Bread & Boxes:
What Grupo Bimbo and DHL Tell Us About Oracle's AI Bet
At Oracle AI World, Grupo Bimbo asked the question: “Is there an undo button? Because it’s going to be my team that’s got to go clean it up.” Stream this episode to hear how Oracle answered, and why DHL won’t let any agent run unsupervised until it proves 95% accuracy.
TRANSCRIPT:
A loaf of bread has about seven days to live.
Some of Grupo Bimbo's products get ten. The longest get twenty-one.
And every hour that loaf rides on the wrong truck, it loses money.
My name is Mark Vigoroso, CEO & Founder of The Enterprise Edge and welcome to a special edition of EdgeBytes, recording live here at Oracle AI World 2026 in Las Vegas!
For the next few minutes, I’d like to share what I learned when I sat down with leaders from Oracle customers Grupo Bimbo and DHL Supply Chain. Different businesses. Same confession, almost word for word.
As you probably gathered, Grupo Bimbo is in the Food and Beverage space. And on the other end of the spectrum is DHL Supply Chain. Two hundred thousand people. 1.7 million supplier invoices a year. An operating margin of 6.5%.
One company makes things that spoil. The other moves things for a living.
To start they both said they are struggling to keep up with the pace of change and velocity of releases at Oracle.
One Oracle executive told DHL that eight months ago, Oracle couldn't have built what it announced this week. When the vendor moves that fast, your ability to absorb change becomes the scarcest resource you have.
When two customers this size tell you the vendor is outrunning them, the constraint has moved. It's no longer the software. It's the organization holding it.
So let's get into it.
Quick context, because the vocabulary changed this fall.
On September 29 in Munich, Oracle introduced Fusion Claw. Think of it as an execution engine for agents that run long, multi-step jobs. A frontier model from Google or OpenAI does the planning. Then Fusion's own deterministic engines do the math and post the transactions. The model never writes to your records directly.
Twenty-five Claw-powered agentic applications went live across ERP, HR, supply chain and customer experience. In supply chain: Supply Planning Allocation, Shipment Load Planning, and Procurement Spend. In finance: Ledger, which works your close exceptions.
At AI World, Oracle added Fusion Companion, which lets you drive Fusion from tools like ChatGPT, and Agentic Home, where you set an objective, delegate it, and track it.
The governance wrapper has three parts. An Operating Envelope holds your policies and approval limits. A Trust Harness fences what each run can touch. And an Outcome Receipt records who authorized what, on what evidence, with what result.
Two honest caveats. Oracle hasn't published prices; the apps draw down "AI Units" as they work. And Gartner has said, Claw is "very much in early piloting mode still."
So, what do the people actually running this stuff think? Five themes came through in both conversations.
Theme 1: The boring decision was the AI decision
Neither company got AI-ready by buying AI. They got there by doing the unglamorous thing years earlier: one platform.
Grupo Bimbo’s story starts with SAP. They implemented it at Canada Bread. Then Grupo Bimbo bought Canada Bread around 2015, and the leader I spoke with helped retire it. Today Bimbo runs five regional instances: North America, Latin America, Brazil, Europe and China. Canada is the lighthouse, about 90% Oracle. The U.S. is next, where Bimbo is retiring AS/400 systems that are 45 years old.
"The old days, everybody just went out and bought the best of breed. It's not a way to be anymore." – said my new colleague from Grupo Bimbo.
The reasoning: an agent can go fetch data from other systems, but every hop "adds timing and capacity." When it all lives in one place, the agent is already there.
DHL Supply Chain made the same bet starting in 2019. Five ERPs consolidated into one Oracle platform. Oracle HR across the entire DHL Group. BlackLine pulled out and replaced with Oracle's own reconciliation tool. They called it a "digital core," built on purpose, before anyone said "agentic."
The payoff showed up this year. When I asked whether agents forced them to rebuild their data models, the answer was no. Nothing. The foundation held.
My read: in 2026, the integration count on your architecture diagram is an AI cost line. Every interface is a toll booth your agents have to pay at, in time and in tokens. Gartner said it this spring: cloud ERP is now valued as "an application for embedded AI," not just a finance foundation.
