AI agents that query your own systems

AI agent, chatbot, assistant or automation: what is the difference?

The terms get mixed up, and the difference affects cost and risk. A rough guide:

What an AI agent connected to your CRM or ERP can do

Agents earn their keep where the answer exists in your systems but getting it means asking someone, running a report or opening three screens:

• Sales: ‘How much has this customer bought this quarter compared with last year, and do they have unpaid invoices?’ answered before a visit.
• Operations: ‘Which orders due this week are short of stock?’ with the list and the purchase orders already raised.
• Finance: ‘Show me debt over 60 days by account manager’ without waiting for month-end MI.
• Customer service: the agent finds the order, delivery and invoice for a customer on the phone in one question.
• Controlled actions: create a follow-up task in the CRM, draft a quote or prepare a purchase order, always waiting for a person to confirm.

How it works and how it avoids making things up

An agent is a language model with a set of tools: specific, tested queries against your systems (‘get open orders for customer X’, ‘get stock for item Y’). The model decides which tool to use; the data comes from your system, not from the model's memory. If no tool returns an answer, the agent says it does not know.

Every answer shows where the data came from, with a link to the record, so anyone can check it. We start read-only. Actions are added one at a time, each with a confirmation step. Before the team uses it, we test it against a set of real questions with known answers, and every question and answer is logged so problems can be traced.

Permissions: the agent sees what the user can see

The safest rule is that the agent works with the identity and permissions of the person asking. If a sales rep cannot see margins in the ERP, the agent cannot show them either. Microsoft's own design for Business Central is a useful example: its MCP server lets AI clients such as Copilot Studio, VS Code or other MCP-compatible tools work with Business Central data; by default it exposes read-only API access, administrators choose which API pages agents can use to create or modify data, and operations run under the signed-in user's identity, so audit trails show who did what.

So an AI agent that queries Business Central can be built on Microsoft's own route or as a bespoke service, depending on what else it has to reach. The same principle applies to Xero, Sage, HubSpot or Salesforce through their APIs, and to older systems through a controlled read layer. Data sent to AI models is handled as described on our private ChatGPT for business page. If the agent will talk to customers in the EU, the EU AI Act's Article 50 requires that people are told they are dealing with an AI system unless it is obvious; we design for that, and your adviser assesses your case.

Copilot Studio or a bespoke AI agent

Microsoft describes Copilot Studio as a low-code studio for building agents, with prebuilt and custom connectors, knowledge sources and publishing to Teams, websites and other channels. If you are a Microsoft business, try it first:

• If the problem is getting systems to talk at all, start with ERP and CRM integration; an agent sits on top of that.

How we would build it and what drives cost

We start with one team and ten to twenty real questions they ask every week, answered read-only. You see a working demo every two weeks, you get one fixed price for the project, and the code is yours, in a repository in your company's name. The rest is on how we work. What moves the cost most:

• Systems connected and how good their APIs are.
• Permission rules: inheriting them from one system or defining your own.
• Actions: each write action needs its own confirmation flow and tests.
• Channel: web, Teams, Slack or inside another app.
• Usage: AI model costs depend on how many questions are asked; market references are in our guide to the cost of implementing AI in a business.