An order arrives as a PDF. Someone in the back office opens it, looks up the address, matches the item numbers, creates the transaction. Two to five minutes, depending on how many lines are on it. At a hundred orders a day, that is one full position doing nothing else. And it is only one of many back office processes that AI can take over today.
Since microtech büro+ gained a GraphQL interface, much of this work can be handed to AI. The open question is how. There are three options, and they differ mainly in how much checking sits between the AI and the ERP, and who is responsible for it.
microtech has laid the groundwork for AI with its GraphQL interface
Microtech büro+ has a GraphQL interface that reads and writes. Addresses, articles, transactions, open items: all of it can be queried, created, changed and posted. That is the door through which an AI can work inside your ERP.
Add the Model Context Protocol, or MCP, the standard through which AI assistants like Claude talk to an ERP. microtech has published open guides for it on GitHub, called skills, which explain to an assistant how büro+ is structured and how a transaction comes about.
One thing is true for all three options: data goes to an AI. The difference lies in which data, who controls it, and what happens before anything is posted in büro+.
Three options for AI automation, compared
| AI assistant over GraphQL | Custom build | Standard solution like Workist | |
|---|---|---|---|
| What you do | connect it, no code | build and run your own software | connect it, set approval rules |
| Time to production | hours to days | months | immediately once connected |
| Data access | everything the linked user can see | whatever you define and secure yourself | limited to the process |
| Check before posting | none | if you build it | on every case, approval threshold adjustable |
| Flexibility for edge cases | high | only what is programmed | high, via WorkI |
| Ongoing maintenance | light, results vary | entirely yours | the vendor's |
| Fits | very small companies, analysis | processes no vendor covers | back offices with recurring work |
Option 1: Connect an AI assistant directly over GraphQL
You connect an assistant like Claude over MCP to the büro+ interface and ask your ERP questions in plain language. Which customers ordered a lot in August and nothing this month? The answer comes from live data. The first time it runs, it is impressive.
The catch: the assistant can do everything the linked user can do, and it decides for itself which data it pulls to answer a question. Whoever operates it sees terms, payment behaviour and transactions of every customer without ever logging into büro+. If it writes a transaction, nobody checks the item number first. For one query a day that does not matter. For a hundred orders it does.
For a five-person business where the owner knows all the numbers anyway, it is a useful tool. A way to automate a process it is not.
Option 2: The custom build
Your developers build their own application on the interface, tailored exactly to your process. That is the right call when your process is truly unique. Companies with a process like that usually know it.
What gets underestimated is rarely the connection but everything after it. Orders arrive in a hundred layouts, product descriptions have to become item numbers, and the ten percent of edge cases eat eighty percent of the time. And what you have built, you maintain: new AI models, new customer forms, the developer who knows it all leaves. For order entry, that means solving a solved problem a second time.
Option 3: Standardised industry solutions for targeted automation
The third option is a solution built for the back office that already knows büro+. Workist is an official Microtech partner and connects over the same GraphQL interface, reading and writing, with no changes to your system.
What you get are proven use cases running at more than 200 companies. The Order Agent reads orders and price enquiries, checks them against your master data and creates orders or quotes in büro+. The Inbox Agent answers customer enquiries straight from the mailbox. The Voice Agent takes orders over the phone. Every case is checked before posting: what is unambiguous goes through, what is not goes to a person for approval.
And beside them stands WorkI, the AI assistant inside Workist. It answers questions about your data and takes on, in chat, whatever does not fit a standard case: drafting a follow-up to a customer, tidying master data, resolving an edge case. The same flexibility as in option 1, but within the access rights you have set for Workist, and with a record of who did what.
That is why most companies end up here. The standard cases run checked and without your involvement, and for everything else there is an assistant that knows your ERP. Data protection is also easiest to establish this way: Workist is a processor under contract, access is limited to the back office, and every step is documented.
Decide which path is right for you
If you are up to ten people, everyone sees everything anyway and a handful of documents arrive each day, plug in an AI assistant.
If your process exists nowhere else, you have your own developers and you are prepared to maintain it for years, build it yourself.
If orders, enquiries and calls come into your back office every day, take a standard solution. You get the proven cases right away and the flexible assistant on top, without handing over a master key to do it.
Questions that come up
Does büro+ need to run in the cloud for AI to reach it?
No. The GraphQL interface works on your own server just as it does in a private cloud. Workist connects the same way in both cases.
Can AI write through the interface as well as read?
Yes. Creating addresses, posting transactions, changing records, all of it is possible. What an access may do is set by you in büro+. The difference between the three options is whether anyone checks before it writes.
Which processes in büro+ can Workist automate?
Creating orders and quotes from emails and PDFs, answering price and availability enquiries, taking orders over the phone, plus ad-hoc tasks via WorkI such as follow-ups and master data maintenance. You will find an overview on the AI platform page for Microtech Büro+.
What sets WorkI apart from an AI assistant plugged straight into the interface?
The boundaries. WorkI works with the same master data and access rights as the standard agents, knows your articles and customers, and logs what it does. A directly connected assistant sees everything the linked user sees, and nobody checks what it writes.
Does data go to an AI with Workist?
Yes, as with any AI solution. The difference: Workist only receives the data the back office process needs, is contractually bound as a processor and is SOC 2 Type 2 certified. Customer data is not used to train public models.
Does Workist require changes to my büro+?
No. The connection runs over the GraphQL interface, with no middleware and no changes inside your tenant. You set up an access for Workist, nothing more.
How long does connecting to büro+ take?
A few weeks. The connection itself is standard; most of the time goes into your master data and agreeing which cases may run through automatically.
What does a custom build actually cost?
The connection is the smaller part. The effort sits in document capture, master data matching, edge cases and permanent maintenance as customer forms and AI models change.
Can the options be combined?
Yes. If you use Workist, though, the back office no longer needs a second assistant, because WorkI covers the ad-hoc cases. A directly connected assistant still makes sense for management reporting.
Read on: The AI platform for sales teams working with Microtech Büro+.
.avif)
