Where’s The Generative AI ROI? Start With The Supply Chain.
An industry that communicates with unstructured documents turns to generative AI to clean up its processes. And it's delivering real returns.
The way our supply chain communicates is a mess. To find out what’s going where, shippers today use faxes, paper contracts, and lots of unstructured documents. Errors, delays, and lengthy wait times can ensue. If you’ve ever lost an item en route to your home, there’s a good chance this is why.
Now, generative AI is starting to solve the problem, delivering meaningful returns for those employing it. While the technology’s more glamorous applications (including ultra-smart chatbots) have yet to turn a profit, it’s delivering real ROI for supply chain operators, enabling them to spot errors, triage, and communicate faster than they ever could previously. The early success indicates where the tech might find near-term results — solving these ‘boring’ but valuable problems — as its long-term promise remains just that.
Alex Nederlof, engineering director at Flexport, a supply chain technology company, told me the company has aggressively deployed generative AI to make sense of the vast amount of unstructured documents it uses to operate. Shipping contracts can run forty pages, with rows and subrows, tables and nested subtables, all filled with crucial information. Bills of lading — documents that say what’s in a container and when it will arrive — aren’t standardized and are difficult for traditional computer systems to ingest. All this adds lots of inefficiency (and cost) to global supply chain.
“We're getting to the end of 2024 and we have physical paper that comes into the office with freight documents that we’re still scanning and keying into computers, which is mental compared to most other industries,” said Nederlof. “We're like the taxi industry, pre-Uber.”
When ChatGPT arrived, Nederlof and his colleagues perked up. The underlying LLM technology became immensely useful once it went multimodal, adding vision capabilities that could examine long documents and parse the information within them. With each successive release — from GPT-3.5 to GPT-4 and eventually GPT-4o — generative AI became better at handling tasks that previously bogged the company down. Nederlof’s team eventually built it directly into Flexport’s technology.
Today, Flexport uses generative AI to examine documents and make sense of them, and input the important information into its systems via an automation layer. Humans check the AI’s work, accounting for the technology’s hallucination problem, before accepting it as final. Then, Flexport’s employees take action when something looks off or requires immediate intervention.
The deployment is already working at scale. Flexport’s AI-powered document parser is processing 15,000 documents per month, saving its operators 80% of the time they’d spend on the same tasks, the company told me. Nederlof is aiming to get to 50% automation by next year, if not more.
“Parsing the document is about five cents on OpenAI with vision and then a couple of prompts,” Nederlof said. “So you go from a document parsing being anywhere between five and ten dollars to spending five cents, maybe 10 cents, worst case, on the automation. Then a human looking at it at one dollar if they can go fast through the documents.”
“We're like the taxi industry, pre-Uber.”
Generative AI is particularly useful in customs work, where shippers apply different rules to different items. Flexport, for instance, could get a bill of lading with t-shirts of all sizes, but have to apply different tariffs to adult and children clothing. Previously, humans would have to key in all this data and apply the particular rules — a lengthy process subject to errors — and now AI can make sense of the documents and apply the rules itself.
With the help of the AI, Flexport can grow faster. It’s not using the technology to cut employees. “We're a growth company,” said Nederlof “We handle more volume. It's that simple.”
Nederlof and his colleagues are now starting to apply the technology to more ambitious projects. One proof of concept they’ve worked on is a voice AI that will call truckers and ask them if they’re available to take on certain jobs. “They have a real conversation from which we can determine if they're available, when they could pick it up, and a bunch of data points we need to extract,” Nederlof said. “It just puts that into the system, and they book a journey.”
As the much-hyped AI advancements like agents, assistants, and human-level intelligence take time — and piles of cash — to develop, it’s these type applications that will likely keep the field humming, and justify further investment.
“It's so boring, but it is so valuable,” Aidan Gomez, CEO of Cohere, which is working to apply generative AI for businesses use cases, told me this week. “People don't understand the actual scale of impact of some of these crucial, banal things. And if we can scale them up, make them more accurate, more reliable, it really is world-changing.”
Is your financial future riding on just one or two stocks? (sponsor)
Are you incurring huge tax bills when you sell your appreciated stocks? The Cache Exchange Fund, benchmarked to the Nasdaq-100, is built to diversify your stocks while deferring taxes. Engineers to CEOs across tech and finance have turned to Cache to manage their concentrated stocks. New investors are onboarded monthly, and their next fund close is on Oct 31st.
Learn more about the Cache Exchange Fund.
Advertise on Big Technology?
Reach 145,000+ plugged-in tech readers with your company’s latest campaign, product, or thought leadership. To learn more, write alex@bigtechnology.com or reply to this email.
What Else I’m Reading, Etc.
Elon Musk introduces Tesla’s autonomous robot taxi [The Verge]
DeepMind co-founder Demis Hassabis wins the Nobel Prize for AI work [NPR]
Roblox accused of inflated numbers, safety issues for kids [WSJ]
Apple Intelligence summarizes man’s breakup text [Ars Technica]
Taylor Lorenz exits the Washington Post [NPR]
With RTO policies, executives long for the good ol’ days (that weren’t really so good) [Bloomberg]
Number of The Week
$14 billion
Potential OpenAI losses in 2026, according to internal documents.
Quote of The Week
“No one wants to read AI-generated news”
OpenAI head of media partnerships Varun Shetty, speaking to publishers at a conference in Brussels this week
This Week on Big Technology Podcast: Google's Head Of Search On Its Big AI Bet — With Liz Reid
Elizabeth Reid is Google's VP of Search and a 20-plus year veteran of the company. Reid joins Big Technology to discuss Google's major AI push and how this new era of computing is transforming its product and business. Tune in to hear about AI-organized search results, the growth of Google Lens, and how AI is changing the way people interact with search. We also cover the potential impact on publishers, Google's approach to responsible AI deployment, and the future of AR technology. Hit play for an insider's look at how the world's largest search engine is evolving in the age of AI.
You can listen on Apple, Spotify, or wherever you get your podcasts.
Thanks again for reading. Please share Big Technology if you like it!
And hit that Like Button
My book Always Day One digs into the tech giants’ inner workings, focusing on automation and culture. I’d be thrilled if you’d give it a read. You can find it here.
Questions? News tips? Email me by responding to this email, or by writing alex@bigtechnology.com Or find me on Signal at 516-695-8680
Is there any information on error rates and general effectiveness, aside from the anecdotal reports cited in this article?
I have largely given up on similar use cases because of GPT's routine mistakes, such as failing to identify all relevant figures in images and tabular data, along with getting the arithmetic wrong even if everything is parsed correctly.
Instead, rather than using AI to deal with garbage, it seems much better to streamline the processes and reduce the amounts of garbage in the first place.
Yeah I’ve come to similar conclusions trying integrations myself. This article really does whet the appetite to look into these kind of document review applications again though. I’d love to understand better how Flexport got this stuff working.