How can an AI agent work with business software?

Your employees spend a significant chunk of their day copy-pasting data between systems that were never designed to talk to each other — ERP, CRM, FileMaker, email, spreadsheets. Every handoff is a chance for a mistake, a delay, or a dropped ball. This article explains exactly how AI agents plug into existing business software, what they actually do inside those systems, and what it takes to make them work reliably in practice.

What is an AI agent, and how is it different from a simple chatbot or automation?

A chatbot answers questions. A workflow automation tool moves data from A to B when triggered. An AI agent does something more interesting: it reasons about a goal, decides which steps to take, uses tools to interact with systems, and adjusts when something unexpected happens — all without a human scripting every branch.

In practical terms, an AI agent might:

• Read an incoming order from FileMaker
• Check stock levels in your ERP
• Decide whether to fulfill immediately or flag a shortage
• Create the order record, update inventory, and message the sales rep on Slack
• Log every action it took, with a reason

No human clicked anything. No fixed workflow broke because the ERP returned an unexpected field. The agent handled it.

The underlying technology — large language models (LLMs) combined with tool-calling, memory, and orchestration frameworks — is what makes this possible. But from a business operations standpoint, the right mental model is: an AI agent is a junior employee who can use your software, follow policies, and know when to escalate.

How does an AI agent actually connect to business software?

This is where most articles stay vague. Here's what the connection actually looks like in practice.

1. Via APIs

Most modern ERP and CRM systems — Exact Online, SAP Business One, HubSpot, Salesforce — expose REST or SOAP APIs. An AI agent calls these APIs the same way a web app would: authenticate, send a request, parse the response, decide what to do next. The agent's "intelligence" sits on top of ordinary API calls.

2. Via database access or middleware

Systems like FileMaker expose data through the FileMaker Data API, ODBC connections, or custom middleware layers. An agent can query records, create new entries, or update fields directly — treating FileMaker as a structured data source it can read from and write to.

3. Via email, file systems, and document parsing

Not everything lives in a clean database. An agent that processes incoming invoices typically reads emails, extracts PDF attachments, uses OCR or document AI to parse line items and totals, and then compares those against purchase orders in the ERP. The "connection" here is a combination of email access (via IMAP or Microsoft Graph API) and document intelligence.

4. Via RPA (robotic process automation) as a fallback

For legacy systems with no API, an agent can use RPA tooling to interact with the UI the same way a human would — navigating screens, filling fields, clicking buttons. This is a last resort (brittle, slow, breaks on UI changes), but it means almost no system is completely unreachable.

Three real scenarios where AI agents change day-to-day operations

Scenario 1: Order processing across FileMaker and ERP

The manual version: A sales rep enters a new customer order into FileMaker. Someone else sees the order, opens the ERP (say, Exact Online), creates the order there manually, checks the warehouse module for stock, and sends an email to the sales team. This takes 15–30 minutes per order and introduces transcription errors on roughly 3–5% of orders — wrong quantities, wrong customer codes, missing product references.

With an AI agent: The moment an order is entered in FileMaker, the agent reads the new record via the FileMaker Data API. It validates the customer against the ERP's customer master, checks real-time stock, creates the sales order in the ERP with the correct line items, updates the inventory reservation, and posts a Slack message to the sales channel: "Order #4821 for Acme BV confirmed. 48 units of SKU-774 reserved. Estimated ship date: Thursday." The whole cycle takes under 60 seconds. Exceptions — a customer not found in the ERP, a stock shortfall — are routed to a human with full context already attached.

Scenario 2: Automated invoice processing and exception routing

The manual version: Accounts payable receives 200 purchase invoices per month by email. Each invoice is downloaded, opened, checked against the relevant purchase order in the ERP, approved or queried, and posted. A full-time employee spends roughly 40% of their week on this.

With an AI agent: The agent monitors the AP inbox. When an invoice arrives, it extracts the supplier, invoice number, amount, line items, and VAT from the PDF using document AI. It queries the ERP for the matching purchase order and three-way matches: purchase order → goods receipt → invoice. If everything aligns within tolerance (e.g. within 2% and under €500 variance), it posts the invoice automatically. If not — say the invoice total is €1,200 higher than the PO — it drafts an exception summary and routes it to the finance manager with one click to approve or reject. Straight-through processing rates of 70–80% are realistic for structured invoice flows, meaning most invoices never need a human to touch them.

Scenario 3: Customer support with CRM and task management integration

The manual version: A customer emails asking about a delayed delivery. The support agent searches the CRM for the customer record, checks the order status in FileMaker or the ERP, formulates a reply, manually updates the CRM ticket, creates a follow-up task for the account manager, and sends a notification. Six steps, six minutes, repeated dozens of times a day.