Theme 2: Latency is a line item
Here's the number that should keep a food CFO up at night. Circana says U.S. food and beverage volume was flat in the first half of 2026. Dollar growth came almost entirely from price.
You can't price your way forward forever. So where does profit come from?
"We have a billion-dollar challenge in supply chain. Supply chain is the place that the money is. Growth is not necessarily like it used to be." Said Grupo Bimbo.
Bimbo's own third quarter backs this up. Net sales fell 1.3%. Net income rose 22.1%.
Now think about what latency means in bread. Clean-label recipes strip out preservatives, so shelf life gets shorter, not longer. Some distribution centers turn their inventory three times a day. A truck can run two or three stops before anyone notices the last drop won't have enough shelf life left to sell.
Today, finding that pattern is a hunt.
"That hunt can take 30 days. By the time you're done, it's so far past that you can't educate the folks on the floor." – said Grupo Bimbo
The company’s prediction is blunt: "Reporting is going to die." Instead of a report telling you four trucks left late last week, you ask, and within minutes the system tells you which driver is about to leave late today.
The math is simple. If a product line runs nine or ten percent waste and you bring it to four, you've funded a year of inflation. And it's not just the bread. It's the freight, the handling, the returns, the disposal. ReFED puts U.S. surplus food at $382 billion a year. That's 31% of the food supply.
DHL lives the finance version of the same problem. Seventy thousand journals a year. 1.7 million invoices. Oracle's Ledger agent lands in DHL's early-adopter environment on November 5 and in production just before Christmas. Its job is to find and resolve close exceptions while the books are still open, not three weeks after.
Same principle, two industries: the money is in the gap between when something happens and when you find out.
Theme 3: Show me the receipts
In a session, Oracle demoed an agent posting general-ledger adjustments for waste. A prompt pops up: are you sure? The presenter clicks yes.
My Grupo Bimbo friend raised his hand. "So, is there an undo button? Because it's going to be my team that's got to go clean it up."
That's the most important question anyone asked all week. An agent that acts at machine speed also makes mistakes at machine speed. The point: you can show the confirmation prompt ten times, and an impatient human will still click through.
What won Bimbo over was the audit trail. Oracle added new fields to the tables, so a record no longer says "batch 123 ran." It says Claw ran it, which user activated it, when, under which procedure. That's the Outcome Receipt in practice.
The auditors are going to love that.
DHL is stricter still. Every agent stops at a human until it's proven. Their rule:"We're not marking our own homework. A grown-up somewhere is going to go and check that."
InfoSec sits in on build calls with standing permission to shout. Agent permissions map directly to Oracle roles. And the bar to loosen the leash is explicit: prove 95% invoice recognition, repeatedly, with data. Only then do you talk about letting anything above an 80% confidence score go straight through.
Then DHL raised a point nobody on stage did. If an agent runs your controls, your auditor relationship changes. The auditor asks for evidence. You hand over an agent's log of the policies it checked and the steps it took. And the auditor asks, how am I supposed to read this?
My takeaway: governance is now a product feature you should test in the demo, like any other. Ask for the undo. Ask for the receipt. Ask your audit firm to read one before you go live.
Theme 4: Stop pitching headcount
This was among the sharpest strategic tips of the week, and it came from DHL.
"If the first thing that comes out of your mouth is efficiency, you've painted yourself into a corner with your senior executives and board. All they're going to think is headcount reduction."
DHL leads with quality, risk, compliance and insight. Efficiency follows, because at 1.7 million invoices it has to. They also asked the question most business cases skip: what will the tokens actually cost? And what happens to your digital workforce during an outage or a cyberattack?
Bimbo comes at ROI from the other side. Bimbo might have 30 waste-reduction projects running at once. When waste drops, which one gets the credit? Today, nobody can say. He expects agents to fix that attribution problem, because the data trail will finally be good enough.
MIT's NANDA study found 95% of generative AI pilots never reach scaled impact, and the best returns sat in back-office work. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 over cost, unclear value or weak risk controls. Translation: the business case that survives is one that names a control, a risk or a cycle time, then counts the savings second.
