Is your business ready for agentic AI in Dynamics 365? A readiness check for mid-market leaders

Most mid-market leaders looking at Dynamics 365 agentic AI readiness are excited about the feature set and have not done the readiness work that makes it safe to let an agent act on its own. Microsoft's 2026 release wave 1 reached general availability in April 2026, and it makes agentic AI, agents that act on data and processes rather than just summarizing them, the organizing idea across Dynamics 365 this year. Gartner expects 40 percent of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5 percent in 2025, and Dynamics 365 is one of the applications leading that shift.

Turning agentic features on is the easy part. Deciding whether your business is ready for them to act on their own is the harder, and more important, question. That is a different kind of decision than turning on a new report or a new form, and it deserves the same decision quality you would apply to any other high-stakes call, applied before implementation starts, not after. This post is a Dynamics 365 agentic AI readiness check: what actually changed in release wave 1, what "agentic" means in plain terms for your ERP and CRM, and four questions to ask before you let an agent act.

What actually changed in Microsoft's 2026 Release Wave 1

Microsoft published its 2026 release wave 1 plan on March 18, 2026, covering new capabilities rolling out between April and September across Dynamics 365, Power Platform, and Copilot Studio. The plan touches nearly every module: Sales, Customer Service, Contact Center, Field Service, Finance, Supply Chain Management, Business Central, Human Resources, and Customer Insights.

The headline is not that Copilot got smarter. It is that several agents moved from preview into general availability, which means they are now doing work instead of just describing it. In Business Central, the Sales Order Agent reads a shared mailbox and drafts sales orders on its own. The Payables Agent categorizes, reconciles, and routes invoices for approval. The Expense Agent is built for people who have never logged into an ERP screen in their life, such as field technicians and drivers submitting a receipt from their phone.

In Field Service, the new Scheduling Operations Agent takes over dispatch decisions that used to require a scheduler manually weighing technician location, skill, and job priority. In Supply Chain Management, warehouse agents handle picking, stock rebalancing, and hands-free scanning. None of this is speculative. It shipped, or is shipping, this year.

What "Agentic" Means for your ERP and CRM, In Plain Terms

Set aside the vendor language for a moment. A Copilot assistant drafts something and waits for a person to review it, send it, or throw it away. An agent does more of the work on its own, inside a boundary someone configured in advance, and only stops when it hits a case it cannot resolve.

In your ERP, that boundary might be a dollar threshold, a vendor allowlist, or a rule about which expense categories get auto-approved versus escalated. In your CRM, it might be which accounts a sales agent can draft outreach for without a rep reviewing the message first, or which service cases a case agent can close without a supervisor sign-off. The difference is permission, not intelligence. An assistant drafts something for you to review. An agent has been authorized to act inside a boundary someone configured in advance.

That permission is the entire story. A Copilot suggestion that is wrong costs someone a few minutes to correct. An agent action that is wrong has already touched a customer record, a general ledger entry, or a scheduled technician visit before anyone sees it.

Four questions to ask before turning on an agent

Is the underlying data clean enough to act on?

An agent that reads a messy vendor master file, inconsistent customer records, or outdated pricing tables will act on that mess with total confidence. Fivetran's 2026 Agentic AI Readiness Index found that 60 percent of enterprises are already investing millions in agentic AI, while only 15 percent report having the data foundation to run it securely and at scale. That gap between spending and readiness is the single biggest predictor of a rocky rollout.

Before you assign an agent to a workflow, check the data it will touch. Is the customer or vendor record it reads from a single, governed source, or a patchwork of spreadsheets and legacy fields nobody has cleaned up since a prior system migration?

Who is accountable when the agent gets something wrong?

Every agent deployment has an owner on paper. Far fewer have a named person who reviews what the agent actually did, on a set schedule, and who is expected to answer for it if something goes sideways. Ownership and accountability get treated as the same thing, and they are not.

Before go-live, name the person who reviews the Payables Agent's exception queue every week, not "the finance team" in the abstract. Do the same for every agent you turn on, in ERP and in CRM alike. If you cannot name a person today, you are not ready to name an agent.

