up an escalation, finding cases that have waited too long for a customer reply. That work rarely shows up in contact center metrics, but it costs real time. From the technical side, two things become new parts of your solution. The Service app and its MCP setup need the same change control as any integration. And your triage and escalation rules, once written as skills, need an owner and a review rhythm just like your knowledge articles.

Agent Accounts, memory, and triggers

In process digitalization there is often a gap between a person starting a workflow and the system running it on its own. These announcements help close it.

Teper wrote that Agent Accounts enter public preview in December 2026, together with memory and triggers, and Jason Moore confirmed the December date on November 18. The accounts give Copilot Studio agents managed credentials and access to approved resources. An agent with its own account gets a Microsoft Entra ID, a mailbox, and a presence in Teams, according to Microsoft Learn. People can hand it work in a document, by email, or in Teams, and it keeps going. Business process memory, announced in October, means that a detail one colleague explains to the agent becomes shared context for everyone else using it. Triggers start the agent when something happens, such as a new email or a new SharePoint item, and the Book of News adds a Work IQ “listen” capability in preview so agents can react to new messages or business events. Work IQ also gets finer policies and human approval options in preview.

Together, this lets you design a process in which an agent is a named participant. Think of a complaints mailbox owned by an agent. It reads each complaint, opens a case, asks for missing details, and passes the case to a person when the rules say so. Its actions are logged under its own name, not under the maker’s account, which is a big step forward for audit.

An agent with a mailbox and a Teams presence will get mail and messages from real customers and colleagues. Decide what it may answer, what it must pass on, and who looks at what it sends before you give it an address.

Microsoft IQ

Microsoft IQ is generally available, and the Book of News points out it comes with “no new SKU or subscription”. It pulls context from four places:

Source What it brings A service example
Work IQ Mail, meetings, chats, files, people, and now Dynamics 365 data An exception agreed in a Teams thread
Fabric IQ Business data with meaning from semantic models Customer value or contract status
Foundry IQ Documents, search, and knowledge graphs Product manuals and warranty terms
Web IQ Public web information with source controls Public notices about outages

Teper wrote that Dynamics 365 data is now natively integrated with Work IQ, so people can update customer records in plain language “while respecting existing permissions and governance”. That is a big convenience for service teams.

It also raises a question about sources. The Customer Conversation harness announcement uses the example of a billing dispute where the agent reads a refund policy in SharePoint and an exception agreed in a Teams thread. For an employee assistant, that is great. For an agent that talks to customers, a Teams thread is usually the last place I want it to read from. Somebody has to decide which sources a customer-facing agent may use and which stay internal, and that decision belongs in the design, not in the go-live checklist.

Business plugins in Microsoft IQ, available in the Frontier program, let experts organize guidance for an organization, a business unit, or a project, with inheritance down the structure and sensitivity labels on top. For a service organization with several countries or brands, that inheritance fits nicely: one global complaints policy, with local additions. And the SharePoint indexed knowledge source for Foundry IQ is now generally available and respects source permissions, which is the cleanest way to use SharePoint-based service knowledge in custom agents.

Copilot Managed Runtime

Copilot Managed Runtime is generally available. It hosts and governs apps built in Copilot Code, Copilot Cowork, Copilot Studio, and compatible third-party tools, and every app shows up in the Microsoft 365 admin center. Teper added that app creation in Copilot Studio is generally available and runs there too. Copilot Code itself is in the Frontier program, with general availability planned for December.

For process teams the more interesting news is the developer story. The first version of Agents and Workflows in Copilot Managed Runtime lets developers build Copilot Studio agents and workflows locally, work on them with their own tools, publish with the CLI, and keep them under version control in GitHub. Apps that need a database or storage can also use Fabric as their backend.

Jason Moore’s November 18 post adds more for teams that build together. Agent sharing is generally available, so several makers can work on the same agent in Copilot Studio. Command-line tools and Copilot Studio MCP are in public preview, so developers and their AI assistants can work with Copilot Studio components from their own tools. Business users can also start agents and workflows in natural language from Copilot Cowork or Copilot Code, in public preview, and those agents land in Copilot Studio where IT can govern them.

This helps speed, but it also creates a support risk. The line between “a tracker someone built over lunch” and “a business application” is getting very thin. That is good for speed. It is bad for whoever has to support it two years later. I would set some simple rules early: who may publish, which data backends are allowed, and when an app has to move into a proper solution with an owner.

Windows 365 for Agents

Most service organizations have at least one old application with important data and no API. Windows 365 for Agents is Microsoft’s way of letting agents use those screens.

The Book of News lists several changes. Windows 365 for Agents is now eligible for Copilot Credits, so its cost shows up with your other agent costs. It is available as a computer-use tool in the standard harness in Copilot Studio and will be in preview in the Microsoft Copilot harness in December. Agents on these Cloud PCs will be able to reach on-premises applications with agent identities, in preview, and persistent Cloud PCs for always-on agents are in preview in the Frontier program.

I see this as a bridge. Screen automation is slower than an API and breaks when the screen changes. It is still a good way to remove manual copying today, as long as the same process stays on the list for a proper integration.

Microsoft-Decision-1.1

A busy contact center makes thousands of small decisions every hour. Which intent is this? Which queue? Is it a complaint? Does it need a supervisor? Sending every one of those to a large reasoning model is slow and expensive.

