Dynamics 365 AI Agents: AI Agents for Dynamics 365 Sales, Finance, and Business Central

They are Microsoft first-party autonomous agents that live inside the Dynamics 365 apps, act on Dataverse and finance and operations data directly, and are metered separately from your Dynamics licences. They are distinct from Copilot features, which are assistive prompts inside a screen, and distinct from agents you build yourself in Copilot Studio, though all three draw on the same Copilot Credit meter.

The distinction that matters commercially is agent versus Copilot capability. Business Central, for example, lists roughly a dozen Copilot capabilities (marketing text, autofill, summarize records, bank reconciliation assist, e-document matching) and exactly two agents: Sales Order Agent and Payables Agent. A Copilot capability waits for a user to press a button. An agent monitors a mailbox or a queue and works on its own. When a vendor deck says "Dynamics 365 has AI everywhere", it is usually counting the first group.

The second thing worth knowing before you plan around them is release status. Microsoft uses a label called production ready preview, which means supported for production use but still preview, so terms, behavior, and pricing can change and the agent may not be covered by the usual service commitments. As of August 2026 that label covers the Account Reconciliation Agent in Finance, the Supplier Communications Agent in Supply Chain Management, and all three Project Operations agents. Field Service scheduling is plain preview. Everything in Sales, Customer Service, and Business Central is generally available.

Read the table above as a coverage map rather than a feature list. The CRM half of Dynamics 365 has a real agent portfolio today. The ERP half has a handful of pilots.

Not in the way most buyers expect. The agents are not a per-user add-on to your Dynamics 365 seats. They consume Copilot Credits, a tenant-wide currency that replaced Copilot Studio "messages" on September 1, 2025, and the same pool feeds every custom agent anyone in your organization builds.

Microsoft publishes the consumption rates. They are worth memorizing because they decide your bill far more than any licence line:

• 1 credit per classic answer (a scripted response)
• 2 credits per generative answer
• 5 credits per agent action, which includes triggers, deep reasoning, topic transitions, and computer-using agents
• 10 credits per tenant graph grounded response
• 13 credits per 100 agent flow actions
• 8 credits per page of content processing
• 10, 35, or 75 credits per minute for classic, generative, and premium generative voice

Credits are sold as a prepaid tenant pack, commonly quoted at $200 a month for 25,000 credits, which works out to $0.008 a credit, or pay-as-you-go through an Azure subscription at about $0.01. So an agent action costs roughly four cents and a tenant graph grounded answer roughly eight.

Those numbers look trivial until you run Microsoft's own worked example. Its documented customer support agent handles four classic answers and two generative answers per conversation for 900 customers a day, which is 7,200 credits a day, or about 216,000 a month. That is more than eight 25,000-credit packs, roughly $1,730 a month at the prepaid rate, for one support agent. Its documented order processing agent, by contrast, fires four agent actions per run at 20 credits a run. Conversational volume is what costs money. Back-office automation is cheap.

Two rules decide whether you land at the cheap end or the expensive end.

Unused credits do not roll over. Capacity is enforced monthly and whatever you do not spend disappears. Buying for your peak month means overpaying for eleven others.

Enforcement is a cliff, not a slope. When a tenant reaches 125% of prepaid capacity, Microsoft disables custom agents. Conversations already running finish, then the agent stops answering and users get "There is a billing issue" or "This agent is currently unavailable. It has reached its usage limit." Agent flows are stricter still: new runs are blocked at 100% of capacity. If you would rather absorb overage than have an agent go dark in front of a customer, link the environment to pay-as-you-go, because environments on the Azure meter are exempt from enforcement.

This is the single biggest lever on Dynamics 365 agent cost and it is buried in a footnote of the billing rates table.

Every rate above is charged at zero for employee-facing use when the person interacting with the agent holds a Microsoft 365 Copilot licence and the agent runs under their authenticated identity. Classic answers, generative answers, agent actions, tenant graph grounding, and agent flows triggered by "When an agent calls the flow" all become no charge. Microsoft describes this as business-to-employee usage included in the Microsoft 365 Copilot user subscription licence, subject to fair usage limits it reserves the right to change.

The practical consequence is that the same agent has two completely different price tags depending on who talks to it. Microsoft's own sales performance agent example makes this explicit: an agent used by 50 licensed users and 100 unlicensed users bills for the 100 and zero-rates the 50.

So the cost planning question is not "which agent" but "who is on the other end":

• Internal agents for licensed staff (finance close, procurement follow-up, project time and expense, internal knowledge) are effectively free if you already pay for Microsoft 365 Copilot. Turn them on.
• Customer-facing agents (a Contact Center agent answering the public, a website support agent, a voice agent) bill every single response. These are the ones to model before you deploy.
• Computer-using agents are never included, even for licensed employees. They bill at the 5-credit agent action rate regardless.

If you already run Microsoft 365 Copilot at $30 a user a month, the internal agent tier you are entitled to is the most underused thing in the licence. The same logic applies across the Microsoft estate, which we cover in the breakdown of Microsoft 365 Copilot pricing and in AI agents for Microsoft Teams.

Sales is where Microsoft has invested most and where the agents are generally available rather than preview. Seven are listed today.

