Head of AI, Air & LCL Forwarding at Maersk · Amsterdam
AI systems organizations actually use
Senior product leader and builder. I ship AI that Maersk’s freight teams use, and lead the PMs who scale it.
Impact Dashboard
Numbers with context, what was delivered, why it mattered, and how it was measured
Value delivered across the AI portfolio at Maersk
The 5-PM team I lead delivered $22M gross profit (Air) and $4M EBIT (LCL) in productivity gains; the automated pricing line I drove personally returned ~$11M EBIT.
Auto-rating adoption scaled from 14% to a peak of 80%
Automated pricing across the air & LCL freight unit, rollout phased by confidence band and gated on override rate, not launched on hope.
Pricing agents cut quote cycle time from 45 to 4 minutes
Productionized AI rating and quoting agents on the live quoting path, speed with guardrails, not autonomy for its own sake.
Drove platform growth at Blibli
Travel SaaS platform serving corporate, agency, and retail clients, the user base grew from 135K to 165K over the same period.
Accommodations across travel platform
OTA coverage expanded to 101 airlines, 3.6M+ accommodations, 49K activities, and 2.3K events, inventory breadth the platform could not have built organically.
Reduced agent handling time at Navan
Redesigned support-agent workflows around a GPT assistant, ~$1M in annual savings on top of $1.2M from the assistant itself.
Metrics shown are directional ranges abstracted for confidentiality. Exact figures are available during interviews under mutual NDA.
How AI Gets Adopted (Or Doesn’t)
The signature story: how automated pricing went from 14% to a peak of 80% adoption
The Situation
We'd shipped automated pricing across Maersk's air & LCL (less-than-container) freight business, a flagship AI bet in the middle of a unit turnaround. It cut quote cycle time from 45 minutes to 4. But adoption stalled at 14%, and on high-confidence outputs (the ones designed to auto-accept) ops teams were overriding 18% of them. That's the number that kills trust in an AI system.
What I Found
The consensus in the room: the model was hallucinating, we needed a better LLM. I root-caused the overrides and found the opposite. The model was reasoning perfectly over wrong facts. The same physical lane was represented three different ways across source systems. There was no canonical lane-and-commodity entity, and no controlled vocabulary tying commodity codes to surcharge rules. The agent retrieved the wrong surcharge and reasoned flawlessly over it.
The Obvious Play (Wrong)
The default playbook was sitting right there: swap in a newer model, tune the prompts, schedule more user training. I stopped the team before we spent a quarter on it. If the model was reasoning correctly over wrong facts, a smarter model would just be wrong faster, and more training would only teach people to distrust it more efficiently.
What Actually Worked
I halted the rollout and fixed the knowledge layer: defined a canonical lane-plus-commodity entity with a controlled vocabulary, and reconciled the master data behind the rate lookup. Then I set the bar at an eval gate, shadow-mode replay against known-good quotes, so nothing went live until accuracy cleared the floor. Override on high-confidence quotes dropped from 18% to ~4% in a single sprint. From there, rollout went region by region through super-users their teams already trusted.
The Lesson
Adoption ramped from 14% to a peak of 80%, unlocking ~$11M in EBIT on the automated pricing line. Same model, same features. The fix was the ontology underneath it, and a rollout gated on evidence instead of hope. The mechanism is domain-agnostic: whether the entities are shipping lanes, hotel inventory, or payment rails, an agent reasoning over a model that was never made canonical fails the same way.
“Adoption is a product problem. If your AI isn't used, it isn't finished.”
Flagship Systems
AI-driven platforms I've designed and shipped at scale
Touchless Booking Platform
Maersk Email-to-Booking Automation
AI pipeline that reads booking emails and attachments, resolves entities against master data, and creates bookings straight in the TMS, now the #1 digital booking channel.
Market Failure
~94% of LCL bookings were keyed by hand from emails and attachments, treated as a back-office cost to offshore, rather than the highest-leverage data-capture moment in the shipment lifecycle.
System Architecture
Outcome
#1 digital booking channel | 44% of bookings (from ~13%) | commercial-invoice match 50% → 98%
GPT-Powered Customer Assistant
Navan CX Automation Platform
Intelligent support automation, GPT query understanding with RAG knowledge retrieval, delivering ~$5M in annual value across assistant, workflows, and platform consolidation.
Market Failure
High-touch customer support operations couldn't scale with user growth, resulting in increased costs and degraded CSAT during peak demand.
System Architecture
Outcome
$1.2M assistant savings | $1M workflow redesign | $2.64M infra savings
Enterprise Travel SaaS Platform
Blibli Multi-Channel Growth Engine
Unified travel marketplace, aggregated 101 airlines + 3.6M accommodations via API layer, drove 137% revenue growth.
Market Failure
Fragmented OTA integrations and limited inventory prevented platform expansion, constraining revenue growth and market competitiveness.
System Architecture
Outcome
137% revenue growth | 3.6M+ accommodations | 22% user growth
Certain metrics, customers, and system details are intentionally abstracted due to NDA and commercial sensitivity.
