The Agentic Supply Chain: Multi-Agent Systems Negotiating the Physical Economy

Nepal's economy runs on supply chains I can see from my window — trucks queued at the Birgunj border, freight-forwarders haggling over customs codes by phone, the same negotiation happening the same way it has for decades. What Salesforce, SAP, Oracle, and DHL are now automating, at least in pieces, is precisely that phone call: the negotiation, the booking, the compliance check, increasingly handed to a software agent instead of a person. I went through the vendor claims one by one before writing this, because "AI agents are transforming the supply chain" is exactly the kind of sentence that's easy to write and hard to verify.

The platforms are shipping, and one number is real money

Salesforce's Agentforce line is the clearest place to start because it has a real, audited revenue number attached to it. Agentforce went generally available October 25, 2024, priced at $2 per conversation, and Salesforce has shipped a new version roughly every quarter since — Agentforce 2 in December 2024, 2dx in March 2025, Agentforce 3 in June 2025, and Agentforce 360 at general availability on October 13, 2025, by which point the company reported 12,000 customers. The number that matters most, though, came in Salesforce's Q4 FY26 earnings release on February 25, 2026: Agentforce ARR reached $800 million, up 169% year-over-year, with 29,000 deals closed, up 50% quarter over quarter. That's Marc Benioff's own reported figure, not a third-party estimate.

It's worth pairing that growth rate with a counter-signal, because a 169% growth rate off an unstated base can sound more dramatic than it is. Stifel's analysis puts actual customer adoption of Agentforce at roughly 5.3% of Salesforce's customer base — meaning the ARR is growing fast among the customers who've adopted it, while the large majority of Salesforce's customer base hasn't yet.

Microsoft, SAP, and Oracle joined the wave

Salesforce isn't alone, and the fact that three other major enterprise vendors independently shipped competing agentic-AI features through 2025 and 2026 is itself a data point — this isn't one company's marketing narrative. Microsoft previewed autonomous agents in Copilot Studio in October 2024 and took them to public preview at Ignite on November 19, 2024; by Ignite 2025 the platform had added GPT-5 model choice, integration with more than 1,400 systems via MCP, and Entra Agent ID governance for tracking what an agent is actually allowed to do.

SAP moved on a similar timeline: at SAP Connect in October 2025 the company detailed 15 Joule agents, with its Production Planning and Operations Agent reaching general availability in Q1 2026 and Supply Chain Orchestration agents scheduled for the first half of 2026. Oracle shipped a steady drumbeat of Fusion SCM agent releases through the first half of 2026 — an Autonomous Sourcing Agent on February 10, 2026, broader Fusion Agentic Applications on March 24, 2026, and four additional supply-chain agents on June 29, 2026. I'm intentionally not repeating a specific SAP "agent count" figure beyond the 15 named at SAP Connect — a higher number attributed to a later event circulates in consultancy blogs but doesn't trace to an SAP press release.

Where this touches freight and customs, for real

The part of this story closest to Nepal's own border-crossing trucks is the logistics layer, and here the clearest primary-sourced example is DHL. On November 11, 2025, DHL announced autonomous voice and email agents built with HappyRobot, handling what the company describes as hundreds of thousands of emails and millions of voice minutes annually. DHL discloses that volume figure, not a dollar ROI — and I'd rather report the number DHL actually published than manufacture a savings percentage to make the story feel more complete.

Freight-specific platforms are moving too. project44 launched an AI Freight Procurement Agent on February 26, 2026 and expanded its agent portfolio at its decision44 event on April 8, 2026, reporting Q1 FY27 new-ARR growth of 34% across a network the company says spans more than 259,000 carriers and 1.5 billion shipments a year. A specific "4.1% spend reduction" figure attributed to project44 circulates in secondary coverage, but project44's own materials describe only "measurable savings" without a percentage — so that's a number I'm leaving out rather than repeating with an asterisk. Flexport's Winter 2026 release added AI agents for customs and compliance work under its Flexport Atlas platform, again without disclosed ROI figures. Unilever and Google Cloud announced a broader agentic-workflow partnership on February 17, 2026, though the public materials describe commerce and marketing use cases rather than a procurement-specific deployment — worth noting as a partnership to watch, not yet a procurement case study.