With an AI agent: The incoming email triggers the agent. It searches the CRM (e.g. HubSpot) for the customer, retrieves the open order from FileMaker, checks the shipping status via the carrier API, and drafts a personalized reply: "Hi Sarah, your order #7341 is currently with DPD and is expected tomorrow before 17:00. Your account manager Tom has been notified." It updates the CRM ticket, creates a follow-up task for Tom due in 24 hours, and notifies him via Teams. If the query is ambiguous — a complaint, a legal question, a complex return — it flags for human handling without sending anything. The support agent sees only the exceptions that actually need them.

What does an AI agent need to function reliably in business software?

Deployment failures almost always trace back to one of these five things being underestimated:

1. Clean, structured data in the source systems. An agent reading FileMaker records that are 30% incomplete, inconsistently formatted, or full of workarounds will make bad decisions. Garbage in, confident garbage out — LLMs don't flag bad data, they reason around it.
2. Well-defined policies and decision rules. The agent needs to know: what's an exception? What's the tolerance for auto-approval? When should it never act without a human? These rules need to be written down and encoded, not left implicit.
3. Reliable API access with proper authentication. OAuth tokens expire. Rate limits get hit. APIs return undocumented errors. A production AI agent needs error handling, retry logic, and monitoring — not a prototype-quality integration.
4. A human-in-the-loop design for edge cases. The question isn't whether edge cases will happen — they will. The question is how the agent recognizes them and what it does: pause, explain, and route, rather than guess and act.
5. Audit logging and explainability. Every action the agent takes in a business system should be logged: what it read, what it decided, what it wrote, and why. This isn't optional — it's what lets you trust the system and investigate when something goes wrong.

How do you get started? A practical checklist

Before building or buying an AI agent integration, work through this:

Readiness check:

• Can you describe the current manual process end-to-end in writing? (If not, map it first.)
• Do the systems involved have accessible APIs or data connections?
• Is the data in those systems consistently structured and reasonably complete?
• Can you define clear rules for what the agent should do autonomously vs. escalate?
• Is there a designated owner (person or team) responsible for monitoring the agent?

Scoping the first use case:

• Choose a process that is high-volume, rule-based, and currently causing measurable pain.
• Start with one system-to-system flow (e.g. FileMaker → ERP), not an end-to-end multi-system orchestration.
• Define success metrics upfront: straight-through rate, error rate, time saved per week.
• Plan for a parallel-run period where agent output is checked against the manual process.

Technical prerequisites:

• API credentials and documentation for all connected systems
• A sandbox or test environment (do not prototype against production data)
• Logging and alerting infrastructure before go-live, not after

FAQ

Does the AI agent need access to all our data to work? No. A well-designed agent has scoped access — it can only read and write the specific records and fields relevant to its task. It should be treated like any other service account: minimum necessary permissions, logged activity, regular access reviews.

What happens when the agent makes a mistake? This depends heavily on how it's designed. If the agent is set up with human-in-the-loop checkpoints for anything above a certain value or outside a known pattern, the blast radius of a mistake is small. Full audit logs mean you can trace exactly what happened and why. The same mistakes happen with human operators — the difference is agents make them consistently (and consistently correctable) while humans make them unpredictably.

Do we need to replace our current software to use AI agents? No. AI agents sit on top of existing systems as an orchestration layer. FileMaker keeps running. Your ERP keeps running. The agent connects them and acts across them. That's actually one of the practical advantages: you get intelligent automation without a rip-and-replace.

How long does it take to implement an AI agent integration? A focused single-process agent (e.g. invoice processing against one ERP) can realistically be built and tested in 6–10 weeks. Multi-system orchestration with complex exception handling takes longer — 3 to 6 months is realistic for a robust production deployment. Beware of vendors promising weeks for enterprise-grade integrations.

Can an AI agent work with FileMaker specifically? Yes. FileMaker exposes data through its native Data API (JSON over HTTPS), which is straightforward for an agent to consume. Custom scripts and webhooks in FileMaker can also trigger agent workflows in real time. For a deeper look at how AI fits inside FileMaker-based operations, see A practical guide to AI agents in business operations.

Is this safe from a GDPR and data privacy perspective? It can be — but it requires deliberate design. Data processed by the LLM (the reasoning layer) should either be anonymized/pseudonymized or handled by a model deployed within your own infrastructure or a compliant EU cloud. Customer names, order details, and invoice data should not flow through consumer AI products without a proper data processing agreement.

---

If your business is running into the friction described in any of these scenarios — manual handoffs between systems, invoices that take days to process, support queues that eat your team's time — the underlying problem is usually a combination of disconnected systems and missing orchestration logic. Loggix works with businesses to map those processes, design the right agent architecture, and build the integrations — whether that means connecting FileMaker to an ERP via API, building a custom processing layer, or adding an AI reasoning component to an existing workflow. If you want to think through where an agent would actually move the needle in your operation, that's a conversation worth having.