Copilot in Dynamics 365 and SAP: What the Product Owner Must Frame Before Activation

Table of Contents

• Introduction: AI Is Already Managing the Last Mile of Your ERP
• The 2026 Wave: What Microsoft Is Actually Shipping in Dynamics 365
• Joule Meets Copilot: The End of the Two-Assistant Problem
• The Great Divide: Triple-Digit ROI, Negative Trust
• What Copilot Reveals (and Doesn’t Create): The Real Governance Risk
• Business Central: When an Anthropic Agent Processes Your Expense Reports
• The Real Job of the Product Owner Facing These Agents
• A Five-Point Framing Method, Tested in the Field
• Three Framing Mistakes That Cost Real Money
• Conclusion: The Technology Is Ready, Governance Rarely Is
• FAQ
• Expert Insight
• Let’s Move Forward
• References

Introduction: AI Is Already Managing the Last Mile of Your ERP

In March 2026, Microsoft published its « 2026 release wave 1 » roadmap for Dynamics 365, Power Platform, and role-based Copilots: a set of capabilities rolling out between April and September 2026, spanning sales, service, finance, supply chain, human resources, and commerce. The central message came straight from Microsoft itself: innovation no longer happens twice a year. It’s continuous now.

For a CIO or CFO running a Dynamics 365 or SAP environment, that continuity isn’t necessarily good news on its own. It’s constant pressure: every quarter brings new agents capable of acting directly inside the ERP — placing orders, reconciling invoices, generating quotes — without a clear governance framework necessarily keeping pace.

I’ve spent the last fifteen years framing Dynamics 365 CRM and Business Central deployments across multi-country environments, certified in Microsoft PL-300 (Power BI Data Analyst) and MB-800 (Dynamics 365 Business Central Functional Consultant). What I see arriving in 2026 isn’t a technical problem. It’s a framing problem — exactly the job of a Product Owner.

This article isn’t trying to convince you to adopt Copilot: competitive pressure and the scale of Microsoft’s investment have largely settled that question for most organizations already. The real issue lies elsewhere — in how you frame that adoption: which data an agent can access, which actions it can execute without validation, and how you document all of it to stay auditable. That’s exactly what 2026’s adoption and trust numbers reveal, once you read between the lines: the technology is moving faster than the governance meant to contain it.

What follows draws on real framing engagements across Dynamics 365 CRM, Business Central, and hybrid SAP/Microsoft environments — not vendor marketing material. The goal is to give a CIO, a CFO, or a fellow Product Owner a concrete, testable framework for the next Copilot rollout on their roadmap, grounded in what’s actually shipping in 2026 rather than what’s promised for some future release.

The 2026 Wave: What Microsoft Is Actually Shipping in Dynamics 365

The Dynamics 365 2026 release wave 1 plan introduces agentic capabilities across nearly every module. Dynamics 365 Sales now lets sellers rely on Copilot to build their pipeline, enrich opportunities, and accelerate deal closure, with unified Copilot experiences that blend CRM data with Microsoft Graph — emails, meeting recaps. Dynamics 365 Customer Service continues extending agentic capabilities across case management, email, customer intent detection, and quality evaluation.

On the ERP side, Business Central is « accelerating the move to intelligent ERP » with AI agents automating sales and purchase scenarios starting in the 2026 release wave 1. Three agents currently structure this automation: the Sales Order Agent, which reads customer requests and prepares quotes or orders for review; the Payables Agent, which processes vendor invoices, matches purchasing records, and prepares transactions for approval; and the Expense Agent, which uses receipt information to classify and prepare employee expenses.

The common thread across these three agents: they don’t replace human decision-making, they prepare work for review. It’s a nuance that a lot of marketing communication tends to blur, and it sits at the exact center of the governance question.

Dynamics 365 Customer Insights — Journeys pushes the same principle into marketing orchestration: end-to-end agentic customer engagements across sales, marketing, and service, letting businesses proactively react to customer behavior via Copilot and AI agents. As with the ERP side, every interaction benefits from shared data and consistent intelligence across CRM applications — which, from a governance standpoint, means a poorly siloed piece of data in one module now propagates automatically to the others.

The release cadence itself has changed in nature. Microsoft says so explicitly: the classic biannual structure is giving way to a continuous cadence, with monthly « Business Applications Update » briefings between the two major waves. For a Product Owner running an AI governance backlog, that means a framing document locked in at initial deployment becomes stale within months — governance itself has to become an iterative process, not a one-time project deliverable.