What I Vibe Code
Side projects where I go from idea to deployed product with AI
FlowFit
Evidence-based wellness with RAG
A live app that generates personalized fitness protocols by retrieving and reasoning over structured scientific literature, every answer ships with an inspectable eval trace, so you can see exactly what the model retrieved and why.
CargoX
AI-driven logistics operations
An end-to-end freight operations demo unifying shipment oversight, data healing, and order management, AI validation with confidence-gated, human-in-the-loop escalation, modeled on the systems I run in production.
Decision Quality Framework
Production guardrails for LLM decisions
A reusable framework to evaluate, score, and govern LLM-driven decisions in production. Focuses on confidence thresholds, failure modes, and human-in-the-loop escalation.
Honest Fit Assessment
Paste a job description below. I'll give you an honest assessment of how well I fit the role, including my gaps.
This tool quickly maps my experience to your requirements, identifies skill gaps, and suggests conversation topics. It saves you time in initial screening while ensuring transparency about fit.
The Person Behind the Metrics
I spent the first decade of my career in airline and travel tech: passenger systems at Emirates and SITA, then travel platforms at Wipro and Blibli. The switch flipped at Navan in 2021, when we put an early GPT-powered assistant in front of real customers. It was accurate, and users still escalated to humans. They didn’t ask “is it right?”, they asked “why should I believe it?”. I’ve been building AI products around that question ever since.
The map behind that: New Delhi → Bangalore → Amsterdam, building for Indonesian travelers, global business travelers, and now freight operators on every continent at Maersk. Working across that many markets taught me that adoption is local. The same booking automation that hit near-universal adoption in Sweden sat near zero in Hong Kong until local super-users made the case in their own words. Trust doesn’t globalize, so I sequence every rollout region by region.
My favorite part of the day is vibe coding: this site, the interactive demos on it, and side projects like FlowFit and CargoX all shipped with Claude Code as my pair programmer. At work that habit became The Build Log, a weekly AI newsletter I publish to 100+ PMs, with a hands-on cohort that builds instead of reading about building. I don’t just ship AI products, I build with them.
How I build
Six tenets I use to build AI systems that scale across teams, users, and time.
Start with the decision, not the model
The model is the easy part. Name the decision it serves, its guardrails, and its fallback before you touch a prompt.
Production reality beats prototype perfection
A demo that wins the room can still die in production. I optimize for the version that survives real traffic.
Measure business impact, not activity
DAU and "AI interactions" are vanity. Every product I own ladders to a business lever finance already tracks.
Adoption is a product problem
If your AI is accurate and nobody uses it, it isn’t finished. Adoption is designed, not announced.
Evals are the new spec
You cannot spec a probabilistic system with a Word doc. The eval set is the requirement, and the gate.
Ship the prototype, then write the spec
Six-month specs bet the world will not change. Ship the prototype, learn from signal, then write the spec it earned.
Experience
10+ years scaling digital platforms across logistics, travel, and SaaS
Head of AI, Forwarding (Air & LCL)
A.P. Moller Maersk
- Lead a 5-PM team and orchestrate a ~40-person cross-functional org across ~6 AI product workstreams, team delivered $22M GP (Air) and $4M EBIT (LCL)
- Built Maersk's first touchless booking platform, now the #1 digital booking channel, creating ~2 in 5 bookings (44%, up from ~13%)
- Scaled automated pricing on an eval-gated rollout from 14% to a peak 80% adoption, unlocking $11M EBIT
- Shipped a file-compliance AI agent lifting commercial-invoice match accuracy ~50% → ~98%; pricing agents cut quote turnaround 45 → 4 min
- Retained a strategic ~$15M ARR account at renewal with AI milestone-quality fixes (ETA-to-port accuracy 50% → 92%)
Senior Product Manager - Experience
Navan (formerly TripActions)
- Launched GPT-powered customer assistant, delivering $1.2M annual savings and improved CSAT
- Redesigned agent workflows, reducing handling time by 10% ($1M annually)
- Consolidated 4 legacy communication systems into a unified CX platform, cutting $2.64M in yearly infrastructure costs
- Scaled product org from 1 → 3 scrum teams
Senior Product Manager - Travel SaaS
Blibli (via Coviam)
- Delivered SaaS travel platform with 137% revenue growth
- Expanded user base from 135K → 165K (22% growth)
- Expanded OTA coverage: 101 airlines, 3.6M+ accommodations, 49K activities
- Implemented advanced auth frameworks strengthening security
Senior Product Manager - Core Platform
Wipro Technologies
- Led Way2Go B2B travel & expense platform (50K users, 600+ trips/day)
- Directed 3 scrum teams (17 devs, 4 QA, 2 PMs, 1 UX)
- Defined platform roadmap and GTM strategy for enterprise clients
Technical Product / Program Manager
SITA
- Delivered large-scale PNR migration programs with 0.02% rejection rate
- Owned product roadmaps with international airline stakeholders
- Delivered scalable and compliant passenger service solutions
Let's Connect
Open to discussing product leadership roles, advisory opportunities, and speaking engagements
Harsh Puri
Senior Product Leader & Builder
Head of AI for Maersk’s air & LCL forwarding unit. 10+ years turning AI into measured P&L across logistics, travel, and SaaS.
For Recruiters: Looking for my resume? Download PDF or reach out directly.