Deep-Insights builds on the DI.AI Platform as its intelligent foundation and AI Agents as the execution units, covering the four core scenarios of warehousing, transportation, freight forwarding, and shipping—giving every link in the supply chain the autonomous closed-loop ability to "sense, decide, act, and self-heal."
> A single mid-scale supply chain disruption costs an enterprise on average 1%–4.5% of annual revenue. By 2026, 40% of enterprise applications will integrate task-oriented AI Agents. — Gartner
>
> Over the past year, "Agent" has been one of the most-mentioned words in the supply chain industry. But very few teams have actually moved Agents off PowerPoint and demos and into real customer warehouses, fleets, and customs brokers.
At Deep-Insights, we have re-organized 22 years of industry know-how, 55+ industry modules, 90+ business scenarios, and the 46 specialized AI Agents we are already running at scale for our customers.
The following is a summary of the keynote by Deep-Insights CEO Liu Bin at the 16th ACE 2026 Supply Chain Innovation Annual Flagship Summit, explaining exactly how we think about Agents.
Why Is Supply Chain Execution Stuck?
Before talking about Agents, let's talk about the problem itself. Today, the execution layer of China's supply chain is being squeezed by four issues at once:
1. Labor bottleneck. Warehouse workers are hard to recruit and dispatchers hard to retain. Labor costs rise more than 10% per year, yet experience cannot be quickly replicated—when a veteran dispatcher leaves, six months of operational turbulence is almost guaranteed.
2. Data silos. WMS, TMS, ERP, and OA run in parallel, with humans reconciling data over and over in Excel. The moment one node goes wrong, the entire chain is affected.
3. Reactive exceptions. Most supply chain organizations only discover problems after they have already happened. But every hour of delay translates into both a worse customer experience and real cost.
4. Optimization ceiling. Even the most experienced people have a cognitive ceiling. In the face of complex, massive data, the human eye simply cannot exhaust the cost-saving opportunities.
These four issues map exactly onto the four problem types that AI Agents are best at—repetitive labor, cross-system collaboration, real-time monitoring, and pattern recognition in massive data sets. This is precisely why we have made AI Agents the core product form for Deep-Insights for the next decade.
How We Think About Agents (I): Four Foundational Technical Problems
The market often blends every "Agent" narrative together. Internally, we break "technical implementation" into four independent—but mutually reinforcing—problems. Any team building Agents must first be clear about which one they are solving.
01 Agents Used by Humans — Human–Machine Collaboration
This is the easiest category to understand and the easiest to ship as a Copilot. The essence is human-in-the-loop: the Agent acts as an expert assistant, freeing people from low-value, highly repetitive actions, while the final decision, confirmation, and sign-off authority remain with the human.
In supply chain, typical examples are: an Agent suggesting the optimal load plan while a dispatcher schedules trucks; an Agent recommending quotes and risk flags while a CSR handles RFQs; an Agent generating charts and insights for managers running analyses.
It solves the "efficiency" problem, with very clear boundaries—humans remain the executor.
02 Agents Used by Agents — Agent Autonomy
This is the watershed between "toys" and "production-grade systems." When a task is complex enough to require multiple sub-capabilities, multiple systems, and multiple steps, a single Agent will not be enough. What is required is multi-agent orchestration: a master Agent decomposes the task, dispatches several sub-Agents, each handling its own role, communicating through structured protocols.
A warehouse Agent calls the cycle-count Agent, the capacity-analysis Agent, and the relocation-recommendation Agent, then aggregates everything back to operations—this is an Agent using Agents.
It solves the "autonomy" problem. Humans no longer need to spell out every step—just goals and boundaries.
03 Agent-Augmented Legacy Systems
This category has been almost ignored, yet its commercial value is the most underestimated. The vast majority of enterprises cannot tear down their WMS, TMS, and ERP overnight—they carry real business operations and the cost of replacement is unbearable.
But that does not mean AI cannot get in. Through the MCP standardized connection protocol, we "Agent-ify" existing systems: without changing the underlying architecture, Agents use these systems just like a new employee would—checking inventory, creating shipments, changing status, issuing instructions. From the customer's perspective, the WMS they bought is still the same WMS, but starting today, that WMS has a "thinking add-on."
It solves the "installed-base" problem and determines whether AI can really enter the daily workflows of most enterprises.
04 Agent-Native Software
The last category is software designed from scratch for Agents. It is not "legacy + AI"—it treats the Agent as a first-class citizen: object models, permission systems, audit logs, and exception handling are all rebuilt around the assumption that Agents will run here.
Inside Deep-Insights, the DI.AI Platform is exactly that foundation—not an upgrade of traditional SaaS, but an operating system purpose-built for a workforce of AI digital employees.
It solves the "future" problem. Five years from now, most newly launched supply chain software will most likely be Agent-native.
These four categories are not mutually exclusive. A mature enterprise AI strategy must run all four paths in parallel: use human–machine collaboration for quick wins, use Agent autonomy to crack complex scenarios, use Agent augmentation to revitalize the installed base, and use Agent-native systems to position for the long term.
How We Think About Agents (II): Scenario Adoption Is the Real Moat
After the technical paths, we have to talk about scenarios—because we are very clear that a generic LLM with a few prompts cannot break through real supply chain operations.
Why "46 Agents ≠ a Wrapper Around an LLM"
Many solutions in the market are essentially a generic LLM dressed up with a chat box and labeled an "industry Agent." They look stunning in a POC, but the moment they touch real customer operations they expose fundamental issues:
- They do not understand the business logic of WMS, TMS, freight forwarding, or shipping—they cannot even articulate how many parties and actions sit behind a single bill of lading.
