Your control tower should be able to drive, not just watch.

This is not competitive speculation. Every number below comes from Palantir's own published pages, its supply chain and food-and-beverage offering materials, and its customer stories, all retrieved 2026-10-08. We cite them because they are the clearest public evidence of where value lands in this category, and because the pattern they form has direct consequences for anyone evaluating a control tower today, including an open-source one.

What the proof points actually are

Read the supply chain evidence back to back and count how much of it is "better visibility" versus "something happened":

• Tyson Foods: $200M in cost savings across 20 use cases within two years.
• Beyond Meat: $20M+ in yearly savings, data unified across 35+ sources, and about half of orders fulfilled automatically.
• Aramark: point-of-sale item matching reached 99% confidence on 30% of initial matches, which was then automated.
• Wendy's QSCC: proactive identification and resolution of product shortages, with supplier emails in unstructured formats folded into the loop.
• Heineken USA: $4.9M in potential out-of-stocks recovered.
• General Mills: roughly $40k per day of avoided loss, annualized by the customer to about $14M.
• Kinder Morgan: multi-source integration collapsed from two-to-three hours to two-to-three minutes, saving hours of analysis daily.
• An industrial customer: $200M saved over 24 months, including $40M of logistics optimization inside 120 days.

Auto-fulfilled orders are writes. Automated matches are writes. Resolved shortages, expedited freight, and re-planned inventory are writes, executed against order books, warehouses, and ERPs. Visibility is how the system knows; the savings are what it did.

The uncomfortable question the case studies skip

None of the published stories explains what happens when the action layer is wrong. If an agent fulfills half your orders, it is writing to your order book thousands of times a day. Before that is safe, someone has to answer, in policy rather than in slides: which actions run straight through, which require a human, who that human is, whether the initiator can approve their own proposal, and how you replay a disputed decision six months later. A tower that cannot answer these is not a control tower; it is a window.

The stakes scale with the automation. A wrong match at 99% confidence is still wrong one time in a hundred, and those writes execute without a human in the loop. A shortage-resolution agent reading unstructured supplier emails is one prompt-injection away from expediting the wrong purchase order. The failure mode you should insist on is the boring one: a misbehaving agent produces a rejected proposal on an audit chain, and nothing else moves. That is an approval-loop property, not a model-quality property.

Three demands to make of any control tower

Whatever you buy, open-source or otherwise, make the vendor demonstrate three things on the execution path:

• Staged proposals on every write. No agent mutates an order book, a warehouse system, or an ERP record directly. The proposal is a record before it is an action.
• Policy evaluated per action, at runtime. Proportionality is a dial: low-risk writes flow, critical ones hard-stop for a qualified approver, and the initiator can never clear their own proposal.
• An evidence chain you own. Approvals, rejections, and confirmations land on an append-only, replayable chain inside your perimeter. If the evidence lives in the vendor's cloud, you have rented your own audit trail.

These are the provable trust properties, applied to freight, purchase orders, and inventory movements rather than to a compliance binder. They are also the difference between a control tower that scales and one that becomes a liability the first time an agent overreaches.

Where Syntheka stands

Supply chains run on ERPs, and most of the large ones run on SAP. Syntheka is built for exactly the layer the case studies monetize: agents that propose actions into your systems of record, approvals with segregation of duties before anything executes, and an evidence chain in your own database. Governed write-back to SAP is a first-class capability, not an afterthought, so the loop the case studies describe, from signal to confirmed execution, closes inside your perimeter. The cross-industry evidence says this is the layer enterprises pay for; self-hosted says you keep the evidence. Run both claims against your own stack before believing either.