Microsoft-Decision-1.1 is generally available in Foundry, and the Book of News describes it as built for “routing, classification, prioritization, verification and workflow control”. It returns confidence scores, so an application can hand the case to a person or a larger model when it is unsure. Microsoft already uses it to route requests in Copilot and to set reasoning effort in Cowork.

It is not a contact center product, but it fits contact center design well. A small model that makes the quick call, a larger model for the hard cases, and the confidence score as a clear rule for when a human steps in. That is a pattern I want to try for intent detection and case classification. Copilot Studio is moving the same way. According to Jason Moore, the classify node in workflows now uses a decision-scoring model by default, for scenarios like email triage and routing.

Governance and security for agents

Agents that talk to customers raise the stakes. They deal with people outside the company, they handle personal data, and some of their actions have legal weight. The security section of the Book of News is long, so I picked what I would take to a service architecture board.

Usage and activity insights for Microsoft pre-built agents and agents built on Microsoft platforms are now generally available in the admin center without an Agent 365 license, although full insights for third-party agents and the security features still need Agent 365 or ME7. Plugins and skills governance is generally available, which is how you control who gets the Customer Service plugin. Governance for MCP servers and tools is in preview, including a view of which plugins break when you block a server. An action center for inactive and ownerless agents is expected in private preview in December. Dynamic Agent Risk, in preview, gives each agent a risk level that moves as the agent acts or gets updated. Entra brings Agent ID to discovered local agents and Autopilot, and an MCP Firewall, both in preview. Purview previews data loss prevention that stops Copilot from using sensitive files even when nobody labeled them.

Overall, agents are now governed much like people. Each one has an identity, an owner, a risk level, and limits. For customer service I would add one rule of my own: every customer-facing agent has a named business owner who signs off on what it does, when it hands over, and what quality it must reach.

Cost management for usage-based AI

Nearly everything in this article that does real work is billed by usage. The Customer Conversation harness and the Microsoft Copilot harness use Copilot Credits, Windows 365 for Agents is now eligible for them, and even the Today Feed in the Copilot app needs Copilot credits for its updates during the day.

Microsoft’s answer is Agent 365 Cost Management. The Book of News describes a central place to see and control usage-based AI services such as Copilot Cowork, Copilot Studio, Copilot Managed Runtime, and Work IQ APIs, with spending policies, limits, alerts, and chargeback through Fabric. It is in preview. Forecasting and recommendations are in the Frontier program this month, and value insights across Microsoft and non-Microsoft agents follow in December. Agent 365 Business Analytics, also in Frontier, tries to show value through measures such as success rate, user satisfaction, and assisted hours.

Service leaders already know their cost per resolved contact. If AI cost per resolved contact can sit right next to it, the business case discussion gets a lot easier. I go deeper into how licenses and Copilot Credits fit together in Copilot pricing, licenses, and credits, and into the cost governance side in Agent FinOps in the Microsoft ecosystem.

Dynamics 365 Activate and Salesforce integration

Two announcements are about the systems already in place. Dynamics 365 Activate now offers a generally available Salesforce to Dynamics 365 migration, and for ERP it starts with SAP ECC discovery, requirements analysis, configuration, and data migration in private preview, with project teams reviewing every recommendation. Teper’s post quotes a partner saying one large customer is on track to move 20,000 users from Salesforce in about five months. That is a partner’s statement, not a benchmark, but it shows where Microsoft wants to compete.

On the other side, Salesforce in Copilot is in preview with more than 38 prebuilt sales skills, Sales Cloud in Teams is in pilot, and Data Federation between Salesforce Data 360 and Microsoft Fabric is generally available without copying data.

For mixed estates this is good news, because you do not have to pick one CRM to use Copilot as the front door. Be aware, though, that the service skills are clearly strongest on Dynamics 365. If your service runs on Salesforce, check exactly which service skills exist before you plan around them.

Reference architecture for customer service

This is my own simplified picture of a customer service organization on the Microsoft stack after Ignite. It is not a Microsoft diagram.

The customer side and the employee side now run on different harnesses, but they share one context layer and one set of controls. If they end up reading different knowledge or following different rules, customers will feel it.

My recommendation is to start with one focused piece of work.

Pick a single service journey that has volume, clear rules, and a cost you already measure. Order status, a change of address, or a simple billing question all work. Then use that journey to answer the questions Ignite has raised, with your own data instead of Microsoft’s slides.

If you run a contact center platform from another vendor, the journey should go through Contact Center Intelligence on that platform. It is the fastest way to find out whether the connector carries what you need and whether the handoff to your people feels right to a customer. If you already run Dynamics 365, start with the Customer Service plugin for your representatives, because it is generally available and it shows you quickly how well the review checkpoints fit your own approval rules.

While you build, make three decisions properly, because they will shape everything that follows. Decide which knowledge a customer-facing agent may read and which stays internal. Decide who owns the agent on the business side and what it must hand over to a person. And decide how the usage cost is measured and charged, so you can put AI cost per resolved contact next to the number your service leaders already track. Get your data protection officer involved early if voice or biometrics are part of the journey.

The Customer Conversation harness is now generally available, so it can carry that pilot. I would keep the scope narrow until per-agent model choice arrives, because it changes the cost per conversation. Agent Accounts and the cost tooling are still in preview or Frontier, so treat them as things to test, not to build on yet.

In my view, this Ignite is less about one big product and more about the pieces finally fitting together for customer service. The organizations that get value first will not be the ones with the most agents. They will be the ones that treated knowledge, ownership, and cost as design decisions from the first day.

Sources

Written by

Holger Imbery