Put Your First ERP Agent on Probation

Two Years In, the Pitch Has Changed

In October 2024, within two weeks of each other, SAP and Microsoft announced that their ERP systems would no longer just answer questions. SAP introduced collaborative agents in Joule. Microsoft announced ten autonomous agents for Dynamics 365. The copilot that summarized a sales order was giving way to software that could act on one.

Two years later the catalogs are enormous. SAP has said it expects to offer more than 400 AI agents by the end of this year, and ERP Today's coverage of SAP Connect this month put its supply chain agents at general availability in the fourth quarter, with the customer rollout only beginning. Infor says it has more than 100 agents. Every major ERP vendor now sells some version of a digital workforce that plans, expedites, reconciles and approves while your people sleep.

Here is the problem with that pitch: you would never hire a hundred people on the same day, give them all system access, and skip the probation period.

Here is the problem with that pitch: you would never hire a hundred people on the same day, give them all system access, and skip the probation period. That is roughly what an agent catalog invites you to do, and the manufacturers getting value from agents this year are doing the opposite. They are hiring one, giving it one job, watching it closely, and keeping the ability to let it go.

What Plants Are Actually Seeing

Interest is not the issue. RSM published a survey of 129 manufacturers in July and found that 56 percent are already using agentic AI in some form. But the same research found that pilots routinely fall short, and the reasons have little to do with the models. Data quality was the most common obstacle, cited by 51 percent, followed by integration with existing systems at 43 percent.

The broader picture is harsher. A survey by the communications platform Sinch, reported by CXM World, found that 74 percent of organizations that had deployed an AI agent had rolled one back or shut it down, and the most common reason, at 31 percent, was exposure of personal data. That is not a manufacturing number, but it describes the same pattern: agents are easy to turn on and hard to trust once they are touching real records.

None of this should surprise anyone who has lived through an ERP implementation. An agent inherits every flaw in the data it reads and every gap in the process it automates. A purchase order with the wrong lead time produces a wrong expedite just as reliably as it produces a wrong MRP suggestion. The difference is that the agent acts on it, at volume, without the planner who would have noticed.

The Bill Now Arrives Differently

There is a second reason to go slowly, and it shows up on the invoice.

The ERP vendors are moving AI to consumption pricing. SAPinsider reported in August that use-based AI pricing became the default in SAP renewals in July, and that an agentic approval workflow can use up more units in a week than a query-based deployment does in a month, with overages billed outside the normal cycle. Microsoft prices Copilot Credits at a cent apiece, which sounds trivial until an agent is running thousands of steps a day unattended. The licensing advisory firm Redress Compliance estimates that a single agent run consumes five to ten times what an interactive prompt does. Gartner analyst Victoria Rowan told The Register in May that pricing per billable action can climb quickly, and SAP chief executive Christian Klein has described the shift, according to ERP Today, as a chance to "completely reset the price level."

Read that last phrase from the buyer's side of the table. A seat license is a cost you can forecast. A per-action meter attached to software designed to take more actions on its own is not, unless you put limits on it before it starts. Most manufacturers will meet this question at their next renewal rather than in a pilot, which means the time to understand it is before the renewal conversation, not during it.

Treat the First Agent Like a New Hire

The discipline that works is one every plant manager already knows. When you bring on a new expediter or a new buyer, you do five things. You give them a defined job. You put them under a supervisor with approval limits. You can see what they did and why. You control what they can spend. And for the first ninety days, you can let them go without a crisis. An ERP agent deserves exactly the same treatment.

The job description is one workflow, not a department. Chasing supplier confirmations on open purchase orders. Triaging MRP exception messages. Working three-way match exceptions in accounts payable. Each of these is high volume, mostly rule-driven, recoverable when it goes wrong, and, critically, already measured. You know how many confirmations go unanswered, how many exception messages age past a week, how many invoices sit on hold. That existing number is the baseline the agent has to beat, and if you cannot write it down before the pilot, you will not be able to prove anything after it.

The supervisor is a named person with approval thresholds. The agent drafts the expedite; the buyer approves anything over a dollar amount or outside a date window. As the agent's error rate proves itself, the thresholds widen. That is how you extend trust to a person, and it works the same way here.

Visibility means a log of every action and the reasoning behind it, in a form someone on your team can actually read. Infor has made visibility into agent reasoning before production a selling point, and every vendor should be held to that standard. If you cannot reconstruct why an agent moved a due date, you cannot defend it to a customer, and you cannot fix it.

The spending limit is a consumption cap, set in the contract and enforced in the system, so that a runaway workflow stops rather than bills. And the termination clause is a rollback plan: the manual process documented and still staffed, a known list of the records the agent can touch, and the ability to turn it off on a Tuesday afternoon and be back to normal by Wednesday.

Pick the Job Before You Pick the Agent

The order matters. The vendor demo starts with the agent and looks for a place to put it. Start with the workflow instead, because the right first job is one where your data is already reasonably clean and your process is already reasonably stable.

That usually rules out the most exciting candidates. Pricing and quoting touch customers directly. Quality disposition carries regulatory weight. Production scheduling depends on information that, in most small and mid-market plants, lives partly in the ERP and partly in a supervisor's head. Those can come later. The first hire should be the one where a mistake costs you an afternoon, not a customer, and where success shows up in a number you already track.

It also helps to accept that the first agent's most valuable output may be a list of data problems. If it cannot expedite because the supplier lead times in your item master are three years stale, the pilot has done its job by finding that. Fixing the data is the step that makes the second agent work.

Write the Probation Into the Contract

Because consumption pricing is arriving through renewals, the probation terms belong in the commercial negotiation as much as in the project plan. Ask for a hard monthly cap on AI consumption, with notice well before you reach it. Ask for unit price benchmarks and protection against repricing during the term. Ask for the right to export the agent's action logs in a usable format. And ask what it costs to turn the agent off, because an agent you cannot afford to stop is not on probation.

UpperEdge, which negotiates software agreements on the buyer's side, has been urging customers to settle the pricing structure, the caps and the benchmarks now, while the vendors still want reference customers more than they want overage revenue. That leverage will not last.

The Bottom Line

The agent era in ERP is real, and the vendors are shipping fast. That is a reason to start, not a reason to start big. The manufacturers who come out ahead will not be the ones who switched on the most agents in 2026. They will be the ones who hired one, measured it against a real baseline, kept it on a short leash and a capped meter, and then hired the second one on evidence instead of a demo.

At Cherry Street, this is the work we do. We are independent of every ERP vendor, so we can help you choose the first workflow, measure the baseline before anything changes, price the same job across SAP, Microsoft, Oracle, Infor and Epicor, write the probation terms into your renewal, and run the first ninety days alongside your team. If your vendor is pitching you a digital workforce, start by giving one of them a job interview.