Which logistics and sales decisions I'd hand to an AI agent today (and which I wouldn't)
Humberto Carrillo draws the line from thirty years in SAP sales and distribution, logistics execution and process industries: the four questions he asks before handing a decision to an AI agent, which order-to-cash and logistics decisions pass the test today, and which should stay with people.
I’ve spent close to 30 years around SAP: seven at SAP itself in the Latin American oil and gas consulting hub, and since 2011 on the partner side across consumer products, retail and energy. Most of that time has been in sales and distribution, logistics execution and process industries, on projects like PDVSA Gas, Bariven and mining operations in Peru. So when people ask me whether AI agents are ready to “run” order-to-cash or the warehouse, my answer is: some decisions, yes, today. Others, not yet. And the line between them is less about the technology than about the decision itself.
Here is how I draw that line.
The test I use
Before handing a decision to an agent, I ask four questions:
- Is the decision frequent and repetitive? Agents earn their keep on volume.
- Is the data it needs already in the system, clean and trusted?
- Is a wrong call cheap and reversible?
- Can a person explain afterwards why the decision was made?
If the answer to all four is yes, I’d hand it over now. If any is no, the agent should prepare the decision, not make it.
What I’d hand to an agent today
Order exceptions in order-to-cash. Credit blocks within tolerance, incomplete orders, delivery blocks with a known cause. These are high volume, rule-heavy and reversible. An agent can clear the routine ones and route the rest with context attached. A typical example: a sales order held for a credit check that is slightly over the customer’s limit but within an agreed tolerance. The agent verifies payment history, releases the block and logs why, instead of the order waiting in a queue for a person to notice it.
Delivery and shipment scheduling within set rules. Grouping deliveries, proposing shipment dates, re-sequencing picks when a truck is late. The constraints are known and the cost of a suboptimal choice is small.
Purchasing approvals below a threshold. Low-value, catalog or contract-backed requests where the agent checks budget, contract and supplier status and approves or escalates.
Customer-service handoffs. Order status, proof of delivery, invoice copies. The agent answers from SAP data and hands the case to a person, with the history, when it goes beyond that.
What I wouldn’t hand over yet
Pricing and discount decisions with key accounts. The data says one thing; the relationship often says another. An agent can model the options, but a person should own the call.
Allocation when supply is short. Deciding which customer gets product when there isn’t enough for everyone is a commercial and sometimes political decision. A typical example: a process plant that can cover only part of its customers’ demand in a given week has to choose between a long-standing key account, a contractual commitment and a new customer. The agent can show the options and their impact, but the choice should stay with a person.
Anything safety- or compliance-critical in process industries. In gas and mining, a logistics decision can carry safety and regulatory weight. Agents should flag and recommend; people decide.
Master data changes. Customer, material and pricing master data feed every downstream decision. A confident agent with write access to master data can spread a mistake faster than any person could.
Why the line sits there
The decisions I’d hand over share three traits: they are bounded by rules the business already agreed on, the data lives in the core system, and mistakes are visible and fixable. The ones I’d hold back depend on context that isn’t in the system yet: relationships, judgment under scarcity, safety. As that context gets captured, and as clean-core and data-platform work brings more of it into one place, the line will move. But it should move one decision at a time, with people watching the results.
Where to start
Pick one exception queue that your team handles every day, measure how it’s handled today, let an agent work it in “recommend” mode for a few weeks, then let it act on the cases where its recommendations matched. That’s a small, reversible step, and it tells you more than any pilot slide.
Humberto writes in a personal capacity; he leads Solution Management for GBM across Central America and the Caribbean.
Next: the platform synthesis this question sits at the heart of — AI agents do not need more data. They need business context. — or meet all sapperment experts.