- They cannot connect into the customer's enterprise systems—they cannot see locations in the WMS or trigger shipments in the TMS.
- They cannot guarantee SLAs or compliance—when something goes wrong, there is no traceability and no accountability.
- The end result: many pilots, very few real deployments.
Deep-Insights takes a fundamentally different path. Our 46 Agents have grown out of 22 years of industry data and processes, built on four pillars:
Industry know-how, internalized. 55+ industry modules and 90+ scenario maps are baked into a unified Ontology object model, so AI truly "understands" the structured relationships among orders, cargo, vessel schedules, and warehouse locations.
MCP standardized connectivity. WMS, TMS, ERP, customs APIs, carrier APIs, and IoT gateways are connected through standardized protocols rather than per-customer custom builds.
Bounded Autonomy. AI executes within explicit business boundaries; anything outside those boundaries is automatically escalated to a human reviewer. Every operation is auditable and traceable—not a black box.
Specialized scenario Agents. Not one giant do-it-all Agent, but warehouse, transportation, freight forwarding, shipping, customs, and commercial Agents—each playing its own role like real digital employees collaborating in a team. The whole foundation is supported by a five-layer architecture, from platform to scenario, weaving platform, connectivity, capability, orchestration, and scenario into a closed loop.
Real Slices of Scenario Adoption
A few Agents already running on customer sites:
1. Warehouse Capacity Analysis Agent
Connects to WMS location data in real time, automatically identifies high- and low-turnover SKUs, and recommends the optimal putaway locations. Location utilization rises 20%–35%, overflow risk drops 40%, and capacity alerts move from T+1 to real time.
2. Inbound Planning Agent
Automatically analyzes appointment time windows, schedules docks and unloading resources, and pushes exception alerts in real time. Inbound waiting time drops 50%, dock utilization rises 60%, and planning runs 24×7 with no human in the loop.
3. ABC Diagnostic Agent
Dynamically updates ABC classification, identifies dead, expiring, and overstocked inventory, and automatically generates inventory optimization recommendations. Turnover rises 30%, inventory carrying cost drops 25%.
4. Smart Load-Planning Agent
Replaces manual truck planning, route selection, and time-window monitoring, generating shipments end-to-end with one click. Picking efficiency rises 30%–50%; new-employee training shrinks from two weeks to two days.
Behind these numbers is a clear value curve:
- 50%+ repetitive labor saved—robots don't take leave, don't get tired, don't make mistakes.
- 95%+ process efficiency improvement—from hours to minutes, from reactive to real time.
- 99%+ SLA assurance—AI guards critical metrics so nothing slips through.
- 80%+ reduction in exception losses—from firefighting to prevention.
The Future of Software Is "Interface-less" — A Rate Negotiation Agent Example
The point I most want to emphasize is this: AI Agents will change supply chain not just by making existing software smarter, but by gradually causing the very concept of "software" to disappear.
The interaction logic of traditional software is "humans open the system, click buttons, fill out forms, submit workflows." Three decades of enterprise software evolution has, in essence, been about optimizing this process—better UI, fewer clicks, smoother flows. But it has never escaped the basic frame of "humans using software."
The real paradigm shift Agents bring is: humans no longer need to open software.
Our Rate Negotiation Agent, currently running on customer sites, is one validation of this view.
In the past, a single FTL tendering cycle looked like this:
- Operations opens the TMS
- Exports historical rates and opens email
- Contacts twenty carriers and consolidates quotes in Excel
- Opens OA to launch approval, then returns to TMS to maintain the new contract
Every back-and-forth, every quote revision, every approval trace was filled with system switching and manual actions. A mid-sized tender typically tied up 2–4 dedicated staff and ran for more than two weeks.
After deploying the Rate Negotiation Agent, this entire process becomes a single instruction: "Launch the annual tender for lane XX, target a 5% cost reduction, deadline next Friday."
Everything that follows is handled by the Agent: it pulls data across TMS, contract systems, email, and OA; runs multi-round negotiations with carriers based on our preset strategy; writes outcomes back to the contract system; auto-launches approvals according to rules; and escalates exceptions to a human. Throughout the process, operations never opened a single application UI—they only opened a chat with the Agent.
Results: reconciliation efficiency up 95%+, exception losses down 80%, 24×7 unattended operation, 2–4 dedicated FTEs saved per month. But the more important point is not the numbers—it is the shape of next-generation supply chain software:
- From "interface" to "conversation" — people drive the system through intent, not clicks.
- From "process" to "goal" — people no longer decompose every step; they only declare the goal and the boundaries.
- From "system" to "employee" — what used to be a piece of software gradually grows into a digital employee that can be "hired."
This is exactly why we define our products as AI Digital Employees rather than "AI features" or "AI modules"—because we believe that in the future, enterprises will hire not only humans, but also a workforce of AI employees that are always on, never tired, and clearly accountable.
Closing Thoughts
We are not wrapping a generic LLM and selling it to customers—we are using 22 years of industry data to train AI that truly understands supply chains.
Deep-Insights builds on the DI.AI Platform as its intelligent foundation and AI Agents as the execution units, covering warehousing, transportation, freight forwarding, and shipping, so that every link in the supply chain has the autonomous closed-loop ability to sense, decide, act, and self-heal.
Let AI Agents become core members of your supply chain team.