TL;DR
Business Central's agents handle back-office ERP work well and front-of-funnel work not at all. The honest split, the licensing math, and how to extend it.
→ See how this applies to your business (free 30-min call)If you run Dynamics 365 Business Central, you've probably been pitched "AI agents" twice: once by Microsoft, and once by a partner reselling Copilot Studio. Both pitches are real. Neither one covers the part of your business that actually loses money.
Here's the split that matters. Microsoft's agents are excellent at structured work that happens inside the ledger — turning an emailed purchase order into a sales order, matching a supplier invoice to a receipt, drafting an item description. They are not designed to touch the part of the process that happens before anything reaches the ERP: a lead calling at 7pm, a quote request sitting in an inbox, a customer who wants to book something today.
That gap is where most of the recoverable money is. Understanding why requires being precise about what each layer is for.
What Microsoft Actually Ships
Business Central's agent story sits on three layers, and they get conflated constantly.
Copilot features inside Business Central. Assistive, in-context, and scoped to a single task: drafting marketing text for an item, suggesting a chart of accounts mapping, summarizing a record, helping you reconcile bank transactions. You invoke them; they respond. Genuinely useful, low risk, low ceiling.
Autonomous agents. The Sales Order Agent is the flagship example: it monitors an inbox, reads inbound customer requests, identifies the customer and items, and drafts a sales order for human review. The Payables Agent works the same shape on the purchase side. These run without being invoked, which is the meaningful distinction — they have a trigger, a task, and a review checkpoint.
Copilot Studio agents you build yourself. A general agent builder that can connect to Business Central through connectors and APIs. This is where custom work lives — an agent that answers "what's the stock position on SKU X across warehouses," or one that chases overdue invoices with the right tone for each customer segment.
The mental model: layer one is a better keyboard. Layer two is a junior clerk with a defined desk. Layer three is whatever you're willing to design and maintain.
ERP agents are good at the work that starts once a customer is already a record in your system. The expensive failures happen before that.
What They Genuinely Do Well
Be fair to the tooling — the wins here are real and worth taking.
If your pain is back-office throughput, start here and stop reading pitches. The built-ins will cover a lot of it.
Where It Stops — and Why That Matters
Business Central is a system of record. Its agents inherit that posture: they act on records that already exist, inside a transactional boundary, with a human approving anything consequential. That's correct design for accounting. It's the wrong shape for the front of the funnel, for four reasons:
No real-time inbound channel. ERP agents watch inboxes and documents. They don't answer a ringing phone, and phone is still where high-intent buyers go.
Latency tolerance is inverted. A payables agent that runs in a batch every fifteen minutes is fine. A lead response that takes fifteen minutes has already lost most of its value.
Conversation isn't the interface. Sales requires a back-and-forth that establishes fit, urgency, and budget. Order agents parse; they don't qualify.
The record doesn't exist yet. The prospect isn't a customer, doesn't have a number, and isn't in the system. The agent has nothing to act on.
This is not a criticism. It's a boundary, and knowing where the boundary sits is what stops you from buying the wrong solution to your actual problem.
The Licensing Math Nobody Puts in the Deck
Business Central agents don't bill like a per-seat feature. Autonomous agent actions and Copilot Studio usage consume message capacity, purchased in packs and drawn down as agents work. A few practical consequences:
The failure mode to avoid is enabling an agent on a broad trigger, discovering the consumption in month two, and disabling the whole program. Scope narrowly, measure, then widen.
The Architecture That Actually Works
The right pattern is layered, not either/or. Business Central stays the system of record. Agents live at two different altitudes:
Inside the ERP: Microsoft's agents handle order creation, payables, and data hygiene. Keep the human approval step. Log every agent action so you can audit what it did and why.
In front of the ERP: a separate, latency-optimized layer that handles inbound leads, qualification, and booking — then writes the qualified result into the CRM and, when it becomes a real order, into Business Central.
Between them: a clean integration boundary. Business Central exposes APIs and OData endpoints; use them deliberately with idempotent writes so a retry doesn't create a duplicate customer or a duplicate order. This is the part that breaks in production, and it breaks quietly.
The reason to keep these layers separate is that they have opposite requirements. The ERP layer optimizes for correctness and auditability, and can afford to be slow. The front layer optimizes for speed and conversation quality, and can afford to be wrong occasionally because a human reviews the booking.
What We Build in Front of an ERP
For businesses running Business Central, the piece we typically add is the one the ERP can't reach: AI caller agents that pick up or call back every inbound lead within 90 seconds — evenings, weekends, holidays — and run a real qualifying conversation. Fit, urgency, budget, timeline. Qualified prospects get booked onto a calendar and pushed into a GoHighLevel pipeline with the full transcript attached. Only when the deal is real does anything cross into Business Central as a customer and an order.
That division of labor is the whole point. The ERP agent doesn't need to learn to sell, and the sales agent doesn't need to learn double-entry accounting.
A Sane Rollout Sequence
Pick one document type. Highest volume, most standardized. Enable the relevant Microsoft agent on that alone.
Keep humans in the approval loop for at least 200 documents. Track the correction rate. Under about five percent, consider loosening. Above fifteen, your master data is the problem, not the agent.
Instrument capacity consumption weekly. Set an alert before you set a budget.
Fix master data before you widen scope. Agents amplify whatever quality your item, customer, and vendor records already have.
Only then look at the front of the funnel — and treat it as a different project with different tooling and different success metrics.
Define the metric per layer. ERP agents: documents processed per hour and correction rate. Front-of-funnel agents: time to first contact, qualification rate, booked appointments.
Teams that run this sequence get compounding wins. Teams that enable everything at once get a capacity bill and a trust problem.
The Honest Summary
Microsoft's Business Central agents are a real productivity gain for back-office document work, priced by consumption, and correctly conservative about autonomy. They will not fix slow lead response, because that was never their job. If your revenue leak is at the front — leads going cold before anyone calls back — the ERP roadmap will not reach it, no matter how many agents you enable.
If you want to see what the front-of-funnel layer looks like sitting cleanly in front of your existing ERP, [book a free strategy call](/book) and we'll map the integration boundary and what it would take to close the response-time gap.
Free Weekly Briefing
One AI Marketing Tactic.
Every Tuesday. Free.
What's actually working across our client accounts right now — ROAS moves, follow-up sequences, creative angles. The stuff that isn't in any blog post yet.
No spam. Unsubscribe anytime. 1,200+ business owners already in.