LLM access to live business data: how to let AI answer from your systems instead of old exports

You can give an LLM (large language model, the type of AI behind chat assistants) access to live business data without copying that data into the model. The model connects to your systems through a controlled layer. That layer looks up current records when someone asks a question, checks that person's permissions and logs what was read. The model reads and searches your data. It is not trained on it.
Most first attempts go differently. Someone exports a list of open orders to a spreadsheet, uploads it to a chat tool and asks questions. The answers look good for a day. After that, the list is out of date and nobody notices.
What counts as "live business data" for an LLM?
Live business data is the current state of your systems: today's stock, open invoices, order status in your ERP, the latest note on a customer in your CRM. It changes every time someone in your team enters something.
Documents work differently. A policy, a contract or a manual changes rarely, so an LLM that searches documents can work from a stored copy. Records change all the time. An LLM that answers questions about records has to look them up at the moment the question is asked.
Why is uploading an export not enough?
An export is a snapshot. It is correct at the moment it is saved and gets less correct after that. It causes three problems:
- Stale answers. The model answers with confidence from yesterday's numbers.
- More copies of your data. Every export is one more file to secure, share and eventually delete.
- No permission check. The export holds what the person who made it could see. The person asking the question may not be allowed to see all of it.
The same problem shows up without AI. When a spreadsheet becomes the place where the real numbers live, it slows everyone down.
How does an LLM get safe access to live data?
Safe access means the model never touches your systems directly. It asks a layer in between, and that layer decides what it may see. A sound setup covers these points:
- Read through a controlled connection. The model asks for specific records, such as the open invoices for one customer. It does not get a full copy of the database.
- Use the asker's permissions. The answer contains only what the person asking may already see in the source system.
- Keep changes with people. Reading is one thing. When the AI proposes a change, such as updating a record, a person approves it first.
- Log every lookup. You can see who asked what, and which records were read to answer.
- Show the source. The answer points to the record it came from, so a person can check it.
This works best when your data sits in one place and is entered once. If the same customer lives in two systems with two spellings, the model sees both. Fixing that, for example by integrating your ERP and CRM so data is entered once, is often the first step.
Local LLM for business or a cloud model?
Both can be connected to live data in the way described above. The difference is where the model runs and where your data goes while it answers.
A local LLM for business runs on your own hardware, so the data stays in your building. That needs suitable hardware, usually a GPU (graphics processing unit) server. How fast it answers depends on that hardware. Running it also costs money for the hardware, maintenance and updates, so plan a budget for those before you start.
A cloud model needs no hardware of your own, but your data travels to the vendor while it answers. Read that vendor's terms to see what happens to it. The GDPR (the EU data protection law) does not by itself require on-premise AI. It does mean you need clear agreements with whoever processes personal data for you.
Sometimes neither option needs custom work. If a package you already use offers an assistant that answers from its own data, and that covers what you need, use it.
Reporting automation is a good first use case
Reporting automation (in Dutch: rapportage automatisering) is a sensible place to start. Many reports exist only to answer a few fixed questions. Which invoices are overdue? Which orders are late? Which customers have not been contacted this quarter? An LLM with live access can answer those questions from current records when someone asks. Nobody has to rebuild a weekly export first.
The rules from the section above still apply. Keep it read-only, keep the permissions and keep the log. Tasks that go past answering questions, such as AI agents that act on records, call for a separate decision about whether to build custom or use off-the-shelf tools.
Where ERIS fits
Upward Technologies is building ERIS, all-in-one business software that runs on the client's own hardware. ERIS is not available yet. The beta starts in November 2026 and the release is planned for January 2027. Companies can sign up as a design partner.
How to start without guessing
Upward Technologies starts every project with a three-week paid scoping phase before anything is built. It maps which systems hold the data, who may see what and which questions are worth answering. It ends with a document you could take to another supplier and a go or no-go decision. It does not produce a prototype or a working agent. If you go ahead, Upward Technologies builds the solution and can also host and run it.
FAQ
Can an LLM access live business data?
Yes. An LLM can answer from live business data when it is connected to your systems through a controlled layer. That layer looks up current records at the moment a question is asked and applies the asker's permissions.
Is a local LLM trained on my company data?
No. A local LLM reads and searches your company data when it answers a question. It is not trained on that data.
What hardware does a local LLM for business need?
A local LLM needs suitable hardware, usually a GPU server. Its speed depends on that hardware. Running it costs money for hardware, maintenance and updates.
Does GDPR require AI to run on-premise?
The GDPR does not by itself require on-premise AI. If a cloud vendor processes personal data for you, check that vendor's terms and put the right agreements in place.
Is ERIS available now?
No. The ERIS beta starts in November 2026 and the release is planned for January 2027.



