Copilot in Business Central: what's actually shipped

Business Central got Copilot features earlier than most of the Microsoft stack — marketing text suggestions shipped back in 2023 — and the list has grown every release wave since. The pattern is consistent: not one big chatbot bolted on the side, but small AI assists embedded in specific tasks, each one suggestion-based and confirmed by a human before anything posts. This guide covers what has actually shipped, how the plumbing works, and where the limits are — because the marketing copy covers everything else.

One framing note up front: Copilot in Business Central is included in the licence. There is no per-user Copilot add-on as with Microsoft 365 Copilot; if you have Essentials or Premium, the features are there to switch on. That changes the adoption maths completely — the question isn't "is it worth the money" but "is it worth ten minutes of setup".

The shipped features

Bank reconciliation assist. The flagship, and the one covered in depth in bank reconciliation in Business Central. After the rules-based auto-matcher runs, Copilot proposes matches for the awkward residue — payments with mangled references, one payment covering several invoices, near-miss amounts and dates — and suggests G/L accounts for statement lines with no counterpart (fees, interest). Every suggestion is accepted or rejected line by line. This is the feature to lead with when demonstrating Copilot to a finance team, because it attacks a task everyone already resents.

Sales line suggestions. On quotes and orders, Copilot turns a rough request into document lines — type or paste "the same as last month's order but 20% more of the blue ones", or attach a customer's file, and it proposes item lines with quantities matched against your item catalogue. The user reviews the proposed lines, adjusts, and inserts. For businesses taking orders by email this removes real keystrokes; for configured-product businesses with heavy pricing logic it's a starting point, not a finisher.

E-document matching. Incoming electronic invoices get mapped to purchase order lines with Copilot proposing the matches — the same "suggest, human confirms" shape as bank reconciliation, applied to the purchase side. If you're adopting e-invoicing mandates anyway, this comes along for the ride.

Analysis assist. Business Central's analysis mode (pivot-style tabs on any list page) accepts a natural-language request — "show sales by customer by month for this year" — and Copilot builds the analysis tab: grouping, columns, filters. The output is a normal, editable analysis tab, so this doubles as a teaching tool: users see how the thing they asked for is actually constructed. For controllers who never learned the feature, this is quietly the highest-leverage assist in the product.

Chat with Copilot. A side-pane chat that finds records ("open unpaid invoices for Adatum"), navigates, and explains concepts and fields, with answers drawn from Business Central's documentation. It's genuinely useful as a "how do I" layer for occasional users, and considerably weaker as an analytical tool — it locates and explains; it doesn't compute across your ledger.

Marketing text suggestions. The original: generate product descriptions from an item's attributes, in a chosen tone, for the web shop or catalogue. Narrow, works, matters mostly to retail and e-commerce users publishing item data outward.

Smaller assists keep arriving wave by wave — suggested number series setups from a natural-language description, record summaries, field autofill — typically landing first as previews you enable per environment. Treat anything marked preview as exactly that: try it in sandbox, don't build a process on it.

Two adjacent things are worth separating from this list. The older sales and inventory forecasting and late-payment prediction features predate Copilot branding — machine-learning extensions, useful but not conversational. And the autonomous agents (a sales order agent that processes incoming order emails end-to-end, with more agents following) are the next chapter: unlike everything above, they act rather than suggest, run as their own identity, and arrive with task-specific approval checkpoints. They deserve a separate evaluation with process-owner involvement, not a casual toggle — the broader agent conversation is in Copilot agents vs Copilot Studio.

The plumbing

The architecture is straightforward and worth being able to explain to a sceptical IT director:

• Grounding is your BC data, scoped to the user. Features read the environment's data through the signed-in user's permissions. Copilot cannot surface a record the user couldn't open, and it has no cross-tenant or cross-environment reach.
• Generation runs on the Azure OpenAI Service operated by Microsoft — not on a public consumer endpoint — and customer data and prompts are not used to train foundation models.
• Admins hold the switches. The Copilot & AI capabilities page in BC toggles each feature per environment, previews are opt-in individually, and there's a data-movement consent for regions where the Azure OpenAI capacity sits in a different geography than the environment. Feature-by-feature control means you can ship bank reconciliation assist to finance without also enabling half-baked previews tenant-wide.
• It's extensible. AL developers can build their own Copilot experiences — the platform ships a prompt-dialog page type and an Azure OpenAI module so ISVs and per-tenant extensions can offer the same suggest-review-accept pattern over custom functionality. Expect your ISV add-ons to grow Copilot features on this rail.

Honest limits

• Suggestion quality tracks data quality. Sales line suggestions against a catalogue full of cryptic item descriptions, or bank matching against customers who pay with no reference, will underwhelm. Copilot amplifies the state of your master data; it doesn't fix it.
• Language coverage is uneven. Features generally work best in English first, with other languages following per feature and wave. If you run BC in a smaller language, verify each feature in it before promising anything to users.
• Chat is not a reporting tool. Users will ask it "what was our margin last quarter" and get a record lookup or an explanation, not analysis. Set that expectation early or the whole Copilot brand takes the blame; point those questions at analysis assist or the reporting stack instead.
• Previews churn. Features rename, move, and occasionally disappear between waves. Anchor training material to shipped features only.

The fair overall read: Copilot in Business Central is a set of well-chosen, low-risk assists that remove tedium from real tasks, free with the licence, and individually switchable. It is not a transformation, and adopting it doesn't need a project — enable bank reconciliation assist and analysis assist, show the right two teams, and let the rest earn its way in wave by wave, tracked like any other change in the release wave rhythm.