For most of its life, Microsoft Dynamics 365 Business Central has been an extremely expensive logbook. Orders go in. Invoices post. Inventory moves. At the end of the day it will tell you, with total confidence and zero opinion, exactly what happened. It is the friend who films your crash in 4K and never once shouts a warning.

Which is useful, in the way a trail map is useful after you have already fallen down the mountain. The next move is turning that record into something that tells you what is happening, why, and what to do about it — and in a few well-chosen cases, doing it for you.

Microsoft’s Business Central AI story splits cleanly in two, and the split is the whole ballgame. Copilot features ride along with a person who is still responsible for the decision: analysis assist, chat, summarise, autofill, bank reconciliation assist, sales line suggestions. Agents go further and actually run a defined process with oversight — the Sales Order, Payables and Expense agents are generally available today.

Drop in on a real problem, not on “AI”

The most reliable way to waste a year and a healthy chunk of budget is to buy the technology first and then go hunting for somewhere to put it. This is the full gaper move: brand new powder board, first day of the season, sheet ice from the gondola to the bottom. Everyone can see it. Nobody wants to say anything.

Better opening question: where are our people burning hours on repetitive work that mostly consists of going and finding information?

That framing matters, because AI earns its lift ticket on a fairly specific set of problems. It is genuinely good when the work involves:

  • Large piles of business information nobody has time to read
  • The same investigation, over and over
  • Unstructured mess sitting next to tidy structured records
  • Questions people would rather just ask out loud
  • Classification and summarising
  • Spotting a pattern across time
  • Repetitive decisions that follow rules you can actually write down

Business Central is absolutely full of these. The rest of this run shows you where they tend to hide.

Read the terrain

Business Central is not short of data and never has been. What it is short of is a single human being with a free afternoon and the will to live long enough to interpret it.

Here is what people actually want to know, and rarely get around to asking:

  • Which customers have quietly shrunk their orders three months running?
  • Which items are we buying more of and selling less of?
  • Which vendors have let their lead times slide while nobody was watching?

Traditional reporting can answer every one of these. What changes is the interface. Instead of knowing which report, which filter, which dimension and which saved view, somebody just describes the question.

This does not let you off the hook on data architecture. It raises the stakes. More people are now standing on that foundation, and they cannot see the cracks from up there.

The morning snow report

The simplest use is also one of the best. An executive wants the overnight picture: yesterday’s sales, open orders, inventory exceptions, overdue receivables, purchasing problems, anything that failed in an integration.

Traditionally that is six reports, two exports and a coffee going cold. An AI-assisted workflow assembles the whole thing before anyone has clicked a single tile.

The win is not that a machine wrote a tidy paragraph. The win is that nobody spent the first hour of their day gathering it by hand.

When something yard-sales

Now it gets interesting. A Business Central job fails. The traditional response is to find the failure, read the error, search the docs, review the configuration, inspect related records, check what changed last week, and eventually work out whether you are looking at bad data, bad config, a broken integration or actual code.

That is thirty minutes of walking uphill picking your gear out of the moguls before you are even allowed to start thinking. Full yard sale. One ski in a tree. An AI-assisted diagnostic can do the picking-up:

A human still calls it. They are simply not spending half an hour on the bootpack first.

Let it run the groomer

Business Central keeps moving toward agent-shaped automation. The work changes shape from this:

Employee → Business Central → employee decides → employee acts

to this:

Business Central → AI evaluates → AI recommends or acts → human reviews the exceptions

The point is not getting rid of people. It is getting people off the magic carpet so they are free for the terrain that actually needs judgement. You do not hire a ski patroller and then put them on the bunny hill handing out trail maps.

Wire it into the lift grid

AI works best as a layer in the architecture you already run, not as one more disconnected app with its own login. Business Central can take part through Power Automate, published APIs, Azure services and — since April 2026 — a generally available MCP server, which lets properly configured agents reach Business Central directly.

That is a materially better arrangement than handing everyone another chat window. It is also exactly where integration and security design stop being optional, which is its own run: building an AI integration layer around Business Central.

AI will not fix your base

No amount of intelligence makes unreliable data reliable. If customer records disagree with each other, the item master has been neglected since the last upgrade, integrations quietly duplicate records, customisations fight standard functionality, or nobody trusts the history — then AI will produce confidently wrong answers much faster than your current process produces slow ones.

Before you hand AI the keys to an ERP, understand the ERP.

This is why AI programmes and modernisation programmes belong on the same trip. Conveniently, AI is rather good at finding the technical debt that would have wrecked it.

A roadmap that does not end in a lawsuit

A sensible Business Central AI programme starts on the greens and works up.

Phase 1 — Assist

Summarise, search, explain, analyse, draft, classify. The human decides. Always.

Phase 2 — Recommend

Spot anomalies, propose actions, prioritise work, surface exceptions.

Phase 3 — Automate

Run defined workflows, create records, trigger processes, clear low-risk exceptions.

Phase 4 — Govern

Permissions, approvals, audit trails, monitoring, data governance, evaluation criteria, and a plan for what happens when it is wrong.

Phase four is the one everybody skips, because “governance” does not sound nearly as good in a board update as “we have deployed agents”. It is also the only thing standing between you and an extremely expensive intern with write access to your general ledger and no concept of consequences.

Where to drop in first

Pick one painful process. Do not try to AI-enable the entire ERP. That is straight-lining a double black on day one with rental gear and a GoPro, and it ends the way you think it ends. Look for something that happens constantly, eats real hours, touches more than one system, has fairly predictable outcomes, and can actually be measured.

Build one small AI-assisted workflow. Measure it. Then take another lap. Ten concrete candidates are a good place to start looking.

The actual opportunity

Business Central already holds the transactional foundation of the business. AI can become the layer that helps people understand it, interpret it and act on it.

The companies that get the most out of this will not be the ones with the biggest AI budget. They will be the ones who understand their own processes well enough to see where intelligence actually removes friction — and honest enough to admit which runs they have no business dropping into yet.