Short answer: Copilot and custom AI differ less in model capability — often the same models sit underneath — than in who designed the process. A Microsoft feature runs a scenario Microsoft designed for every customer. A custom solution runs yours. Start with Copilot: it costs a fraction and covers real ground.
Terms first
Confusion here starts because "Copilot" covers several different things that get mixed up in conversation.
- Copilot features inside Business Central — assistance on pages, reconciliation proposals, generated item descriptions. These are prompts in the interface, not agents: they do not run multi-step work on their own.
- Microsoft's agents in Business Central — separate entities that run a process end to end. Microsoft keeps extending this set, so the current list is worth checking in their documentation rather than in anyone's marketing, this article included.
- Copilot Studio — a tool in which you build the agent yourself. Formally that is already custom development, just with a lower barrier to entry.
- Microsoft 365 Copilot — a separate product over mail, documents and meetings. Only indirectly related to Business Central.
When someone asks "Copilot or custom", they usually mean the second item against the fourth: a vendor's ready-made agent against one designed around your company.
What Copilot covers
The vendor's logic is sound: cover processes that look the same in most companies. Typical inbound order handling, bank statement reconciliation, standard document preparation — differences between companies are small here, and a product feature is economically justified.
Its advantages are real and underrated. It is maintained by the vendor, survives updates without your involvement, sits inside the licensing model and requires no project. If your process matches the one Microsoft designed, take it — and a discovery that ends with that conclusion has done its job. Our detailed look: Copilot in Business Central.
The limitation is equally structural: a vendor's agent implements an average process. It does not know that in your company orders from category B customers are approved by a division head rather than a manager, or that one supplier's lead times are always calculated with an allowance for their habitual delay.
Where it stops
A practical marker that takes five minutes to check. If you cannot describe the process without using a term specific to your company, a product feature will not run it.
"Process an inbound order" can be described in general terms, and a product solution fits. "Process an inbound order, check the customer is not on our stop list by our criteria, and if the amount is above the division limit send it to the division head unless there is a framework agreement" is a description in which three things exist only at your company.
The second boundary is data outside Business Central. If the decision needs a look into a production system, another CRM or an external service, a product feature will not reach it. A custom one will, but that is an integration project and should be scoped as such.
The third is access rules. Where different roles must see different data, that is solved by your integration's authentication model — see security, permissions and data residency.
Comparison
| Criterion | Microsoft agent | Custom agent |
|---|---|---|
| Who designed the process | Microsoft, for all customers | You, for your specifics |
| Your terminology and rules | Standard | Yours |
| Data outside Business Central | Microsoft ecosystem | Any reachable system |
| Stopping points | As the vendor anticipated | Where you decided |
| Updates | Microsoft | Checked after releases |
| Getting started | Switch on and configure | Discovery, build, testing |
| Responsibility for behaviour | Vendor | You and your partner |
Different cost structures
Comparing the figures directly does not work, because these are different kinds of spend. A vendor agent is licensing on Microsoft's terms: predictable, no upfront investment, scaling with usage. A custom solution is a project with a one-off cost (from $3,000 for an assistant, from $6,000 for an agent) plus running costs for hosting and model usage.
So the decision is not made on price. A product agent that does not run your process is zero result at any cost. Custom development duplicating an existing feature is spend without benefit. The question is always whether the process the vendor designed matches yours. Detail: how much AI for Business Central costs.
How to decide in half an hour
Five questions instead of general reasoning. Three or more "yes" answers mean a product agent will not be enough.
- Does the process description contain terms that exist only in your company?
- Does the decision need data from systems outside Business Central?
- Must different roles see different data within this process?
- Does your process have steps in an order that is not the obvious one?
- Are there exceptions frequent enough that they have to be handled?
If all five are "no", you almost certainly do not need custom development, and an honest partner should say so. The cheapest possible move is to switch on what is already licensed and use it for a month with real questions — the list of what it does not cover becomes your specification, written for free and from facts.
Frequently asked questions
Will Microsoft's agents replace custom solutions over time?
Some of them, and that is fine: every scenario the vendor closes gets cheaper for everyone. But a product agent by definition implements a process designed for all customers and does not know your internal rules. That boundary moves rather than disappears.
Can we run both at once?
Yes, and it is the most common working setup. Product agents cover high-volume standard operations, custom ones cover the narrow places where your specifics create value. They sit over the same database and do not conflict.
How is Copilot Studio different from custom development?
Copilot Studio is a platform where you assemble an agent from ready-made blocks. That is already custom development, with a lower barrier and less flexibility. Enough for simple scenarios; not for complex logic or non-standard integrations.
We are on Business Central on-premises. Does that change things?
Substantially. Microsoft's cloud AI capabilities target cloud environments, so the available set is narrower on-premises. Custom integration is possible but needs web services published and reachable, which is checked before quoting.
Where do we start on a limited budget?
Switch on everything already included in your licensing and use it for a month with real tasks. What it fails to cover becomes the specification for a custom solution — assembled from facts and at no cost.