The moment a company starts poking at AI in Business Central, the same question comes skidding down the hill: do we use Microsoft’s built-in Copilot, or do we build our own?
The honest answer is usually both, for different runs. Microsoft’s current model spans Copilot for assistive experiences and agents for more autonomous process work, and Business Central also supports custom AI and external agent architectures. Three tools, three jobs. Nobody sensible straight-lines all three on day one.
Copilot: the buddy riding along
Copilot is the freshly groomed blue at 9am — the obvious first lap when you want help inside Business Central and you want it without building anything. Current Copilot capabilities include chat, summarise, analysis assist, autofill, bank reconciliation assist, sales line suggestions and marketing text suggestions.
All of that suits work where a human is still the one holding the decision. Copilot rides along and points at things. It does not take the wheel, and it does not carry the blame.
Choose Copilot when
- The task is already supported
- You want minimal development
- Users need assistance rather than autonomy
- The workflow lives inside Business Central
- Microsoft-supported functionality meets the requirement
Agents: one run, one job
Agents step past assistance and actually handle defined tasks with limited human input. Microsoft ships the Sales Order, Payables and Expense agents as generally available capabilities, with a configurable Custom Agent in preview.
The load-bearing phrase there is defined tasks. An agent with woolly instructions and broad ERP permissions is a first-timer pointed down a double black because nobody wanted to be rude about it. What you want is a constrained worker with one run, one job and a visible finish line.
Custom AI: where the map runs out
Custom AI starts looking attractive when the problem is nowhere on the trail map. A manufacturer might want AI to weigh production schedules, inventory, customer demand, vendor lead times, capacity and historical sales, then recommend production adjustments.
That is not a standard Copilot scenario. That is sidecountry — and sidecountry means you bring your own architecture, your own kit and your own plan for getting back.
Five questions before you drop in
Five questions, in order, the way you check edges and bindings before the first lift. Most requirements never make it past question two.
1. Does Business Central already do it?
Then use standard functionality. Building AI because you can is a gaper move with a purchase order attached.
2. Does Copilot already solve it?
Then start there. It is a free lap. Take it.
3. Is the process repetitive and well defined?
Consider an agent — or ordinary automation, which grooms the same run for a good deal less money.
4. Does it require unique business intelligence?
Now custom AI is genuinely on the table.
5. Does it need information from several systems?
Then you have stopped choosing an AI tool and started doing integration architecture, which is a different discipline with its own design requirements and its own ways to put you in the trees.
One quiver will not cover the whole mountain
In real environments this is almost never a single choice. It is a layered kit, for the same reason nobody takes one pair of skis to every condition:
| Layer | What it is good at |
|---|---|
| Business Central Copilot | Assisting employees in context |
| Business Central agents | Defined ERP processes |
| Power Automate | Workflow orchestration |
| Business Central APIs | Structured, scoped data access |
| Azure and AI services | Custom intelligence |
| CRM, warehouse, other systems | Broader enterprise context |
Layer it like that and each technology gets to do the thing it is actually good at — while the expensive, high-maintenance option stays in the bag until a problem genuinely earns it.
Custom AI is a season pass, not a day ticket
Build your own and you own development, hosting, security, authentication, monitoring, model selection, prompt management, evaluation, data governance, maintenance, failure handling and version changes. All of it. Every season.
That is a standing commitment, not a project. It should be closing a real gap, not decorating a board slide.
When custom AI does make sense
- The process is unique. Your advantage comes from specialised business logic.
- The data crosses systems. The answer needs more than Business Central.
- You need specialised reasoning. The analysis is not available as standard.
- You need a custom experience. It has to work through Teams, a portal or an internal app.
- You need specific controls. Approval, security or governance behaviour is non-negotiable.
The better question at the top of the lift
Not “Copilot or custom AI?” but what is the smallest technology stack that solves this problem reliably?
That question builds better architecture, because it makes you look at the run instead of the gear. Sometimes the answer is Copilot. Sometimes a Business Central workflow. Sometimes Power Automate. Sometimes an agent. Sometimes custom AI.
And sometimes it is a spreadsheet — deeply unfashionable, still parked at the bottom of the hill, and still, annoyingly, correct.
