AI on the systems you already have: ERP, CRM and documents

The most useful AI in a company is rarely a separate application. Usually it is a small part built in where people already work: in the ERP, the CRM, the mailbox, the document archive. That is why "which model" matters less than "how will it connect to what we have".

Start from what you actually have

According to Eurostat, in 2025 53% of EU enterprises used specialised business software (ERP, CRM or BI). In Bulgaria the share was 31%, the lowest in the EU. In many Bulgarian companies the important information lives in spreadsheets, emails, folders and on paper.

That is not an obstacle but a starting point. AI can work with documents and email too. But integration starts with an honest map: where the data is, who maintains it and which version is the right one.

The four places where AI connects

Documents. Contracts, internal policies, technical documentation. The assistant searches them and answers with a reference to the source, so a person can check.

Incoming flow. Emails, requests, invoices. AI reads them, classifies them and extracts data, and a person approves.

Business systems. ERP and CRM through their programming interface (API). AI reads a customer, an order or stock levels and prepares the next action.

Writing back. The most sensitive step: AI creates or changes a record in the system. Always start here with human approval.

What has to be ready

AI-ready data. Gartner predicts that through 2026 organisations will abandon 60% of AI projects not supported by such data. In its survey 63% of organisations either do not have, or are unsure whether they have, the right data management practices. In practice "ready" means a clear source of truth, up-to-date records and a known owner.

Access rights. An AI assistant should see only what the user asking it can see. If someone in the warehouse cannot open the payroll, the assistant must not be able to show it to them either. This is the most common gap in quick rollouts.

Personal data. If the solution processes personal data, the model provider usually acts as a processor and you need a contract with it under Article 28 of the GDPR. Also check where the data is processed and whether it is used to train models.

A trail. Every AI answer and action should be logged: what was asked, what came back, who approved. Without it, errors cannot be investigated.

Three steps: read, suggest, act

The safest path has three stages. First AI only reads and answers, changing nothing. Then it suggests an action and a person approves it with one click. Only when its suggestions have been right for long enough can some of them run automatically, for defined cases and within clear limits.

Each step brings value, and the risk grows only once trust has been proven with data.

What usually breaks

Rights inherited without review: the assistant sees more than it should. Stale data: the document index is not refreshed and answers quietly become wrong. Changes in the systems: a new ERP field, a new API or model version, and the integration stops. No monitoring: nobody notices when quality drops.

MIT's 2025 study reaches a similar conclusion: in failed projects the problem is rarely the model itself. More often the solution is not built into the workflow and does not learn from it.

What it costs

In most projects integration is the largest one-off line, larger than the model itself. Every system has its quirks, and rights and data have to be put in order. We cover costs in detail in the article "What AI implementation costs and what drives the price".

When you do not need integration

If your people mainly need help with writing and summarising, a ready-made business tool without integration may be quite enough. Integration makes sense when a task repeats many times and needs data from your systems.

If you would like to look together at which of your systems are ready and where it makes sense to start, write to us. The first conversation is about the map, not the quote.