By Laas NuryEvery conversation about AI in a company arrives at the same question sooner or later: "We already have Copilot. Or ChatGPT. Why would we need anything else?"
It is a good question, and often the honest answer is: you don't, yet. Off-the-shelf assistants are cheap, quick to roll out and genuinely useful. But they do one kind of work well, and a lot of the work that costs a company the most time is a different kind. This article compares the three routes plainly, and gives you a way to tell which one your case needs. We build custom solutions for a living, so read it knowing that; we will still tell you when you do not need one.
The short answer
- If the work is one person doing a task (writing, summarising, searching, drafting, analysing a file), start with Copilot or ChatGPT. A custom build would be overkill.
- If the work is a process (every order, every invoice, every planning run, handled the same way, with your systems, your rules and your exceptions), an assistant will help the person doing it, but it will not run it. That is where a custom build belongs.
- Most companies end up with both: assistants for everyone, and one or two custom builds on the processes where the volume and the money are.
The three routes, briefly
Microsoft 365 Copilot lives inside Outlook, Word, Excel, Teams and the rest of Microsoft 365. It works with the e-mails, chats and documents the user already has permission to open. Its strength is that it is where your people already work.
ChatGPT (Business or Enterprise) is a general assistant in its own window, with strong writing and analysis, file uploads and custom GPTs your team can set up. Its strength is breadth: it is good at a very wide range of tasks.
A custom AI build is a solution made for one of your processes. It reads from and writes to your own systems (ERP, CRM, planning, your mailbox), follows your rules, sends exceptions to the right person and runs whether or not someone opens a chat window. Its strength is that it does the work instead of helping with it.
Side by side
| Copilot | ChatGPT | Custom build | |
|---|---|---|---|
| Best at | Helping one person in Microsoft 365 | Helping one person with any text or analysis task | Running one process end to end |
| Starts when | Someone asks | Someone asks | Something arrives (an e-mail, an order, a date) |
| Works with your systems | Microsoft 365 content; others through connectors | Through uploads, connectors and custom GPTs | Built into them: reads and writes where the work happens |
| Your rules and exceptions | In the prompt, each time | In the prompt or a custom GPT's instructions | Built in, tested, the same every time |
| Data | Within Microsoft 365's service boundary, not used for training | Business plans: not used for training | Where you decide, including EU-only hosting |
| Cost | Licence per user per month (about EUR 20 to 30 at 2026 list prices, plus a Microsoft 365 licence) | Licence per user per month (about EUR 20 to 30 at 2026 list prices) | A project, then a monthly running cost |
| Time to start | Days | Days | Weeks for a well-defined process |
| Where it stops | Anything outside Microsoft 365, and work that must run without a person | Anything that must write reliably into your systems | Work too varied to define: that stays with people and assistants |
When Copilot or ChatGPT is enough
Stay with an assistant, and put your effort into good use rules and training, when:
- The work varies every time and depends on judgement: a proposal, a policy, an analysis for a board meeting.
- One person does it, occasionally, and checks the result anyway.
- Nothing has to be written back into a system automatically.
- The volume is low. Ten documents a week do not need a process.
For this work, a custom build would cost more than it saves. Buy the licences for the people who will use them (not everyone), and spend a day setting up how they should be used.
When you need a custom build
These are the signals we look for. Two or three together usually settle it:
- It is a process, not a person. The same kind of work arrives again and again: orders, invoices, claims, delivery changes, planning requests.
- It lives in your systems. The work is only done when something is booked in the ERP, updated in the CRM or changed in the planning, not when a text is written.
- It has rules and exceptions. Customer-specific prices, approval limits, what to do when the article number does not exist. An assistant needs to be told every time; a build knows.
- Nobody should have to start it. It should run when the e-mail arrives at 06:40, not when someone opens a chat.
- You need to prove what happened. An audit trail of what the AI read, decided and changed, and who approved it.
- The accuracy can be measured. You can check whether an order was booked correctly, so you can test the build before it goes live and watch it after.
The middle route: build-it-yourself assistants
Copilot Studio agents and custom GPTs sit between the two. Your team configures an assistant with instructions, knowledge files and sometimes connectors to other systems.
They are a good way to learn and fine for simple, contained jobs, such as an assistant that answers questions from your employee handbook. They get difficult when the assistant has to write into business systems reliably, handle the exceptions, keep an audit trail and keep working when the colleague who built it moves on. That is where most do-it-yourself assistants stall: not at the demo, but in the third month.
The costs people forget
With Copilot: it shows users everything they already have access to. In many companies that includes folders nobody meant to share widely. Cleaning up permissions before a wide roll-out is real work, and a common reason roll-outs get paused.
With ChatGPT: the time spent writing and re-writing prompts, and the copy-paste between the chat and your systems. If a task needs twenty pastes a day, that time is the cost.
With a custom build: the running cost and the maintenance. A build is a small piece of your operation that needs an owner, monitoring and an occasional change when your process changes. Budget for it from the start, or it will quietly degrade.
An example: order intake at a wholesaler
Take an illustrative case: a wholesaler receives a few hundred orders a week by e-mail, as PDFs, spreadsheets and plain text, and an order desk types them into the ERP.
- ChatGPT can read an order when someone pastes it in and turn it into a clean list. Useful, but the person still does the pasting, the checking and the typing.
- Copilot can summarise the order inbox and draft replies. Useful, but the booking still happens by hand.
- A custom build reads each order as it arrives, matches the customer and the article numbers against the ERP, applies the customer's agreed prices, prepares the order in the ERP, and puts only the exceptions in front of a person: the unknown article, the price that does not match, the delivery date that cannot be met.
The first two make the order desk a little faster. The third changes what the order desk does all day. That difference is the whole decision.
Six questions to decide
- Is this a task one person does, or a process that repeats?
- Is the work finished when a text is written, or when something changes in a system?
- How many times a week does it happen?
- What are the rules and the exceptions, and could you write them down?
- Does anyone need to prove later what was decided and why?
- What does an error cost?
If your answers are "one person, a text, a few times, few rules, no, little", buy the assistant. If they point the other way, it is worth looking at a build.
One regulatory note
When you build an AI system and put it into use under your own name, you become its provider under the EU AI Act. For most operations work, such as orders, invoices and planning, that adds little, because those uses are not high-risk. It matters more if the process makes decisions about people. Our overview of the EU AI Act for companies that use AI explains the difference.