Theme 5: Small crew, short leash
How many people does DHL have building agents? Two or three at a time.
Their first requisition agent was built in one afternoon by a functional consultant who had never opened Agent Studio. Not a developer. Another agent now pushes 150 payment-term changes from a spreadsheet in one pass. Work that once took weeks of specs, SQL and testing now takes hours.
So why not hand the tool to everyone? DHL said no, on purpose. They remember the era when everyone could build reports, and every meeting started with an argument over whose KPI was right.
They said, "We do not want to be developers. We want to deploy change management."
On jobs, Bimbo expects the effect in supply chain to be roughly "net neutral." A DC that needed 100 trucks a day needs 90. You may stop running 24/7. But you still need people on the floor to sort things out and fix them.
The prognosis: Oracle in food and logistics
These are my forecasts, not Oracle's.
One. Finance agents hit scaled production before supply chain agents do. DHL's first live Claw workload is Ledger, not logistics. Bimbo is still finishing its ERP foundation. Expect 2027 to be the year of the close, the payables desk and collections. Supply chain follows wherever master data is clean.
Two. Oracle has roughly a two-quarter lead on shipping governed, outcome-based agents. Like bread, that lead has a shelf life. Competitors’ supply chain agents arrive this quarter. Oracle's edge holds only if customers like DHL turn early pilots into published, measured results in 2027. Oracle has yet to publish measured production results for Claw.
Three. The suite pressures the specialists in the middle of the process first. Reconciliation and close tools are most exposed; DHL already swapped out BlackLine. The warehouse floor is a different story. Bimbo still runs a separate handheld system for its field operators. DHL uses Oracle warehouse management in "bits and pieces." Blue Yonder and Manhattan keep that ground through 2027. Transportation is where Oracle is strongest, and in logistics and CPG that's the lane that matters most.
What's your edge?
For CEOs.
First, treat your ability to absorb change as the constraint, and staff it. Both customers said the vendor is outrunning them. Name one owner for each cross-functional outcome before you switch on an agent. Oracle has drafted about 100 "value paths" that split work between people and agents. Use them as a starting map, not an answer.
Second, aim AI at your slowest feedback loop in your biggest margin pool. Circana has U.S. food volume flat in 2026. Bimbo grew net income 22% on lower sales. In a business with seven-day shelf life, the 30-day hunt is the target. Pick one number, waste percentage or days to close, and make it the scoreboard.
For CFOs.
First, write the business case in DHL's order: quality, risk, compliance, insight, then efficiency. Gartner expects over 40% of agentic projects to be canceled by the end of 2027 for cost or unclear value. Gartner also predicts CFOs who manage AI as a portfolio, not as isolated pilots, will add 10 margin points of growth by 2029.
Second, put a ceiling on run cost before you scale. Claw apps consume AI Units as they work. Get a measured consumption baseline from a pilot and a price cap in the contract. Oracle's design lets the model plan while Fusion's own engines do the computation. Gartner notes that avoids feeding millions of records through a model's context window. Make your account team show you what that saves in units.
Third, call your auditor now. Have them read a sample Outcome Receipt before Ledger goes live. And tag every waste or working-capital project, so when the savings show up you can finally say which project earned them.
For CIOs.
First, count your interfaces and start retiring them. Bimbo is retiring 45-year-old AS/400s. DHL removed BlackLine. MIT's NANDA research found purchased AI tools succeed about two-thirds of the time; internal builds about one-third. Start with the vendor's agents, and use Agent Studio for the gaps.
Second, keep agent building small and governed. Run a crew of two or three, map every agent to an existing ERP role, and put InfoSec in the build calls. Keep a human in the loop until the data proves 95% accuracy, then set confidence thresholds.
Third, test the undo and the outage. Ask for the rollback procedure on every agent that posts transactions. Write the runbook for when your digital workforce goes down. And take Gartner's advice: put Fusion Claw in your agent RFPs, alongside a clear exit plan.
That’s all for now. Thank you for joining me for a few minutes here at the Venetian. Would love to hear your reactions, experiences, and other thoughts. Leave a like, share this video or drop a comment below. See you on the next episode of Edge Bytes, where we deliver Signal over noise. Take care everybody.