Does your team know what the agent is and isn't allowed to do?

Most Dynamics 365 agents ship with default permission settings tuned for a demo, not for your risk tolerance. Sales Order Agent, Payables Agent, and Expense Agent all have configurable thresholds inside the agent designer, and those thresholds should start narrower than the default, not wider.

The team working alongside the agent needs to know the boundary too, in plain language. A finance clerk should be able to tell you, without opening an admin panel, which invoices the Payables Agent handles alone and which ones land back on a human desk.

Is there a rollback plan?

Traditional ERP controls assume a named person initiated a transaction and a second named person approved it. An agent that drafts a sales order or posts an expense report from a mailbox does not fit that model cleanly, which is exactly why regulators and auditors are paying attention.

The EU AI Act's high-risk obligations, which cover AI systems affecting financial decisions, were originally set to take effect on August 2, 2026. A provisional agreement reached in May 2026 pushed full compliance out to December 2027 for stand-alone systems and August 2028 for AI embedded in regulated products, but the underlying requirement for documented human oversight and record-keeping is still coming. Build your rollback and audit trail now, on your own timeline, rather than waiting for a regulatory deadline to force the issue.

Ask what happens the day an agent gets something wrong at scale, not just once. Can you see every action it took, reverse the ones that were incorrect, and turn the agent off without breaking the workflow underneath it? If the honest answer is "we would figure it out," you are not ready yet.

Where agentic features create real value today (vs. where they're still early)

The value is real in specific, contained places. Field Service scheduling, warehouse picking and stock rebalancing, invoice categorization and matching, and sales reps getting a pulled-together view of an account from CRM and email signals are narrow, well-bounded tasks with a clear right answer. Dynamics 365 customers are getting measurable time back from them today.

The early part is anywhere the task is broad, judgment-heavy, or crosses departments. Fully autonomous order creation with no review step is one example. Multi-agent handoffs between sales, finance, and supply chain are another. So is a contact center agent handling a complex complaint end to end. Each of these still needs a human in the loop more often than the marketing suggests.

Adoption numbers back this up from two different angles. Nearly 90 percent of the Fortune 500 have built at least one active agent using Microsoft's low-code tooling, according to Microsoft's own April 2026 earnings commentary. That measures whether an agent exists, not whether it is running unsupervised in production at scale. Pair that with Fivetran's finding that only 15 percent of enterprises have the data foundation to support agentic AI, and a pattern emerges. Building an agent is easy. Trusting it with judgment calls is still the hard part.

Risk also varies by who the agent touches. An agent built for finance professionals who already understand approval workflows is a different risk than the Expense Agent, which is explicitly designed for non-ERP users like field staff. Pilot the second kind with a single team before opening it to everyone with a mailbox.

How to tell if you're ready

You do not need a perfect answer to all four questions above before you pilot anything. You need honest answers. Roughly speaking, organizations fall into one of three positions once they work through the questions.

If your data is governed, you can name an accountable owner, your permission boundaries are documented, and you have tested what happens when you turn an agent off, you are in a position to pilot now, in a contained scope. If the use case is promising but one or two of those four answers are shaky, spend a few weeks preparing before you flip anything on. If the business problem an agent would solve is not clearly defined yet, or nobody can describe what "wrong" looks like for that workflow, wait. Turning on an agent to see what happens is not a strategy.

If you want to go deeper than a self-check, a formal Copilot readiness assessment, the kind that reviews tenant security, data governance, licensing, and change management, typically runs three to six months for a mid-market organization. The four questions above are not a substitute for that work. They are the gate you clear first, so the time and budget you put into a formal assessment or a release-wave rollout goes toward a use case that is actually worth it.

Next step, a 20-minute Red Flag Review

If you want a structured way to work through this before your next release wave decision, we offer a 20-minute Red Flag Review with our team. Limited slots only so book yours below:

20-Minute Red Flag Review

We look at your specific environment across data readiness, accountability, permissions, and rollback planning, and tell you plainly where the flags are red, yellow, or green before you commit budget or timeline to an agent rollout.