Joule Meets Copilot: The End of the Two-Assistant Problem

The most structurally significant announcement of 2026 doesn’t come from Microsoft alone. At SAP Sapphire and Microsoft Build 2026, the two vendors unveiled bidirectional AI agent interoperability between Joule, SAP’s AI assistant, and Microsoft Copilot — with production-ready integrations shipping across Teams, Excel, Outlook, and Microsoft Fabric.

To understand the scale of the shift, it helps to revisit the problem it solves. Historically, an employee working in an SAP context switched to Joule; the same employee, working in a Microsoft 365 context, used Copilot. Neither assistant had any real awareness of the other’s domain. That boundary is dissolving. Joule is now available as a native Teams app, approved for enterprise deployment through the Microsoft Teams App Store.

For organizations running both technology stacks — an estimated 70% of Fortune 500 companies — this integration is anything but incremental. It eliminates the most painful part of the « two AI assistants » problem: duplicated context, re-entering data, confusion over which assistant to query for which piece of information. But it also raises, more sharply, the question of who governs what when two vendors’ AIs access the same enterprise data.

It’s worth pausing on that 70% figure. It doesn’t describe a niche overlap — it describes the default enterprise IT posture for the majority of large organizations globally, the product of decades of SAP as the operational backbone and Microsoft as the collaboration layer sitting on top of it. Joule-Copilot interoperability isn’t a feature aimed at a narrow segment; it’s infrastructure for how most Fortune 500 companies already work, which is exactly why the governance gap it opens deserves more scrutiny than a typical feature launch would.

Concretely, a salesperson preparing for a client meeting in Teams will soon be able to query, in a single conversation, both SAP order data and Dynamics 365 pipeline data — without technically knowing, or needing to know, which of the two AIs processed each fragment of the response. That’s a genuine operational efficiency gain. But from a data governance standpoint, it meaningfully complicates traceability: when incorrect information circulates, tracing it back to its source — Joule or Copilot, SAP data or Microsoft data — becomes a more complex technical exercise than before, at a moment when both vendors are communicating mostly about fluidity, rarely about the auditability of that fluidity.

This is a point I now recommend writing into the scope of any SAP/Microsoft convergence project in 2026: require, from the framing stage onward, a traceability mechanism that identifies the source of every composite response. Without that requirement set upfront, it never appears spontaneously in a systems integrator’s standard deliverables.

The Great Divide: Triple-Digit ROI, Negative Trust

Copilot’s 2026 adoption numbers tell a two-speed story. On one side, the scale is remarkable: Microsoft reported 420 million monthly active Copilot users in Q1 2026, up from 230 million a year earlier. More than 90% of Fortune 500 companies use Copilot in some form, and 64% have an active deployment — up from 40% at the end of 2024. A Forrester Consulting study commissioned by Microsoft puts ROI at 116% for large enterprises and up to 353% for SMBs over three years, with more than nine hours of productivity gained per user per month.

On the other side, a much more uncomfortable signal circulates in the same body of research. Real penetration — paid seats relative to the total Microsoft 365 user base — sits at roughly 6.5%. And more tellingly: according to a Recon Analytics analysis, 44.2% of users who abandon Copilot cite distrust of its answers as their primary reason. The tool’s accuracy Net Promoter Score sat at -19.8 as of January 2026 — a negative score means users who try the tool are statistically more likely to distrust it than to recommend it.

These two realities coexist because they measure different things. Mass adoption measures tool availability and competitive pressure to install it. Trust measures the quality of the framing that preceded deployment — and that’s precisely where the role of a Product Owner or Business Analyst changes a project’s trajectory.

The breakdown by company size confirms this reading. According to a market study published in early 2026, large enterprises above 5,000 employees show a 64% adoption rate — rising sharply — while the mid-market segment (500 to 4,999 employees) plateaus at 31%, and SMBs under 500 employees at just 12%, concentrated mainly in tech and professional services. This hierarchy isn’t accidental: larger organizations have dedicated data and governance functions capable of overseeing deployment, while smaller structures more often activate Copilot without that upfront framing capacity — and end up statistically overrepresented among the 44.2% who churn out of distrust.

The number of active agents in the Microsoft 365 ecosystem has itself grown 15x year over year, rising to 18x in large enterprises. This isn’t just a quantitative shift: before 2025, most agents were experimental and manually triggered. They now own multi-step workflows end-to-end, without intervention at every step. That exact shift — from reactive assistant to acting agent — is what makes upfront framing non-negotiable: an experimental agent you can interrupt at any time doesn’t carry the same governance requirements as an agent running a full accounting reconciliation chain without continuous supervision.

