Y CombinatorRyan Petersen: How Flexport Uses AI to Cut Freight Costs
Through hackathons targeting customs broker emails and container routing; Flexport automated decisions that made ocean freight measurably cheaper for importers.
CHAPTERS
- 0:00 – 0:52
How AI reduces ocean freight cost while improving transit time
Ryan opens with the core claim: logistics is scale-driven, and AI-driven automation can materially lower shipping prices. He shares an early concrete result—2% lower ocean freight spend alongside a 20% improvement in transit time, a trade-off that usually doesn’t happen.
- •Logistics economics: bigger scale drives lower unit cost
- •Goal: 8–10% cheaper ocean container shipping over the next few years
- •AI already saved ~2% of ocean freight spend
- •Automation improved transit time ~20% while cutting costs
- •Framing: faster vs cheaper is usually a trade-off—AI changes that
- 0:52 – 3:16
What Flexport is and where AI fits across the workflow
Ryan defines Flexport as a global logistics company built on a modern tech stack spanning ocean, air, trucking, and rail. He describes AI as a broad layer across customer experience, planning, container loading, contracting, and automating email/phone-based work.
- •Flexport moves cargo end-to-end across multiple transport modes
- •AI applied to customer data access and user experience
- •Optimizing container loading and ship selection for cost and time
- •Automating high-volume “email/phone” operational work
- •Challenge: messy inputs (e.g., giant Excel contracts) require intelligent parsing pipelines
- 3:16 – 6:27
From ChatGPT fascination to company-wide urgency (and incumbent advantages)
The conversation shifts to when AI became “serious” inside a pre-AI, scaled company. Ryan explains how post–ChatGPT momentum spread unevenly at first, and why incumbents can win with AI due to data, domain expertise, and distribution—especially if they control their own codebase.
- •Ryan’s personal inflection point: ChatGPT launch (late 2022)
- •Internal adoption required leadership pressure to avoid “boomer company” inertia
- •Incumbent advantages: data scale, domain knowledge, and distribution
- •Flexport advantage vs competitors: owned tech stack enables deep AI integration
- •Competitors often buy IT tools and can’t easily modify workflows with AI
- 6:27 – 9:37
Hackathons as the AI product engine (bottoms-up innovation)
Ryan describes Flexport’s hackathons as the main conduit for AI experimentation turning into real features. He contrasts his shift toward “founder mode” (more directive leadership) with the need to preserve room for bottom-up, frontline-driven AI ideas.
- •Flexport runs 1–2 hackathons per year; now “religious” about two annually
- •Recent hackathons are ~90% LLM-based projects (vs far fewer 18 months earlier)
- •LLM era makes hackathon prototypes more likely to become shippable features
- •Leadership tension: top-down roadmap vs emergent frontline applications
- •Process idea: schedule hackathons before roadmap budgeting to fund best ideas
- 9:37 – 12:17
Upskilling non-engineers: turning domain experts into AI builders
Ryan outlines an internal program that gives non-engineers time and training to build AI-driven automations. The goal is to convert repetitive operational work into lightweight apps and workflows built by the people closest to the problem.
- •90-day program: one day/week for non-engineers to learn AI and “vibe coding”
- •Tools mentioned: Cursor, Streamlit-style app building, workflow automation
- •Freight forwarding as “freight email forwarding”: ripe for repetitive task automation
- •Domain experts identify automations best because they do the work daily
- •Program began in Amsterdam and is being rolled out globally
- 12:17 – 13:40
Most impactful AI projects: natural-language analytics and reporting automation
Ryan shares customer-facing wins, especially making supply chain data accessible via natural language. This reduces the burden of building dashboards/SQL queries and cuts a major chunk of account management time spent generating reports.
- •Flexport workflow: POs → factories → bookings → execution and delivery
- •Customers care about on-time performance, SKU metrics, cost, and customs/tariffs data
- •Hackathon-born feature: ask questions in natural language to generate charts/tables
- •Eliminates need for SQL/dashboards for many users
- •Reduces account-management workload (report generation was ~25% of their time)
- 13:40 – 15:37
Planning optimization at scale: why software beats humans in logistics timing
Ryan explains a long-running ML/optimization system that chooses shipping plans based on contracts, schedules, and variability. He highlights a key dynamic—continuous replanning triggered by cancellations—that humans could not execute frequently enough to capture big transit-time gains.
- •Pre-LLM AI/ML used for planning: selecting ship/routes given price and timing
- •Results: ~2% ocean spend reduction and ~20% faster transit times
- •Key operational reality: ~2,000 container cancellations per week
- •System replans repeatedly (e.g., 10x/day), pulling shipments forward when slots open
- •Human replanning at that frequency would require prohibitive labor cost
- 15:37 – 19:07
Agentic AI in operations: tool use, email/voice workflows, and error reduction
The group explores what changes when LLMs can use tools and communicate like humans—emailing customers, calling warehouses, and validating details. Ryan gives examples of automations that either replace work or make previously “too expensive” work feasible.
