Agents in Action: What It Takes to Deploy AI Agents in Real Supply Chains

In Project44’s “Agents in Action” webinar, Mark Strassman (Chief Product Officer, Project44) moderated a practical, production-focused panel with Jesse Buckingham (Founder, Vooma), Pablo Palafox (Co-founder & CEO, HappyRobot), and Kristopher Glotzbach (CRO, LunaPath). The conversation was refreshingly direct:

AI agents are no longer a “do they work?” question. Now it’s “how do we deploy them effectively?”

Below is a summary of the most actionable lessons for shippers, 3PLs, and brokerages evaluating (or scaling) agent automation.

Key Takeaways

1. The Agent Conversation Has Shifted: From Pilots to Production

Strassman described the last 18 months as a turning point, with companies moving past curiosity into execution. The focus now is on:

• Which workflows are ready today?

• What breaks in the real world?

• How do you scale beyond a single agent into a coordinated "agent workforce"?

Critical insight: Avoid the "one super-agent" trap. The market is converging on specialized agents, narrow, tuned for specific outcomes, because real operations demand speed, precision, and repeatability.

2. Exception Management Is a Decision Problem Disguised as a Communication Problem

When shipments go wrong, the challenge isn't just "who do we message?" It's deciding:

• Who to contact (driver vs. dispatcher vs. facility)

• Which channel (voice, SMS, email)

• Which language

• How urgent the issue is

• What action is most likely to resolve it quickly

Project44's orchestration framework:

• Analyze: Identify root cause, assess resolvability, gather context

• Optimize: Select channel, agent type, timing, and approach for highest success

• Orchestrate: Execute, log outcomes, and feed results back in real time

3. What "AI Agents" Actually Are (In Plain Terms)

Buckingham provided a clear definition: Traditional automation is deterministic "if-this-then-that" logic. An agent places a language model inside a workflow and gives it access to tools (email, texting, portals, browsers, APIs), so it can reason through ambiguity and complete objectives end-to-end.

Practical implication: Agents excel in workflows with edge cases that can't be fully pre-scripted.

4. Which Workflows Are Most Ready Right Now?

The most production-ready workflows share common traits: high-volume, repeatable, time-sensitive, and measurable.

Examples discussed:

• Visibility/track & trace interventions (status collection, early issue identification, proactive outreach)

• Documentation workflows (POD collection, invoice ingestion, reconciliation)

• Appointment scheduling

• Procurement-adjacent follow-up (supplier management, PO follow-ups)

These map to tactical communications: check calls, status cadences, POD/document retrieval, and multi-channel follow-ups that write clean updates back into systems of record.

5. Shippers vs. Brokers: Why Complexity Changes the Playbook

Glotzbach highlighted a key distinction:

• Broker workflows tend to be more linear and execution-heavy (downstream), with fewer branching outcomes

• Shipper environments often require reasoning across multiple systems and global constraints—inventory positioning, approval chains, multimodal requirements, tariffs, HTS codes, Incoterms, and more

The takeaway: Moving upstream increases potential returns but also demands normalized context and interoperability.

6. Data Quality: The Non-Negotiable (And the Practical Nuance)

The panel aligned on a central truth: agent decision quality depends on context quality.

Key perspectives:

• Buckingham: Combine clear SOPs ("what good looks like") with proper context and enable agents to detect and fetch missing information

• Glotzbach: A normalized visibility/data layer enables pattern detection and consistent execution at scale

• Palafox: Don't wait for perfect governance - start, learn where gaps exist, and improve as you scale. Agents can help extract "tribal knowledge" from frontline workflows

This supports the "start small, instrument outcomes, then scale" approach used by successful agent deployments.

7. Multi-Agent Orchestration: What Becomes Possible Next

The most forward-looking discussion centered on agent collaboration:

• Teams move from isolated point solutions to end-to-end lifecycle automation

• Insights in one area (maintenance capacity, equipment readiness, delay prediction) can trigger coordinated action elsewhere

• Operations shift from reactive firefighting to prescriptive action; not just detecting problems, but sequencing resolutions across systems and stakeholders

  
  

• Build vs. Buy: "Can You Build an Agent, or Can You Build a Workforce?"

Palafox reframed the build/buy debate with a compelling analogy: Most companies wouldn't build their own AWS. Similarly, building one agent is achievable, but building an observable, secure, continuously improving AI workforce is the real challenge.

This implies most teams will:

• Buy or partner for infrastructure (security, observability, memory, real-time data flows)

• Deploy specialists for specific workflows

• Orchestrate across partners via a platform layer

9. How to Get Value Fast: The Panel's Three-Point "Do This Now" Checklist

Immediate action steps:

1. Pick a narrow but meaningful workflow (high volume, measurable, clear success criteria)

2. Define "done" with metrics (SOW/PRD, baseline comparison, success thresholds)

3. Instrument, learn, and scale: build organizational muscle, then expand to adjacent workflows

This approach closes the demo-to-production gap and prevents pilots from stalling.

If your operation spends significant time on check calls, status updates, POD chasing, and exception follow-ups, the fastest path forward is to start with one workflow, prove value, and expand from there. To see what tactical automation looks like in practice, you can request a LunaPath demo here.