What Copilot Reveals (and Doesn’t Create): The Real Governance Risk

A widely circulated 2026 analysis pinpoints a precise mechanism: Copilot doesn’t create new security vulnerabilities. It amplifies access-rights mismanagement that was already sitting on internal networks. Concretely, a poorly configured SharePoint or Teams permission that had stayed invisible for years becomes, with Copilot, immediately exploitable: the tool makes sensitive data instantly accessible to employees who should never have been able to see it, simply because they technically held a misconfigured read permission.

This finding pushed Microsoft to invest heavily in the governance layer. The Copilot Control System gives IT and security teams tools to detect oversharing risks, apply permission-correction recommendations, and anticipate problems before they materialize. In May 2026, Microsoft launched Agent 365, a governance layer dedicated specifically to AI agents, priced at $15 per user per month as a standalone offering — a sign that agent governance has become a market in its own right, not a side feature.

A technical example illustrates well why functional expertise remains irreplaceable. A consultant who audited a Copilot-generated DAX measure for a financial services client discovered the tool had used the SUM function where a semi-additive measure required LASTNONBLANK — an error that would have overstated account balances by 300 to 400%. The documented finding from Power BI experts is unambiguous: Copilot reliably handles common patterns (time intelligence, variance measures, classification), but struggles with semi-additive measures, complex filter manipulation, and row-level-security-aware measures. On these cases, human expertise remains essential.

This kind of error isn’t trivial in a regulated context. An account balance measure overstated by 300 to 400%, surfacing unchecked in a report headed to an executive committee or a regulator, would constitute a serious incident — not just an operational inconvenience, and potentially a reportable one depending on the jurisdiction and the nature of the entity involved. This is exactly the kind of scenario that upfront semantic model preparation is meant to prevent: Microsoft explicitly recommends documenting naming conventions, field descriptions, and modeling rules before opening up Copilot access — work that typically falls to a Business Analyst or a data Product Owner, not a system administrator.

Microsoft’s Power BI governance documentation also insists on a simple principle that’s often skipped out of urgency: grant Copilot access only through security groups, and only after a user has completed training or demonstrated understanding of the tool’s limitations. This incremental, phased rollout approach contrasts sharply with what’s observed in practice — a global one-click activation, followed by incident management on the fly.

There’s a broader lesson in the SUM/LASTNONBLANK example that goes beyond Power BI specifically. Every generative tool trained primarily on common, well-represented patterns will default to those patterns when the underlying data model doesn’t clearly signal otherwise. A semi-additive balance measure looks, syntactically, almost identical to a simple additive one — the difference lives entirely in business logic that a well-documented semantic model encodes and an undocumented one doesn’t. This is precisely why the framing work described later in this article treats semantic model preparation as a governance control, not a data-quality nice-to-have: it’s the layer that determines whether Copilot’s statistical defaults align with your actual accounting rules or silently diverge from them.

Business Central: When an Anthropic Agent Processes Your Expense Reports

A technical detail often goes unnoticed in mainstream deployment guides: Business Central’s Expense Agent doesn’t run on the same infrastructure as the other Copilot agents. Unlike the Sales Order Agent and the Payables Agent, it runs on Anthropic models, under a formal subprocessor approval, and is billed per receipt processed through Copilot Credits rather than bundled into a standard user license. As of mid-2026 it was still in preview, initially limited to English-language use in the United States, with UK and Canadian availability expanding from July 2026 onward.

This detail isn’t anecdotal for a Product Owner: it means the AI subprocessor map of a Dynamics 365 environment is no longer homogeneous. A GDPR compliance audit or a contractual subprocessing clause now has to distinguish, agent by agent, which model processes which data — rather than simply validating « Microsoft » as a single subprocessor.

Beyond this specific case, Business Central’s 2026 version introduces the ability to design custom agents through the AI Development Toolkit. Goals and instructions are defined in natural language rather than complex AL code, testable in sandbox environments, with assignable profiles and permissions, and reviewable execution logs to understand the reasoning followed. Agent definitions can be exported as JSON and version-controlled in a system like GitHub — a software engineering discipline now applied to objects that, just two years ago, fell squarely under functional configuration.

