Enterprise AI Stuck in Chat Mode? MCP Cuts Invoice Costs to $2.65

Why are enterprise AI investments failing to deliver transformation?

Many enterprise AI investments fail to deliver promised transformation because LLMs are disconnected from core business systems. Without access to ERP, vendor databases, or contracts, an LLM can only summarize or answer questions—it cannot take action, making it essentially a chatbot with polished writing.

Executives have heard that AI will transform work, but initial investments often fall short. The root cause is that AI models lack the context needed to make reliable decisions. For example, a finance team processing invoices needs to match them to purchase orders, check payment status, and verify GL coding—all data that lives outside the AI tool.

An isolated LLM might summarize an invoice accurately, but it cannot determine if the invoice matches the PO or if it has been paid. This limitation means AI remains a passive assistant rather than an active participant in business processes.

To move beyond chatbots, AI needs secure connections to the systems where business context resides. This is where the Model Context Protocol (MCP) comes in, providing a standardized way for AI to access external data and act on it.

What is the Model Context Protocol and how does it enable agentic AI?

The Model Context Protocol (MCP) is a standardized way for AI applications to connect with external systems and data sources. It allows AI agents to pull relevant data, make decisions, and take action within defined boundaries, including sending information back to core business systems.

MCP simplifies integration by providing a common protocol, so developers don't need to create custom connections for every AI application and system. This is crucial as businesses add more AI tools and data sources.

With MCP, AI agents can access ERP, vendor databases, and other systems to gather context. They can then use that information to make decisions—such as flagging a price discrepancy—and take actions like requesting a credit or routing an exception for approval.

MCP supports bidirectional connections, meaning AI can also update records or trigger next steps. This turns AI from a passive information provider into an active workflow participant.

1. AI agent pulls invoice data from ERP.
2. Compares invoice to purchase order and receiving records.
3. Identifies discrepancies (e.g., quantity or price mismatch).
4. Takes action: request credit, route for approval, or clear for payment.

How does connected AI transform accounts payable processing?

Connected AI transforms accounts payable by enabling touchless processing, where invoices are captured, validated, coded, approved, and paid without human intervention. This reduces the average cost per invoice from $9.84 to $2.65 for best-in-class organizations.

In AP, an AI agent with access to purchase orders, receiving records, contracts, and payment history can perform the same checks as a human employee. For example, if a restaurant receives an invoice for 20 cases of chicken but only 18 were delivered, the agent can detect the discrepancy and the $5 per case price increase.

Instead of routing the invoice to an exception queue, the agent can investigate further and take action based on company policies. It might request a credit from the vendor, route the discrepancy to an employee with recommended next steps, or clear the invoice for payment if everything matches.

This same concept applies to other AP decisions: identifying duplicate invoices, determining GL codes, or routing for approval. The key is that the agent's ability to decide and act depends on access to the right information.

What is touchless processing and why is it a key metric for AI success?

Touchless processing is when an invoice is captured, validated, coded, approved, and paid without any human intervention. It is a clear metric for AI success in AP because there is no in-between: either the AI completed all steps accurately or it didn't.

Touchless processing provides a straightforward way to measure whether AI agents are truly working. If a person had to step in at any point, the invoice was not processed touchlessly. This binary outcome makes AP an ideal proving ground for connected AI.

Mistakes in AP can have real consequences, such as manual rework, overpayments, missed discounts, and inaccurate accounting records. Therefore, successful touchless processing demonstrates that AI can be trusted with critical business processes.

In other areas, AI's success can be harder to measure. For instance, a chatbot might save an employee time, but quantifying that savings is difficult. AP offers concrete numbers and clear ROI.

Do AI agents need full autonomy to be effective in business processes?

No, AI agents do not need full autonomy to be effective. They can take on a more active role in business processes with proper guardrails and human oversight, as long as they have access to the necessary data and can take actions within defined boundaries.

As Meng Liu, Principal Analyst at Forrester, recently said, "agentic branding is widespread, but real autonomy is not." The level of autonomy AI can handle varies by task and implementation.

In accounts payable, AI agents can handle many steps without full autonomy. For example, they can investigate discrepancies and recommend actions, but final approval might still require a human. This hybrid approach balances efficiency with control.

The key is to define clear boundaries and policies for AI agents. They should be able to access data and take actions within those boundaries, but not exceed them without human intervention when necessary.

Bottom line

Enterprise AI must move beyond chatbots to deliver real transformation. By adopting the Model Context Protocol to connect AI agents to core business systems, companies can enable agentic workflows that cut costs and boost efficiency. Accounts payable offers a proven starting point with measurable results. Businesses should prioritize MCP integration and define clear guardrails to unlock AI's full potential.