Two specific procurement "case studies" that show up in aggregator coverage — autonomous procurement deployments credited to Maersk and Walmart — don't hold up under a primary-source check. Maersk's "Captain Peter" system, as far as I can verify, performs monitoring rather than autonomous negotiation, and I couldn't locate a Walmart source describing an agentic procurement program matching what's being claimed. Both are left out here.

Gartner's warning: this is still mostly a promise

The sobering counterweight to all of the above comes from the analyst firm tracking this space most closely. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024 — both explicitly framed as forecasts, not observed facts. A nearer-term prediction, published August 26, 2025, puts 40% of enterprise applications integrated with task-specific AI agents by the end of 2026, up from less than 5%.

The prediction that matters most for judging how much of this survives, though, is Gartner's warning, published June 25, 2025: over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, and inadequate risk controls. Read against the platform momentum above, the honest shape of this story is that the vendors are real and the money is real, but the fully autonomous, negotiating supply chain most of this marketing implies is still a 2026-2028 bet — and Gartner's own analysts think a plurality of the specific projects placing that bet won't make it.

I'm also leaving out a set of framework-adoption statistics that circulate confidently around open-source agent orchestration tools — specific Fortune 500 adoption percentages for CrewAI, adoption percentages attributed to Gartner for LangGraph, and named-company usage claims (Klarna, Uber, LinkedIn) tied to those percentages. The frameworks themselves are real and widely used, in the same broad category I covered in an earlier look at how AI agents actually network and coordinate with each other — but the specific percentages attached to them here are inconsistent across the secondary sources reporting them, and I couldn't trace any of them to a primary filing or vendor disclosure.

Worth distinguishing this piece from two others on this site that sound adjacent: this is about business-process agents — procurement, sourcing, logistics — running on cloud infrastructure covered in the broader agentic-cloud spending story, not the physical robots covered in my look at embodied AI and physical agents.

The view from Kathmandu

What strikes me most, watching this from a country where a customs form still gets filled out by hand at Birgunj, is how narrow the gap actually is between "agent negotiates a freight rate" and "agent fills out the form a person used to fill out." Most of what's shipping in 2026 — DHL's voice and email agents, project44's freight procurement agent, Oracle's sourcing agent — is closer to the second thing than the first: a genuine productivity layer on top of a process that still has a human approving the outcome, not an autonomous economic actor negotiating unsupervised. That's not a smaller story than the marketing implies. It's a more honest one, and it's the one Gartner's own numbers — real momentum, real cancellations coming — actually support.

Sources

• salesforce.com — Agentforce GA (Oct 25, 2024), version releases (Dec 2024–Oct 2025), Q4 FY26 earnings release (Feb 25, 2026): $800M ARR, +169% YoY, 29,000 deals
• Stifel (via Ask AI) — Agentforce customer-adoption estimate (~5.3% of Salesforce customer base)
• microsoft.com — Copilot Studio autonomous agents, Ignite Nov 19, 2024 and Ignite Nov 2025 (GPT-5 model choice, 1,400+ MCP-connected systems, Entra Agent ID)
• news.sap.com — SAP Connect, Oct 2025 (15 Joule agents; Production Planning & Operations Agent GA Q1 2026; Supply Chain Orchestration H1 2026)
• oracle.com — Fusion SCM agent press releases: Autonomous Sourcing Agent (Feb 10, 2026), Fusion Agentic Applications (Mar 24, 2026), four additional SCM agents (Jun 29, 2026)
• group.dhl.com — DHL and HappyRobot autonomous voice/email agents, Nov 11, 2025
• project44.com — AI Freight Procurement Agent (Feb 26, 2026), decision44 agent portfolio (Apr 8, 2026), Q1 FY27 ARR growth
• flexport.com — Winter 2026 Release, Flexport Atlas customs/compliance agents
• unilever.com; googlecloudpresscorner.com — Unilever/Google Cloud agentic-workflow announcement, Feb 17, 2026
• gartner.com — press releases, June 25, 2025 (33%/15% by 2028; 40%+ project cancellations by 2027) and August 26, 2025 (40% of enterprise apps by end of 2026)

Written by Abhishek Kushwaha, founder and writer at Global Tech Search, based in Kathmandu, Nepal.