- •Future pattern: LLM as an agent orchestrating tools (solver as one tool)
- •Agents could proactively contact customers to confirm plan changes
- •LLM agent validates warehouse addresses and secures delivery appointments via email/voice
- •Automation enables tasks that were uneconomic for humans to do every time
- •AI monitors customer messages for sentiment and escalates unhappy customers to managers
- 19:07 – 20:11
Will automation make goods cheaper? Freight forwarding labor share and price impact
Ryan quantifies the cost structure: the freight-forwarding labor layer is about 10% of the end-to-end cost importers/exporters pay for containerized ocean freight. Full automation could largely compress that layer, translating into an estimated ~8–10% reduction in shipping costs over time.
- •Freight forwarding labor is ~10% of the end cost paid by shipper
- •Automation can reduce the cost of moving goods, not necessarily the goods themselves
- •Flexport estimate: ~8–10% ocean shipping price reduction over the next few years
- •Lower shipping costs could have broader macroeconomic effects (trade/GDP), though tariffs complicate measurement
- •Cost reductions are driven by scale + automation reinforcing each other
- 20:11 – 23:51
AI and society: GDP growth, jobs fears, and an ‘Axial Age’ analogy
Ryan argues that companies exist to deliver goods/services cheaply, not primarily to employ people, and that productivity gains will redirect human effort rather than end it. He then pivots to a historical analogy: technology-driven shifts (coins, the internet, AI) can disrupt trust and social norms, requiring new philosophical frameworks.
- •“White pill” framing: AI could drive sustained GDP growth (e.g., 7%/year)
- •Job-loss concerns are framed as misunderstanding the purpose of firms and competition
- •Humans will continue to want more, seek meaning, and keep producing/contributing
- •Axial Age analogy: coins made transactions impersonal, eroding trust and shifting society
- •Parallel to internet/AI: society may need new moral/philosophical adaptation to tech
- 23:51 – 26:38
How AI reshapes company structure: humans-in-the-loop, liability, and relationships
The hosts propose a future where companies run as AI cores with humans as mandated approvers and relationship managers. Ryan agrees that as long as humans remain the end customer, relationships and accountability will keep humans relevant—especially in regulated domains like customs.
- •Regulation can require humans-in-the-loop (example: fintech loan approvals)
- •Customs brokerage also requires human approval before clearing customs
- •Risk: “Accept all changes” behavior turns humans into rubber-stamp liability sinks
- •Ryan’s view: humans still prefer human relationships in commerce (for a long time)
- •Example: AI ‘spellchecker’ reduces customs filing mistakes (e.g., Austria vs Australia codes)
- 26:38 – 27:43
If Ryan started Flexport today: don’t over-automate the messy tail
Ryan says he wouldn’t build Flexport as a purely software company even now. Logistics still demands operational grit—phone calls, port visits, and bespoke problem solving—while automation should focus on the repeatable core rather than edge-case “tails.”
- •Flexport’s edge: willingness to do real-world ops (phone calls, port visits)
- •Example: handling unusual customer needs (e.g., cranes, special unloading) with hands-on execution
- •Warning against “no API, can’t do it” mentality in traditional industries
- •Don’t automate the long tail prematurely; use humans where variability is high
- •Pragmatism: combine tech leverage with operational excellence
- 27:43 – 30:21
Founder advice on big funding rounds: discipline, control, and avoiding headcount bloat
Ryan discusses how to think about raising capital: optimize for price per share and maintain control, but beware that money pushes organizations toward spending. He recommends imposing a hiring freeze after raising to prevent cultural drift toward “money solves problems.”
- •Raising can be good if price per share increases and control is maintained
- •Control matters legally, culturally, and operationally
- •Common failure mode: using money to solve problems by reflexively hiring
- •Advice: raise the round, then do a 90-day hiring freeze to set discipline
- •Headcount bloat slows execution and creates a bad problem-solving culture
- 30:21 – 33:40
Flexport’s 2035 vision: global utility logistics powered by automation
Ryan closes with the long-term product vision: logistics should become a utility—easy, reliable, low-cost, and accessible via APIs/voice—so companies can focus on building products. Achieving it requires both deep automation and broader global coverage with Flexport’s own teams on the ground.
- •Vision: ship anything anywhere, any mode/quantity, executed “via code”
- •Goal: make logistics as invisible as the electrical grid for customers
- •Current reality: shipped to/from 147 countries, employees in 22 countries
- •Near-term roadmap: cover ~95% of container trade with in-country teams by 2028
- •Competitive stance: strong tech lead; expansion challenge is global operational footprint