Sales Qualification Agent researches inbound leads on its own, decides whether they fit, sends outreach email, and engages with replies. It reached general availability in late October 2025 and is the closest thing Microsoft ships to a dedicated AI SDR. Sales Opportunity Agent works further down the funnel, researching open opportunities, flagging emerging risk, and surfacing the ones worth attention. Sales Close Agent runs the end-to-end cycle: engaging customers, recommending products, handling objections, and driving to a transaction using templated personalization.

Sales Research Agent answers complex questions against your sales data in natural language, which is the analyst job rather than the seller job. Recommended Actions Agent prioritizes what a seller should do next. The Sales agent in Microsoft 365 Copilot sits outside Dynamics, letting sellers search and act on CRM data from inside Teams and Outlook, and AI-powered Data Enrichment quietly fills opportunity records from email interactions.

For a Dynamics-standardized sales team this is a genuinely complete set, and because sellers are usually the group that already holds Microsoft 365 Copilot, most of this usage is zero-rated. The honest caveat is the same one that applies to every autonomous outbound tool: booked meetings is a flattering metric and meetings held is the real one. We go through what the published benchmarks actually show in the comparison of AI SDR agents and what they cost, and AI agents for Salesforce covers the equivalent decision on the other major CRM.

The service side has the second-deepest portfolio, and all of it is generally available.

Case Management Agent automates the case lifecycle end to end: creating, updating, resolving, and closing, including autonomous follow-up and closure. Customer Knowledge Management Agent reads closed cases, extracts what was learned, fills gaps in your knowledge base, and checks compliance, which is the maintenance job that kills most knowledge programs. Quality Evaluation Agent scores interactions against a framework supervisors define, replacing the sampled manual QA most teams do on two percent of contacts.

Customer Intent Agent and its voice counterpart discover the intents actually present in your conversations rather than the ones someone guessed at design time, and intent-driven routing uses those to route contacts. A rollout manager phases agent adoption and an agent insights dashboard gives supervisors live performance visibility.

Two cautions. First, this is the expensive corner of the credit model, because customers are by definition unlicensed users and every response bills. Model a realistic month before you switch a public-facing agent on. Second, be skeptical of deflection projections: vendors across this category advertise 30% to 60% deflection while independent measurement puts the median nearer 22%, which we set out in full in the analysis of AI ticket deflection rates. If your business case needs 50% deflection to work, it probably does not work. Compare against the wider market on AI agents for customer service.

This is the section most buyers arrive looking for, and it is the thinnest.

Dynamics 365 Finance has one agent. The Account Reconciliation Agent automates matching and clearing between subledgers and the general ledger, notifies you of differences, proposes resolutions, and can resolve them for you. It is aimed at accountants and controllers closing a period, and it is in production ready preview.

Dynamics 365 Supply Chain Management has one agent. The Supplier Communications Agent handles inbound supplier email, follows up with vendors, and updates purchase orders and change requests against rules you define, automating a slice of procure-to-pay. Also production ready preview.

Business Central has two, and both are generally available. Sales Order Agent reads customer emails, finds the customer, and takes an order from request to confirmation. Payables Agent monitors a mailbox for vendor invoices, analyzes them, and puts drafts in front of a human supervisor for review. Alongside them Business Central has real Copilot assistance for bank reconciliation and for matching e-document lines to purchase orders, which is useful but is not autonomous.

Project Operations has three, all production ready preview: Time Entry Agent drafts time from bookings and historical patterns, Expense Agent extracts and itemizes receipt data from a forwarded mailbox or a phone photo, and Approvals Agent checks incoming time and expense against your uploaded policy documents and marks each record ready for approval or needs review.

Set against a full ERP footprint, the gaps are obvious: no collections agent, no AP agent in Finance and Operations, no close checklist agent, no intercompany agent, no demand planning or inventory exception agent, no dispatch agent in Field Service beyond scheduling. If one of those is the process costing you a headcount, the choice is to wait for a Microsoft roadmap you do not control or to run a ready-made agent against it now. The specifics by function are on AI agents for finance, AI agents for supply chain, and AI agents for procurement. Teams weighing Dynamics against the other big ERP suite will find the same pattern documented on AI agents for SAP.

This is not an either-or, and anyone selling it as one is selling something. The two options fail in different places, so the useful thing is knowing which failure you are buying.

Use the Microsoft first-party agent when it exists for your process, is generally available rather than preview, and the humans using it hold Microsoft 365 Copilot licences. Native agents see your Dataverse security model, inherit your Dynamics permissions, and cost nothing extra for internal use. Nothing beats that.

Use a ready-made marketplace agent when the process you need covered is not on Microsoft's list, when the only option is a preview agent and you need production commitments, when the workload is customer-facing at volume and the credit math does not work, or when the work spans systems Dynamics does not own. A lot of real finance and operations work involves a mailbox, a spreadsheet, a bank portal, and a vendor system that has nothing to do with Dataverse.

Build in Copilot Studio when the process is genuinely specific to your company and worth maintaining. Be honest that this is a build: someone owns the topics, the knowledge sources, the testing, and the credit consumption forever, and it draws on the same pool your first-party agents use.

A practical starting sequence: turn on the generally available Microsoft agents that cover your staff-facing work, since those are close to free. List the processes left over. Take the single most expensive one and run a ready-made agent against it for a month on a flat subscription. If it earns its keep, keep it and move to the next. That order costs almost nothing to be wrong about, which is more than can be said for an ERP AI implementation project. Our checklist for evaluating any agent before it touches a live system is in how to vet an AI agent, and the wider market pricing picture is on AI agent pricing.