Every ERP conversation in 2026 eventually arrives at the same question: “Where does AI fit into this?” It’s the right question — asked, usually, at the wrong altitude. The useful answer isn’t a technology list. It’s an economic one: AI belongs where your people repeatedly transform unstructured input into structured records, or repeatedly answer the same questions from data the system already has.
That definition sounds narrow. In a real business, it’s enormous.
The patterns that pay for themselves
1. Conversational self-service on channels customers already use. The clearest ROI I’ve measured didn’t come from a chatbot on a website. It came from putting service workflows on WhatsApp — where customers in most of the world already live. In one deployment, 4,000+ customers now handle subscriptions, service requests, visit reports, and payments in a guided conversation; it replaced roughly half the call-center workload. The “AI” sophistication varies — from structured flows to LLM-assisted responses — but the economics come from the channel and the integration, not the model.
2. Document-to-record automation. Vendor bills, purchase orders, delivery notes, bank statements: businesses pay skilled people to retype documents into the ERP. LLM-based extraction has made this automation genuinely reliable — if it’s wired into the ERP’s validation logic so a human confirms exceptions rather than processing everything. The pattern that works: extract → validate against master data → post automatically above a confidence threshold → queue the rest for one-click review.
3. Workflow agents for the in-between work. The expensive time in operations isn’t the big decisions; it’s the connective tissue — chasing approvals, drafting follow-ups, summarizing a customer’s history before a call, triaging inbound requests to the right queue. Agents that read ERP context and prepare (not send — prepare) these actions compound quietly. This is also where new standards like MCP matter: they give AI tools a disciplined way to access business systems instead of the copy-paste chaos of early adoption.
4. AI-accelerated delivery of the ERP itself. The least glamorous pattern with the most proven numbers: using AI tooling in the development process. Introducing AI-driven development workflows across my team measurably raised productivity around 30%. Clients don’t buy this as a feature — they receive it as speed.
Where AI is just expensive
- Forecasting theater. Demand prediction on top of unreliable inventory data automates noise. Fix data discipline first; the boring truth is that a clean reorder rule beats a sophisticated model fed garbage.
- Chatbots bolted onto broken processes. If your service process frustrates customers, an AI face on it frustrates them faster.
- “AI strategy” without a workflow. If a proposal can’t name the specific workflow, the input, the output, and who reviews exceptions — it’s a demo, not a system.
The sequencing that works
AI multiplies whatever your ERP already is. Clean processes and trusted data get multiplied into leverage; chaos gets multiplied into faster chaos. The sequence I hold clients to:
- Stabilize the core — records people trust, workflows people follow.
- Automate the deterministic — the rule-based work that needs no intelligence, only integration.
- Add AI where judgment is repetitive — documents, conversations, triage, preparation.
- Keep humans on exceptions — the system earns autonomy gradually, with an audit trail.
Businesses that follow that order get compounding returns. Businesses that start at step 3 get invoices.
Wondering which of your workflows are actually AI-ready? That’s a 30-minute conversation with your process list on the table — book a call or see AI & Automation services.