Most enterprise AI initiatives share a common sticking point. It's reached when someone asks the obvious question about impressive AI pilots: how does it connect to our ERP, our custom order management system or the finance platform we've run for 18 years?
The question that has been quietly killing AI projects across industries was addressed head-on in the latest Frends Spotlight.
During the webinar, Frends CTO Asmo Urpilainen walked attendees through the newly released Frends Enterprise MCP and showed, in a live demo, what it looks like when AI can finally talk to the systems that matter.
36% of organizations cite business orchestration and integration challenges as the primary blocker keeping AI out of production. That number comes from the State of Integration & AI report, and it tracks with what Asmo hears consistently in conversations with enterprise architects, CTOs and CIOs.
"The problematic area is not going to be the new off-the-shelf systems and AI-native solutions that are appearing in the market," Asmo explained. "But the ERP system from 10 or 20 years ago or the custom in-house built CRM system, and how you get those very critical parts of the enterprise to talk with any AI solution."
The irony is that the systems no one wants to touch are often the ones powering the most critical processes. And for as long as those systems sit outside the reach of AI, the productivity gains and automation benefits stay theoretical.
MCP — Model Context Protocol — is an open standard that gives AI a governed, standardized way to interact with tools, systems and data. "A good way to think about the MCP protocol in general — it's kind of this USB connector for AI, which basically standardizes the way AI can interact with any existing system, data or business processes," says Asmo.
Frends Enterprise MCP builds on that standard in a way that's specifically designed for enterprise conditions: legacy systems, compliance requirements, audit obligations and the kind of process complexity that off-the-shelf AI tooling was never built to handle.
The mechanism is straightforward, explained Asmo. A Frends Process sits in front of any existing system — no changes to the underlying code required — and exposes that process as a governed, callable tool. This means you can connect Claude, ChatGPT or your own AI of choice to that toolkit via the MCP protocol, and the system is immediately reachable.
To make this concrete, Asmo built a live demo around a finance team's most familiar headache: invoice and purchase order approval.
The scenario involved a finance system almost 20 years old, one that still powers core processes and had no AI compatibility whatsoever. Using Frends Enterprise MCP, four tools were built on top of it: one to retrieve pending invoices, one to fetch vendor details, one to pull full invoice specifics, and one to approve or hold an invoice in the system.
With those tools exposed via MCP, Claude was connected to the toolkit. What followed demonstrated what governed AI access to a legacy system looks like.
Asked whether there were open purchase orders to process, the AI retrieved the pending invoices, ran analysis on the data and flagged two anomalies: one invoice suspected as a duplicate from the same vendor for the same amount, and one where the invoice total didn't match the authorized purchase order. The discrepancy in the second case came to €2,400, traced to 20 extra boxes invoiced beyond what was ordered.
At that point, the AI prompted for what came next: hold the flagged invoices, approve the clean ones.
"This is really where the intelligence in AI comes into play," Asmo noted, "because the AI now has access to that data in real time in my system. It's able to say that this particular invoice has been flagged as a suspected duplicate, and the reason is that it's from the same vendor for the same amount."
But the demo didn't end with the conversation. Every action taken by the AI was mapped directly to a process instance in Frends, with full execution logs, raw data visible at every step and a complete trail from the first tool call to the final approval. Guardrails, data masking, per-identity access controls and time-limited authorizations are all configurable at that layer.
The second act of the demo removed the human from the loop entirely.
Using the Frends AI Connector, the same four tools were provided to an agent running inside a Frends process. The same logic that played out through the Claude conversation was now executed automatically with the AI reasoning through the invoice set, identifying the same anomalies and producing a structured output in 13 tool calls, end to end.
For situations where human input is still needed, the Frends long-running process and checkpoint capabilities handle the handoff. The agent pauses, the state is saved, and the process resumes when a human provides input, whether through a Teams message or a record update in a system.
"We should really consider wrapping all of our systems, especially our legacy and custom systems, into AI-compatible tools," Asmo said.
He continued: "We can then securely expose those tools for our employees, perhaps even business partners, where they can then use AI in this human-in-the-loop guided manner using their preferred AI client. We can even take it one step further and automate the entire process."
Asmo's advice for teams already using AI but still copying data manually between systems was direct: "If you identify any processes with manual data entry, or you are already using AI but you see yourself copy-pasting from your system to your AI, these are the lowest-hanging fruits. With just a few hours of work, they can bring a really, really big impact."
Frends Enterprise MCP is available in Frends 6.3. Documentation can be found at docs.frends.com.
The Enterprise MCP release is the first iteration. A full Frends MCP server is in development, which will allow AI to interact with the Frends platform itself: developing new integrations, debugging existing ones, monitoring failures, deploying fixes and running architectural analysis.
As a hint of what's ahead, Asmo confirmed that the demos shown in the webinar were built by AI using the Frends MCP Server, not by Asmo himself, and completed in minutes earlier the same morning.
The next Frends Spotlight, scheduled for September 15, will go deeper on those capabilities. Registration is open; secure your spot here.