Your ERP Vendor Just Announced AI Agents. Here Is What to Ask.

The Season of Agents

This autumn, the large enterprise software vendors converged on the same idea.

In September, Oracle announced more than fifty role-based AI agents across its Fusion Cloud applications. On October 8, SAP announced collaborative AI agents for its Joule assistant, designed to work together across business functions. On October 21, Microsoft announced ten autonomous agents for Dynamics 365, including a Supplier Communications Agent meant to track supplier performance and act on delays, along with tools for companies to build their own agents. Salesforce made its Agentforce platform generally available in late October, priced at two dollars per conversation.

The details differ, but the direction is the same.

The details differ, but the direction is the same. The vendors are moving from AI that answers questions to AI that does work: reading documents, updating records, sending messages and starting transactions, sometimes without a person approving each step.

What an Agent Is

The distinction matters. The generative AI assistants that appeared in business software over the past two years mostly help a person do something: summarize a document, draft an email, answer a question about data. The person remains in control and takes the action.

An agent is given a goal and the ability to act. It can decide which steps to take, call other systems, and carry out transactions within the permissions it has been given. An agent watching for late supplier deliveries might detect a delay, check alternative suppliers, draft and send a message, and update the expected receipt date in the ERP.

That is a powerful idea for manufacturers, whose operations are full of routine coordination work of exactly this kind. It is also a new kind of risk, because the software is now making changes to the system of record.

Agents Run on Your Data

The first thing to understand is that an agent is only as good as the data and processes it works with. An agent that expedites purchase orders based on lead times will expedite the wrong ones if the lead times are wrong. An agent that answers customer questions about order status will give wrong answers if the order data is out of date. An agent that matches invoices will make errors if supplier records are duplicated.

The data quality problems that manufacturers have tolerated for years, because people quietly worked around them, become much more costly when software acts on them directly and at speed. Before any agent pilot, look honestly at the data the agent will rely on.

The Questions to Ask

When a vendor presents its agents, ask what each agent can actually change in the system, not just what it can read. Ask how its permissions are set and whether they can be limited to the minimum needed. Ask whether every action is logged in a way you can review, and whether a person can be required to approve actions above a threshold.

Ask how the agent is priced, and what happens to cost as usage grows. Per-conversation and per-action pricing is easy to approve for a pilot and hard to forecast for a full deployment.

Ask what happens when the agent is wrong. How is a bad action detected and reversed? Who is accountable? And ask which agents are generally available today, which are in preview, and which are announced for next year. The gap between an announcement and a product you can run in a plant can be long.

Start Narrow

The most sensible first step is a narrow pilot on a single, well-understood process with clear measures of success. Choose a task that is frequent, rules-based and low risk if a mistake is made, such as monitoring supplier confirmations, drafting follow-ups for late deliveries, or flagging invoice mismatches for a person to review. Measure the baseline first, keep a person approving actions at the start, and expand only when the results justify it.

The Bottom Line

The announcements of the past two months mark a real change: enterprise software vendors are building AI that acts inside business systems, not just AI that answers questions. For manufacturers, agents could take on a great deal of routine coordination work. They will also act on whatever data they are given, with the permissions they are granted. Clean the data first, insist on limited permissions, logging and human approval, understand the pricing, and start with one narrow, measurable process.

At Cherry Street, this is the work we do. We help manufacturers separate what vendors have announced from what they can run today, prepare the ERP data agents will depend on, and design pilots that are narrow enough to measure and safe enough to trust.