The best place to drop in is nearly always duller than the brochure. Marketing wants to sell you the heli drop. The money is on the cat track everybody rides forty times a week without thinking about it. So find the thing your people do over and over, then ask the unglamorous question: could AI do part of this reliably?
Business Central is stacked with candidates, most of them hiding in plain sight. Here are ten worth a look, roughly in the order they tend to justify the lift ticket.
1. Invoice processing
Accounts payable is where hours go to die. AI can pull the information off the document, classify it, work out which vendor it belongs to, line it up against ERP records and have a transaction sitting ready for approval before anyone has got their gloves off.
Note the verb: prepare, not blindly post. Microsoft’s Payables agent works in this territory and is generally available — which is not the same as saying it should run without review on day one. Generally available means the lift is open. It does not mean the run is in condition.
2. Customer and vendor communication
AI can draft the message straight out of Business Central data — say, an email to every customer whose order is running more than five days late, assembled from the customer, the order, the item, the shipment and the expected delivery date. The work was never the writing. The work was going and finding all five of those things.
A person approves before anything sends. Always. The only thing worse than a late order is a confidently wrong apology for one.
3. Inventory exception analysis
Inventory is a natural, because the volume beats manual review the way sheet ice beats a brand new edge. Nobody is really scanning hundreds of lines, whatever the process document claims. Instead of asking someone to do that, AI can surface unusual demand, overstock, stockouts, slow-moving lines, sudden purchasing changes and adjustments nobody expected.
The goal is not letting AI decide what to buy. It is making sure the human who does decide sees the important exceptions first, rather than meeting them in a month-end variance.
4. Accounts receivable prioritisation
A workflow can weigh outstanding receivables against amount overdue, customer history, payment behaviour, credit status, recent communication and order activity, then hand back a prioritised collection list. Which is a meaningful upgrade on the traditional method: chase the biggest number, or whoever shouted most recently.
5. Failed integration diagnosis
A Business Central environment is typically wired into CRM, e-commerce, payroll, shipping, payment providers, warehouse, tax and manufacturing systems. That is a lot of lift cable. When one of them stops, somebody has to bootpack up there and work out why.
AI can do the collecting — error messages, payloads, timestamps, related records, integration logs — and then point at the pattern running through all of it. A human still calls it. They are just not the one spending half an hour picking gear out of the trees first.
6. Job queue monitoring
Job queues throw off noise the way a gondola throws off fogged goggles. The distinction worth automating is the one people are worst at making at seven in the morning: is this an expected transient failure, or is something systemically broken?
7. Sales opportunity analysis
AI can read customer and sales history for declining or increasing purchasing, dormant customers, unusual order patterns, cross-selling opportunities, lost customers and high-value accounts that have gone quiet. That last one is the expensive one, because nobody escalates a customer who has stopped complaining.
The output is not a prediction, and anybody selling it as one is straight-lining. It is a shortlist of where a salesperson should spend Tuesday.
8. Report and data explanation
Plenty of ERP users can tell you exactly what a report says and absolutely nothing about why it says it. Ask why gross margin slid this month and AI can go through sales, costs, product mix, discounts, vendors and inventory adjustments and summarise the likely contributors — the difference between watching the crash back on video and being told which edge you caught.
9. ERP documentation
AI can draft and maintain documentation for integrations, workflows, extensions, reports, business processes, configuration, custom fields and APIs. Nobody has ever volunteered for this job, in the history of ERP. Which is exactly what makes it a good candidate.
It pays out disproportionately during modernisation, when “what does this actually do?” becomes the most expensive question in the room. It is also the natural way in to finding technical debt.
10. Daily executive briefings
A five-minute ERP briefing waiting at first light — yesterday’s revenue, orders, new customers, inventory exceptions, ageing receivables, failed integrations — followed by the two or three items that genuinely need a person. Call it the morning snow report for your balance sheet.
None of which is science fiction. It is an orchestration problem, and mostly it reduces to one unromantic question: is the data reachable through APIs?
Start with human-in-the-loop
Every one of these ten can sit at a different level of autonomy, and the level matters far more than the use case does, because the level is what decides how far you fall.
| Level | AI role | Who is accountable |
|---|---|---|
| Assist | Suggests | The person |
| Analyze | Explains | The person |
| Recommend | Proposes an action | The person |
| Approve | Waits for validation | The person, explicitly |
| Execute | Performs the action | The process owner |
| Autonomous | Handles the defined process | The process owner, with audit |
What makes a good first run?
The best candidate usually has four things going for it: it happens frequently, the expected result is reasonably predictable, employees burn meaningful time on it, and you can measure the improvement. Skip that fourth one and you have not automated anything — you have moved the work somewhere nobody is counting.
A process that takes five people two hours every Monday is a far better opportunity than some obscure ritual performed twice a year, however much more interesting the obscure one is to build. Ride the lap that is actually open.
Do not groom a rock field
If a process is broken, AI will help it be broken faster and at greater volume. That is not transformation. That is industrialised confusion, billed monthly.
So before automating anything, write down the trigger, the inputs, the decision points, the business rules, the exceptions, the required approvals, the final action and the audit requirements. If that document turns out to be hard to write, congratulations: you have found the real problem, and it was never a technology one.
Then work out which parts actually need AI. Some of this is standard Business Central functionality already doing its job. Some belongs in Power Automate. Some needs AL or an API. AI earns its place where interpretation or adaptive judgement adds something — a line drawn more carefully in Copilot versus custom AI.
