Conversational BI on your ERP database

The ERP holds most of what leadership wants to know: sales, inventory, payables and receivables, payroll, production. The data is already there. What usually goes missing is a way to reach it without a queue.

ERP reports are fixed. Someone decided years ago which columns, filters and totals were worth building, and any question outside that design turns into a request to IT, to the system vendor or to the analyst who can write a query. The question a manager asks on Wednesday morning is usually the one no report answers: which customer bought less this quarter than last, which items have sat in stock longer than turnover justifies, which invoices fall due this week and cluster around a few customers.

Asking the database directly solves that without a spreadsheet export and without building a data warehouse first. A person types the question in plain language, Entendo generates the SQL query, runs it in read-only mode and returns a number, a table or a chart. The SQL and the reasoning show on screen, so someone who knows the business can check how the answer was built.

This does not replace the ERP. The system remains the place where operations happen. Entendo sits beside it, only reading, for the questions the reports do not cover. For the broader concept, see what conversational BI is.

The path from question to database has three parts.

The first is Entendo Gateway, an agent installed on the company's network, near the database server. It opens an encrypted outbound connection to Entendo, so the company does not open an inbound firewall port or whitelist an IP address. The database credential stays inside the company's network, and what travels to the cloud is the query result.

The second is read-only access. Before running any SQL, Entendo blocks write commands. We recommend creating a read-only database user for Entendo, so the restriction holds on the database side as well. Every question and every SQL statement is logged for audit.

The third part is the semantic layer, and it decides whether an answer is any good. ERP databases store data in tables and columns with technical, abbreviated or inherited names. A value column may carry a name only the person who deployed the system understands. The semantic layer translates those names into business terms: what "revenue" means, which table holds the customer master, which filter separates a closed sale from a quote, what counts as "canceled". It also stores targets and indicators the AI uses to interpret the question.

Building that layer is the core of implementation work, and there is no point pretending otherwise. Without it, an AI writes plausible SQL over columns it does not know and gets things wrong convincingly. With it, the question "how much did we bill in September by salesperson" lands on the right tables and filters for that system. That is why implementation is assisted by our team and billed separately.

Entendo connects to PostgreSQL, MySQL, SQL Server, Oracle, Firebird and ClickHouse, plus Google Sheets, CSV and Excel. The table below organizes the most common ERP scenarios. None of them relies on an official vendor connector: today the connection goes to the database, and for systems without database access we build the integration on demand.

We do not claim which database each product uses in every installation. It varies by version and by customer choice. Confirm with your vendor or your IT team before you plan.

Connecting to a production database calls for a few simple precautions. Agree on all of them with IT before the first query.

• A read-only credential. Create a database user with query permission only, limited to the tables Entendo needs to see. Entendo blocks writes, but a restriction in the database is the second lock.
• Load on the production database. Analytical queries can be heavier than the ERP's day-to-day traffic. When a read replica or a recent staging copy exists, pointing Entendo at it takes pressure off production. When none exists, agree on time windows and watch performance during the first weeks.
• Permissions per person. Decide who can ask what. Payroll and margin by customer should not show up for every user.
• Data protection. Personal data about customers, suppliers and employees sits in much of an ERP database. Define with your data protection officer which tables and columns are included, what stays out of the AI's reach and who has access. The gateway keeps the credential on the company's network, but query results travel to the platform, and that belongs in the company's privacy review.
• Auditing. Because every question and SQL statement is logged, the company can review what was queried.

Implementation follows a predictable order, even though the timing varies with the size and organization of the database.

1. Access. Install the gateway on the network, create the read credential and confirm that Entendo can see the database.
2. Table mapping. Identify the tables that matter for the first questions (sales, customers, products, finance) and understand how they relate.
3. Glossary. Record the business terms and the rules that define each one, with help from the people who use the ERP every day.
4. Validation. Ask questions whose answers the team already knows and compare. Where they differ, adjust the semantic layer. This step repeats until the team trusts the numbers.
5. Rollout by area. Open access to sales, finance or operations, one at a time, with the agreed permissions.

The most important part of validation is having someone nearby who knows the number for real, because only that person can tell whether the answer is right. We do not promise a timeline before we see your database. Schema size, naming quality and your team's availability weigh more than any generic estimate.

These examples show the kind of question Entendo answers once the semantic layer covers the right tables. Answers depend on each company's data and rules.

• Sales: which salespeople sold less this month than last month? Which active customers have not ordered in over ninety days? What is the product mix by region?
• Finance: which invoices fall due this week? How do receivables look by days overdue? Which suppliers account for most of this month's payables?
• Inventory and purchasing: which items have been in stock the longest? Which products have coverage below the reorder lead time? Which supplier delivers latest?
• Operations and logistics: how many orders shipped late, by carrier? What is the return rate by product?
• Leadership: how do this quarter's sales compare with the same period last year? What does the forecast say about the coming weeks, based on history?

Each question can become a chart, join a dashboard with filters and drill-down, and be exported as CSV, XLSX, PNG or PDF. To sell this to your own customers under your brand, see the white-label model for software vendors and consultancies.