The Real Job of the Product Owner Facing These Agents

Faced with this landscape, the question that comes up in most of the engagements I frame is never « should we activate Copilot? » The answer is almost always yes: competitive pressure and documented ROI make a wait-and-see approach untenable. The real question is: which actions do we hand to the agent without human validation, and which ones stay behind a checkpoint?

That’s a product trade-off, not a technical setting. It requires knowing the business well enough to recognize that a semi-additive finance measure deserves double validation, that a misconfigured SharePoint access right becomes a risk the moment Copilot goes live, or that an expense-report agent running on a distinct subprocessor changes the compliance map. That’s exactly the interface between business, data, and IT that a senior Product Owner or Business Analyst owns — and it’s a role no tool, however powerful, replaces.

A Five-Point Framing Method, Tested in the Field

Across my Dynamics 365 and SAP framing engagements, five points come up systematically before any production Copilot activation:

1. Audit access rights before activation. Map who has access to what across SharePoint, Teams, and Power BI semantic models — before Copilot goes live, not after. This is Microsoft’s own explicit recommendation in its Power BI Copilot documentation, and it’s the most frequently skipped step.
2. Prepare the semantic model. Business-friendly naming conventions, field descriptions, modeling best practices. Microsoft states it plainly: skipping this preparation produces less useful and less accurate Copilot outputs.
3. Map subprocessors by agent. Identify which model (OpenAI, Anthropic, other) processes which data, agent by agent, to properly document GDPR compliance.
4. Build an autonomy matrix by action. For each activated agent — Sales Order, Payables, Expense, or other — explicitly define what it can execute alone and what stays subject to human validation, with particular attention to semi-additive financial measures and access-rights-sensitive calculations.
5. Roll out incrementally, with training first. Grant Copilot access by security group, only after training or proof of understanding of the tool’s limitations — the approach Microsoft explicitly recommends in its own governance documentation.

These five points aren’t another bureaucratic checklist: each one directly answers a documented failure mode covered earlier in this article. The access audit answers the finding that Copilot amplifies misconfigured permissions. Semantic model preparation answers the risk of errors like the SUM/LASTNONBLANK case. Subprocessor mapping answers the fragmentation introduced by agents like the Expense Agent running on separate infrastructure. The autonomy matrix answers the need to distinguish low-risk from high-risk agents. Incremental rollout directly answers Microsoft’s own most explicit governance recommendation.

On the ground, this framing typically takes shape as a short document — two to three pages — but a living one, revisited with every new release wave rather than frozen at launch. That’s the difference between AI governance that actually keeps pace with Microsoft’s continuous update cycle, and governance that goes stale the quarter after it’s written.

Three Framing Mistakes That Cost Real Money

Activating Copilot before auditing access rights. This is the most frequent and the costliest mistake: the tool only exposes permissions that were already misconfigured, but it does so at a speed and scale no manual audit can match.

Treating all agents as equivalent from a governance standpoint. An agent that drafts an email and an agent that reconciles vendor invoices don’t call for the same level of oversight — nor, as seen with the Expense Agent, necessarily the same technical subprocessor.

Confusing seat deployment with real adoption. The 2026 statistics make this unambiguous: tool availability (provisioned seats) and measurable actual usage (active users, prompts processed) are two distinct curves. Managing purely against license counts purchased masks the real success indicator.

Conclusion: The Technology Is Ready, Governance Rarely Is

The 2026 Dynamics 365 and Copilot ecosystem landscape has little left of the experimental about it. Agents are already processing invoices, preparing quotes, and reconciling accounts in production, across tens of thousands of companies. The technology has moved well past proof-of-concept status.

What hasn’t kept pace is the framing: who decides what an agent can do alone, how the AI subprocessor map gets documented, which financial measures deserve double validation. That’s the work of a Product Owner, a Business Analyst, or a senior functional consultant — not a checkbox in an admin panel. Organizations that invest in this framing before activation capture the ROI Forrester documented. Those that skip it feed the distrust and abandonment statistics.

The Joule-Copilot convergence announced in 2026 will accelerate this dynamic in both directions. It will make generative AI even more seamless and more powerful in the daily work of SAP and Microsoft teams — and in the same movement, it will complicate traceability for anyone who hasn’t anticipated the question at the scoping stage. The organizations that win on this ground in 2026 and 2027 won’t necessarily be the ones that deployed Copilot fastest, but the ones that managed to document, measure, and evolve their AI governance at the same pace as their vendors’ quarterly updates.

