Blog
Copilot and agents in Business Central: hype vs reality
Across the 2025 and 2026 releases, Business Central has gained a steady stream of Copilot features and, more recently, autonomous agents. The marketing is breathless. The reality is more interesting, and more useful. Our advice splits the features into two camps: turn on now, and pilot carefully.
What to turn on now
Several Copilot features have crossed from demo to daily driver. They share a trait: the human stays in control, and the AI removes drudgery rather than making the decision.
- Autofill suggests field values as you create records, so you type less and copy-paste less.
- Summarize turns a long record (a customer, a posted document, a history) into a readable paragraph.
- Tell Me / chat lets users find pages, actions and answers in natural language instead of hunting through menus.
- Bank reconciliation assist proposes matches between statement lines and ledger entries, leaving you to confirm.
These are low-risk, high-frequency wins. They speed up work people already do, the user reviews the output anyway, and there is no autonomous action to audit. For most clients this is exactly where to start.
What to pilot carefully
The newer autonomous agents are a different proposition. The Sales Order Agent can read incoming order requests and draft sales orders; the Payables Agent can process vendor documents from a mailbox. These act on your behalf, across multiple steps, with far less per-action human input.
Two things deserve attention before you switch them on. First, they are billed via Copilot Studio messages: agent activity consumes capacity, and that consumption needs modelling against expected volume, not discovering on an invoice. Second, they take real actions: a drafted order or a processed payable touches your data and your process, so the audit trail and the review step matter far more than with a summary. An agent processing the wrong supplier emails is a problem you want to find in a pilot, not in production.
Human-in-the-loop is not optional
The most reliable pattern we have seen, and the one BC’s own tooling now supports, is human-in-the-loop: the agent prepares, a person reviews and commits. The 2026 releases added task panes, in-context review and a way to stop all agent tasks at once, precisely because oversight is what makes agents trustworthy. Finance and operations teams will not, and should not, hand control to an unmonitored agent. Keeping a person on every action for the first weeks builds both the audit trail and the team’s confidence.
Model the consumption before you commit
Because agents are metered, treat consumption like any other operating cost. Estimate message volume from real document counts, run a bounded pilot, measure actual usage, and only then decide where wider rollout pays off. What separates an agent that saves money from one that surprises the CFO is a spreadsheet built before go-live, not after.
Where we land
Turn on the assistive Copilot features broadly; they are safe and immediately useful. Approach autonomous agents as scoped pilots: one mailbox, one process, a person reviewing every action, and a clear read on consumption. Widen scope only once the audit trail and the numbers both check out.
The technology is genuinely good and getting better with each wave. Adopted with judgement, it removes real work; adopted blindly, it adds cost and risk.
Wondering which camp each feature falls into for your business? Tell us your story and we will tell you plainly.