It’s a project that demands a dual skill set, rarely combined in a single profile: genuine functional mastery of the ERP — sharp enough to spot a miscalculated semi-additive measure or a poorly calibrated access right — and the ability to translate that expertise into clear product trade-offs for leadership. That dual skill set, built on real deployments rather than theory, is what separates a Copilot project that delivers on its promises from one that swells the abandonment statistics.

None of this is a call to slow down. If anything, the opposite is true: the organizations best positioned for the next release waves are the ones that treat governance as a competitive advantage rather than a compliance tax. A well-framed Copilot deployment ships faster, not slower, because fewer incidents mean fewer rollbacks, fewer emergency access reviews, and less erosion of the very trust that determines whether an organization’s Copilot investment shows up as one of Forrester’s 116% ROI cases — or one of Recon Analytics’ churn statistics.

For a Product Owner walking into a Dynamics 365 or SAP engagement in late 2026, the brief is no longer « explain what Copilot can do. » Most stakeholders have already seen the demos. The brief is: build the framing artifact — the access audit, the autonomy matrix, the subprocessor map — that turns a promising pilot into a production deployment leadership can actually trust, agent by agent, release wave by release wave.

FAQ

Does Copilot in Dynamics 365 replace a Product Owner or Business Analyst?
No. Copilot agents prepare actions (quotes, reconciliations, classifications) for human validation; they don’t set governance policy, arbitrate business risk, or own product prioritization — that’s the core of the Product Owner’s role.

Should access rights be audited before or after activating Copilot?
Before, systematically. Copilot doesn’t introduce a new vulnerability: it makes already-misconfigured permissions immediately exploitable.

Does Business Central’s Expense Agent use the same model as the other Copilot agents?
No. It runs on Anthropic models under a separate subprocessor approval, unlike the Sales Order Agent and the Payables Agent — a point that needs specific documentation in a compliance audit.

What’s the real enterprise Copilot adoption rate in 2026?
It depends on the metric: more than 90% of Fortune 500 companies use it in some form, but real penetration relative to the total Microsoft 365 user base remains around 6.5%.

Expert Insight

What strikes me most about the 2026 cycle isn’t the speed at which Microsoft and SAP are shipping agents — it’s the persistent gap between that speed and the governance maturity of the organizations receiving them. I’ve seen IT departments activate Copilot in a few clicks without ever having audited their SharePoint permissions in years. The result is almost never a dramatic incident; it’s a slow erosion of trust, which is exactly what that negative NPS reflects. The Product Owner’s role in this context isn’t to slow adoption down — it’s to make it measurable, reversible, and documented, agent by agent.

Let’s Move Forward

Running a Dynamics 365, Business Central, or SAP deployment under pressure to activate Copilot fast? Upfront framing — access-rights audit, per-agent autonomy matrix, AI subprocessor mapping — is what separates documented ROI from adoption that erodes over time. Let’s talk about your Dynamics 365 context, or if your environment runs on Business Central, let’s discuss your ERP framing.

References

• Microsoft Learn — Dynamics 365 2026 release wave 1 plan
• Microsoft Dynamics 365 Blog — 2026 release wave 1 plans for Dynamics 365, Power Platform and Copilot Studio
• SAVIC Technologies — SAP Microsoft Copilot Integration 2026, CIO Guide
• Rand Group — Exploring the Microsoft Dynamics 365 Copilot 2026 release wave 1
• DynamicsSmartz — Microsoft 365 Copilot: Features, Pricing, Licensing & ROI, 2026 Guide (Forrester TEI data)
• StackMatix — Microsoft Copilot Enterprise Adoption in 2026: What the Data Shows
• StackMatix — Microsoft Copilot Adoption Statistics & Trends 2026
• AIBusinessWeekly — Microsoft Copilot Statistics 2026: Users, Seats & Revenue (Recon Analytics NPS data)
• HumanizeAI — Microsoft Copilot Adoption Trends 2026
• ERP Software Blog — Dynamics 365 Business Central AI Agents: 2026 MCP Guide
• MSDynamicsWorld — What’s Coming to Dynamics 365 Business Central in 2026
• Microsoft — Agentic ERP Software Solutions, Dynamics 365 (Forrester Total Economic Impact study, Feb-March 2026)
• ia-info.fr — Microsoft Copilot security risks analysis, 2026
• Microsoft Security — Securing and governing Microsoft 365 Copilot (Purview)
• Power BI Consulting — Power BI Copilot: AI Report Creation Guide (SUM/LASTNONBLANK DAX case)
• Learn Microsoft — Copilot in Power BI: overview and benefits of the integration